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/file.h>
19 #include <linux/slab.h>
20 #include <linux/sched/clock.h>
21 #include <linux/sched/signal.h>
22 #include <linux/vmalloc.h>
24 #include <linux/net.h>
25 #include <linux/if_packet.h>
26 #include <linux/if_arp.h>
27 #include <linux/if_tun.h>
28 #include <linux/if_macvlan.h>
29 #include <linux/if_tap.h>
30 #include <linux/if_vlan.h>
31 #include <linux/skb_array.h>
32 #include <linux/skbuff.h>
38 static int experimental_zcopytx
= 1;
39 module_param(experimental_zcopytx
, int, 0444);
40 MODULE_PARM_DESC(experimental_zcopytx
, "Enable Zero Copy TX;"
41 " 1 -Enable; 0 - Disable");
43 /* Max number of bytes transferred before requeueing the job.
44 * Using this limit prevents one virtqueue from starving others. */
45 #define VHOST_NET_WEIGHT 0x80000
47 /* Max number of packets transferred before requeueing the job.
48 * Using this limit prevents one virtqueue from starving others with small
51 #define VHOST_NET_PKT_WEIGHT 256
53 /* MAX number of TX used buffers for outstanding zerocopy */
54 #define VHOST_MAX_PEND 128
55 #define VHOST_GOODCOPY_LEN 256
58 * For transmit, used buffer len is unused; we override it to track buffer
59 * status internally; used for zerocopy tx only.
61 /* Lower device DMA failed */
62 #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3)
63 /* Lower device DMA done */
64 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
65 /* Lower device DMA in progress */
66 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
68 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
70 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
73 VHOST_NET_FEATURES
= VHOST_FEATURES
|
74 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR
) |
75 (1ULL << VIRTIO_NET_F_MRG_RXBUF
) |
76 (1ULL << VIRTIO_F_IOMMU_PLATFORM
)
85 struct vhost_net_ubuf_ref
{
86 /* refcount follows semantics similar to kref:
87 * 0: object is released
88 * 1: no outstanding ubufs
89 * >1: outstanding ubufs
92 wait_queue_head_t wait
;
93 struct vhost_virtqueue
*vq
;
96 #define VHOST_RX_BATCH 64
97 struct vhost_net_buf
{
98 struct sk_buff
**queue
;
103 struct vhost_net_virtqueue
{
104 struct vhost_virtqueue vq
;
107 /* vhost zerocopy support fields below: */
108 /* last used idx for outstanding DMA zerocopy buffers */
110 /* first used idx for DMA done zerocopy buffers */
112 /* an array of userspace buffers info */
113 struct ubuf_info
*ubuf_info
;
114 /* Reference counting for outstanding ubufs.
115 * Protected by vq mutex. Writers must also take device mutex. */
116 struct vhost_net_ubuf_ref
*ubufs
;
117 struct skb_array
*rx_array
;
118 struct vhost_net_buf rxq
;
122 struct vhost_dev dev
;
123 struct vhost_net_virtqueue vqs
[VHOST_NET_VQ_MAX
];
124 struct vhost_poll poll
[VHOST_NET_VQ_MAX
];
125 /* Number of TX recently submitted.
126 * Protected by tx vq lock. */
128 /* Number of times zerocopy TX recently failed.
129 * Protected by tx vq lock. */
130 unsigned tx_zcopy_err
;
131 /* Flush in progress. Protected by tx vq lock. */
135 static unsigned vhost_net_zcopy_mask __read_mostly
;
137 static void *vhost_net_buf_get_ptr(struct vhost_net_buf
*rxq
)
139 if (rxq
->tail
!= rxq
->head
)
140 return rxq
->queue
[rxq
->head
];
145 static int vhost_net_buf_get_size(struct vhost_net_buf
*rxq
)
147 return rxq
->tail
- rxq
->head
;
150 static int vhost_net_buf_is_empty(struct vhost_net_buf
*rxq
)
152 return rxq
->tail
== rxq
->head
;
155 static void *vhost_net_buf_consume(struct vhost_net_buf
*rxq
)
157 void *ret
= vhost_net_buf_get_ptr(rxq
);
162 static int vhost_net_buf_produce(struct vhost_net_virtqueue
*nvq
)
164 struct vhost_net_buf
*rxq
= &nvq
->rxq
;
167 rxq
->tail
= skb_array_consume_batched(nvq
->rx_array
, rxq
->queue
,
172 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue
*nvq
)
174 struct vhost_net_buf
*rxq
= &nvq
->rxq
;
176 if (nvq
->rx_array
&& !vhost_net_buf_is_empty(rxq
)) {
177 skb_array_unconsume(nvq
->rx_array
, rxq
->queue
+ rxq
->head
,
178 vhost_net_buf_get_size(rxq
));
179 rxq
->head
= rxq
->tail
= 0;
183 static int vhost_net_buf_peek(struct vhost_net_virtqueue
*nvq
)
185 struct vhost_net_buf
*rxq
= &nvq
->rxq
;
187 if (!vhost_net_buf_is_empty(rxq
))
190 if (!vhost_net_buf_produce(nvq
))
194 return __skb_array_len_with_tag(vhost_net_buf_get_ptr(rxq
));
197 static void vhost_net_buf_init(struct vhost_net_buf
*rxq
)
199 rxq
->head
= rxq
->tail
= 0;
202 static void vhost_net_enable_zcopy(int vq
)
204 vhost_net_zcopy_mask
|= 0x1 << vq
;
207 static struct vhost_net_ubuf_ref
*
208 vhost_net_ubuf_alloc(struct vhost_virtqueue
*vq
, bool zcopy
)
210 struct vhost_net_ubuf_ref
*ubufs
;
211 /* No zero copy backend? Nothing to count. */
214 ubufs
= kmalloc(sizeof(*ubufs
), GFP_KERNEL
);
216 return ERR_PTR(-ENOMEM
);
217 atomic_set(&ubufs
->refcount
, 1);
218 init_waitqueue_head(&ubufs
->wait
);
223 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref
*ubufs
)
225 int r
= atomic_sub_return(1, &ubufs
->refcount
);
227 wake_up(&ubufs
->wait
);
231 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref
*ubufs
)
233 vhost_net_ubuf_put(ubufs
);
234 wait_event(ubufs
->wait
, !atomic_read(&ubufs
->refcount
));
237 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref
*ubufs
)
239 vhost_net_ubuf_put_and_wait(ubufs
);
243 static void vhost_net_clear_ubuf_info(struct vhost_net
*n
)
247 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
248 kfree(n
->vqs
[i
].ubuf_info
);
249 n
->vqs
[i
].ubuf_info
= NULL
;
253 static int vhost_net_set_ubuf_info(struct vhost_net
*n
)
258 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
259 zcopy
= vhost_net_zcopy_mask
& (0x1 << i
);
262 n
->vqs
[i
].ubuf_info
= kmalloc(sizeof(*n
->vqs
[i
].ubuf_info
) *
263 UIO_MAXIOV
, GFP_KERNEL
);
264 if (!n
->vqs
[i
].ubuf_info
)
270 vhost_net_clear_ubuf_info(n
);
274 static void vhost_net_vq_reset(struct vhost_net
*n
)
278 vhost_net_clear_ubuf_info(n
);
280 for (i
= 0; i
< VHOST_NET_VQ_MAX
; i
++) {
281 n
->vqs
[i
].done_idx
= 0;
282 n
->vqs
[i
].upend_idx
= 0;
283 n
->vqs
[i
].ubufs
= NULL
;
284 n
->vqs
[i
].vhost_hlen
= 0;
285 n
->vqs
[i
].sock_hlen
= 0;
286 vhost_net_buf_init(&n
->vqs
[i
].rxq
);
291 static void vhost_net_tx_packet(struct vhost_net
*net
)
294 if (net
->tx_packets
< 1024)
297 net
->tx_zcopy_err
= 0;
300 static void vhost_net_tx_err(struct vhost_net
*net
)
305 static bool vhost_net_tx_select_zcopy(struct vhost_net
*net
)
307 /* TX flush waits for outstanding DMAs to be done.
308 * Don't start new DMAs.
310 return !net
->tx_flush
&&
311 net
->tx_packets
/ 64 >= net
->tx_zcopy_err
;
314 static bool vhost_sock_zcopy(struct socket
*sock
)
316 return unlikely(experimental_zcopytx
) &&
317 sock_flag(sock
->sk
, SOCK_ZEROCOPY
);
320 /* In case of DMA done not in order in lower device driver for some reason.
321 * upend_idx is used to track end of used idx, done_idx is used to track head
322 * of used idx. Once lower device DMA done contiguously, we will signal KVM
325 static void vhost_zerocopy_signal_used(struct vhost_net
*net
,
326 struct vhost_virtqueue
*vq
)
328 struct vhost_net_virtqueue
*nvq
=
329 container_of(vq
, struct vhost_net_virtqueue
, vq
);
333 for (i
= nvq
->done_idx
; i
!= nvq
->upend_idx
; i
= (i
+ 1) % UIO_MAXIOV
) {
334 if (vq
->heads
[i
].len
== VHOST_DMA_FAILED_LEN
)
335 vhost_net_tx_err(net
);
336 if (VHOST_DMA_IS_DONE(vq
->heads
[i
].len
)) {
337 vq
->heads
[i
].len
= VHOST_DMA_CLEAR_LEN
;
343 add
= min(UIO_MAXIOV
- nvq
->done_idx
, j
);
344 vhost_add_used_and_signal_n(vq
->dev
, vq
,
345 &vq
->heads
[nvq
->done_idx
], add
);
346 nvq
->done_idx
= (nvq
->done_idx
+ add
) % UIO_MAXIOV
;
351 static void vhost_zerocopy_callback(struct ubuf_info
*ubuf
, bool success
)
353 struct vhost_net_ubuf_ref
*ubufs
= ubuf
->ctx
;
354 struct vhost_virtqueue
*vq
= ubufs
->vq
;
359 /* set len to mark this desc buffers done DMA */
360 vq
->heads
[ubuf
->desc
].len
= success
?
361 VHOST_DMA_DONE_LEN
: VHOST_DMA_FAILED_LEN
;
362 cnt
= vhost_net_ubuf_put(ubufs
);
365 * Trigger polling thread if guest stopped submitting new buffers:
366 * in this case, the refcount after decrement will eventually reach 1.
367 * We also trigger polling periodically after each 16 packets
368 * (the value 16 here is more or less arbitrary, it's tuned to trigger
369 * less than 10% of times).
371 if (cnt
<= 1 || !(cnt
% 16))
372 vhost_poll_queue(&vq
->poll
);
374 rcu_read_unlock_bh();
377 static inline unsigned long busy_clock(void)
379 return local_clock() >> 10;
382 static bool vhost_can_busy_poll(unsigned long endtime
)
384 return likely(!need_resched() && !time_after(busy_clock(), endtime
) &&
385 !signal_pending(current
));
388 static void vhost_net_disable_vq(struct vhost_net
*n
,
389 struct vhost_virtqueue
*vq
)
391 struct vhost_net_virtqueue
*nvq
=
392 container_of(vq
, struct vhost_net_virtqueue
, vq
);
393 struct vhost_poll
*poll
= n
->poll
+ (nvq
- n
->vqs
);
394 if (!vq
->private_data
)
396 vhost_poll_stop(poll
);
399 static int vhost_net_enable_vq(struct vhost_net
*n
,
400 struct vhost_virtqueue
*vq
)
402 struct vhost_net_virtqueue
*nvq
=
403 container_of(vq
, struct vhost_net_virtqueue
, vq
);
404 struct vhost_poll
*poll
= n
->poll
+ (nvq
- n
->vqs
);
407 sock
= vq
->private_data
;
411 return vhost_poll_start(poll
, sock
->file
);
414 static int vhost_net_tx_get_vq_desc(struct vhost_net
*net
,
415 struct vhost_virtqueue
*vq
,
416 struct iovec iov
[], unsigned int iov_size
,
417 unsigned int *out_num
, unsigned int *in_num
,
420 unsigned long uninitialized_var(endtime
);
421 int r
= vhost_get_vq_desc(vq
, vq
->iov
, ARRAY_SIZE(vq
->iov
),
422 out_num
, in_num
, NULL
, NULL
);
424 if (r
== vq
->num
&& vq
->busyloop_timeout
) {
426 endtime
= busy_clock() + vq
->busyloop_timeout
;
427 while (vhost_can_busy_poll(endtime
)) {
428 if (vhost_has_work(vq
->dev
)) {
429 *busyloop_intr
= true;
432 if (!vhost_vq_avail_empty(vq
->dev
, vq
))
437 r
= vhost_get_vq_desc(vq
, vq
->iov
, ARRAY_SIZE(vq
->iov
),
438 out_num
, in_num
, NULL
, NULL
);
444 static bool vhost_exceeds_maxpend(struct vhost_net
*net
)
446 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_TX
];
447 struct vhost_virtqueue
*vq
= &nvq
->vq
;
449 return (nvq
->upend_idx
+ UIO_MAXIOV
- nvq
->done_idx
) % UIO_MAXIOV
>
450 min_t(unsigned int, VHOST_MAX_PEND
, vq
->num
>> 2);
453 /* Expects to be always run from workqueue - which acts as
454 * read-size critical section for our kind of RCU. */
455 static void handle_tx(struct vhost_net
*net
)
457 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_TX
];
458 struct vhost_virtqueue
*vq
= &nvq
->vq
;
461 struct msghdr msg
= {
466 .msg_flags
= MSG_DONTWAIT
,
468 size_t len
, total_len
= 0;
472 struct vhost_net_ubuf_ref
*uninitialized_var(ubufs
);
473 bool zcopy
, zcopy_used
;
476 mutex_lock(&vq
->mutex
);
477 sock
= vq
->private_data
;
481 if (!vq_iotlb_prefetch(vq
))
484 vhost_disable_notify(&net
->dev
, vq
);
485 vhost_net_disable_vq(net
, vq
);
487 hdr_size
= nvq
->vhost_hlen
;
493 /* Release DMAs done buffers first */
495 vhost_zerocopy_signal_used(net
, vq
);
497 busyloop_intr
= false;
498 head
= vhost_net_tx_get_vq_desc(net
, vq
, vq
->iov
,
500 &out
, &in
, &busyloop_intr
);
501 /* On error, stop handling until the next kick. */
502 if (unlikely(head
< 0))
504 /* Nothing new? Wait for eventfd to tell us they refilled. */
505 if (head
== vq
->num
) {
506 if (unlikely(busyloop_intr
)) {
507 vhost_poll_queue(&vq
->poll
);
508 } else if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
509 vhost_disable_notify(&net
->dev
, vq
);
515 vq_err(vq
, "Unexpected descriptor format for TX: "
516 "out %d, int %d\n", out
, in
);
519 /* Skip header. TODO: support TSO. */
520 len
= iov_length(vq
->iov
, out
);
521 iov_iter_init(&msg
.msg_iter
, WRITE
, vq
->iov
, out
, len
);
522 iov_iter_advance(&msg
.msg_iter
, hdr_size
);
524 if (!msg_data_left(&msg
)) {
525 vq_err(vq
, "Unexpected header len for TX: "
526 "%zd expected %zd\n",
530 len
= msg_data_left(&msg
);
532 zcopy_used
= zcopy
&& len
>= VHOST_GOODCOPY_LEN
533 && !vhost_exceeds_maxpend(net
)
534 && vhost_net_tx_select_zcopy(net
);
536 /* use msg_control to pass vhost zerocopy ubuf info to skb */
538 struct ubuf_info
*ubuf
;
539 ubuf
= nvq
->ubuf_info
+ nvq
->upend_idx
;
541 vq
->heads
[nvq
->upend_idx
].id
= cpu_to_vhost32(vq
, head
);
542 vq
->heads
[nvq
->upend_idx
].len
= VHOST_DMA_IN_PROGRESS
;
543 ubuf
->callback
= vhost_zerocopy_callback
;
544 ubuf
->ctx
= nvq
->ubufs
;
545 ubuf
->desc
= nvq
->upend_idx
;
546 refcount_set(&ubuf
->refcnt
, 1);
547 msg
.msg_control
= ubuf
;
548 msg
.msg_controllen
= sizeof(ubuf
);
550 atomic_inc(&ubufs
->refcount
);
551 nvq
->upend_idx
= (nvq
->upend_idx
+ 1) % UIO_MAXIOV
;
553 msg
.msg_control
= NULL
;
558 if (total_len
< VHOST_NET_WEIGHT
&&
559 !vhost_vq_avail_empty(&net
->dev
, vq
) &&
560 likely(!vhost_exceeds_maxpend(net
))) {
561 msg
.msg_flags
|= MSG_MORE
;
563 msg
.msg_flags
&= ~MSG_MORE
;
566 /* TODO: Check specific error and bomb out unless ENOBUFS? */
567 err
= sock
->ops
->sendmsg(sock
, &msg
, len
);
568 if (unlikely(err
< 0)) {
570 vhost_net_ubuf_put(ubufs
);
571 nvq
->upend_idx
= ((unsigned)nvq
->upend_idx
- 1)
574 vhost_discard_vq_desc(vq
, 1);
575 vhost_net_enable_vq(net
, vq
);
579 pr_debug("Truncated TX packet: "
580 " len %d != %zd\n", err
, len
);
582 vhost_add_used_and_signal(&net
->dev
, vq
, head
, 0);
584 vhost_zerocopy_signal_used(net
, vq
);
585 vhost_net_tx_packet(net
);
586 if (unlikely(total_len
>= VHOST_NET_WEIGHT
) ||
587 unlikely(++sent_pkts
>= VHOST_NET_PKT_WEIGHT
)) {
588 vhost_poll_queue(&vq
->poll
);
593 mutex_unlock(&vq
->mutex
);
596 static int peek_head_len(struct vhost_net_virtqueue
*rvq
, struct sock
*sk
)
598 struct sk_buff
*head
;
603 return vhost_net_buf_peek(rvq
);
605 spin_lock_irqsave(&sk
->sk_receive_queue
.lock
, flags
);
606 head
= skb_peek(&sk
->sk_receive_queue
);
609 if (skb_vlan_tag_present(head
))
613 spin_unlock_irqrestore(&sk
->sk_receive_queue
.lock
, flags
);
617 static int sk_has_rx_data(struct sock
*sk
)
619 struct socket
*sock
= sk
->sk_socket
;
621 if (sock
->ops
->peek_len
)
622 return sock
->ops
->peek_len(sock
);
624 return skb_queue_empty(&sk
->sk_receive_queue
);
627 static int vhost_net_rx_peek_head_len(struct vhost_net
*net
, struct sock
*sk
)
629 struct vhost_net_virtqueue
*rvq
= &net
->vqs
[VHOST_NET_VQ_RX
];
630 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_TX
];
631 struct vhost_virtqueue
*vq
= &nvq
->vq
;
632 unsigned long uninitialized_var(endtime
);
633 int len
= peek_head_len(rvq
, sk
);
635 if (!len
&& vq
->busyloop_timeout
) {
636 /* Both tx vq and rx socket were polled here */
637 mutex_lock_nested(&vq
->mutex
, 1);
638 vhost_disable_notify(&net
->dev
, vq
);
641 endtime
= busy_clock() + vq
->busyloop_timeout
;
643 while (vhost_can_busy_poll(endtime
) &&
644 !vhost_has_work(&net
->dev
) &&
645 !sk_has_rx_data(sk
) &&
646 vhost_vq_avail_empty(&net
->dev
, vq
))
651 if (!vhost_vq_avail_empty(&net
->dev
, vq
))
652 vhost_poll_queue(&vq
->poll
);
653 else if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
654 vhost_disable_notify(&net
->dev
, vq
);
655 vhost_poll_queue(&vq
->poll
);
658 mutex_unlock(&vq
->mutex
);
660 len
= peek_head_len(rvq
, sk
);
666 /* This is a multi-buffer version of vhost_get_desc, that works if
667 * vq has read descriptors only.
668 * @vq - the relevant virtqueue
669 * @datalen - data length we'll be reading
670 * @iovcount - returned count of io vectors we fill
672 * @log_num - log offset
673 * @quota - headcount quota, 1 for big buffer
674 * returns number of buffer heads allocated, negative on error
676 static int get_rx_bufs(struct vhost_virtqueue
*vq
,
677 struct vring_used_elem
*heads
,
680 struct vhost_log
*log
,
684 unsigned int out
, in
;
689 /* len is always initialized before use since we are always called with
692 u32
uninitialized_var(len
);
694 while (datalen
> 0 && headcount
< quota
) {
695 if (unlikely(seg
>= UIO_MAXIOV
)) {
699 r
= vhost_get_vq_desc(vq
, vq
->iov
+ seg
,
700 ARRAY_SIZE(vq
->iov
) - seg
, &out
,
710 if (unlikely(out
|| in
<= 0)) {
711 vq_err(vq
, "unexpected descriptor format for RX: "
712 "out %d, in %d\n", out
, in
);
720 heads
[headcount
].id
= cpu_to_vhost32(vq
, d
);
721 len
= iov_length(vq
->iov
+ seg
, in
);
722 heads
[headcount
].len
= cpu_to_vhost32(vq
, len
);
727 heads
[headcount
- 1].len
= cpu_to_vhost32(vq
, len
+ datalen
);
733 if (unlikely(datalen
> 0)) {
739 vhost_discard_vq_desc(vq
, headcount
);
743 /* Expects to be always run from workqueue - which acts as
744 * read-size critical section for our kind of RCU. */
745 static void handle_rx(struct vhost_net
*net
)
747 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_RX
];
748 struct vhost_virtqueue
*vq
= &nvq
->vq
;
749 unsigned uninitialized_var(in
), log
;
750 struct vhost_log
*vq_log
;
751 struct msghdr msg
= {
754 .msg_control
= NULL
, /* FIXME: get and handle RX aux data. */
756 .msg_flags
= MSG_DONTWAIT
,
758 struct virtio_net_hdr hdr
= {
760 .gso_type
= VIRTIO_NET_HDR_GSO_NONE
762 size_t total_len
= 0;
765 size_t vhost_hlen
, sock_hlen
;
766 size_t vhost_len
, sock_len
;
768 struct iov_iter fixup
;
769 __virtio16 num_buffers
;
772 mutex_lock_nested(&vq
->mutex
, 0);
773 sock
= vq
->private_data
;
777 if (!vq_iotlb_prefetch(vq
))
780 vhost_disable_notify(&net
->dev
, vq
);
781 vhost_net_disable_vq(net
, vq
);
783 vhost_hlen
= nvq
->vhost_hlen
;
784 sock_hlen
= nvq
->sock_hlen
;
786 vq_log
= unlikely(vhost_has_feature(vq
, VHOST_F_LOG_ALL
)) ?
788 mergeable
= vhost_has_feature(vq
, VIRTIO_NET_F_MRG_RXBUF
);
790 while ((sock_len
= vhost_net_rx_peek_head_len(net
, sock
->sk
))) {
791 sock_len
+= sock_hlen
;
792 vhost_len
= sock_len
+ vhost_hlen
;
793 headcount
= get_rx_bufs(vq
, vq
->heads
, vhost_len
,
795 likely(mergeable
) ? UIO_MAXIOV
: 1);
796 /* On error, stop handling until the next kick. */
797 if (unlikely(headcount
< 0))
799 /* OK, now we need to know about added descriptors. */
801 if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
802 /* They have slipped one in as we were
803 * doing that: check again. */
804 vhost_disable_notify(&net
->dev
, vq
);
807 /* Nothing new? Wait for eventfd to tell us
812 msg
.msg_control
= vhost_net_buf_consume(&nvq
->rxq
);
813 /* On overrun, truncate and discard */
814 if (unlikely(headcount
> UIO_MAXIOV
)) {
815 iov_iter_init(&msg
.msg_iter
, READ
, vq
->iov
, 1, 1);
816 err
= sock
->ops
->recvmsg(sock
, &msg
,
817 1, MSG_DONTWAIT
| MSG_TRUNC
);
818 pr_debug("Discarded rx packet: len %zd\n", sock_len
);
821 /* We don't need to be notified again. */
822 iov_iter_init(&msg
.msg_iter
, READ
, vq
->iov
, in
, vhost_len
);
823 fixup
= msg
.msg_iter
;
824 if (unlikely((vhost_hlen
))) {
825 /* We will supply the header ourselves
828 iov_iter_advance(&msg
.msg_iter
, vhost_hlen
);
830 err
= sock
->ops
->recvmsg(sock
, &msg
,
831 sock_len
, MSG_DONTWAIT
| MSG_TRUNC
);
832 /* Userspace might have consumed the packet meanwhile:
833 * it's not supposed to do this usually, but might be hard
834 * to prevent. Discard data we got (if any) and keep going. */
835 if (unlikely(err
!= sock_len
)) {
836 pr_debug("Discarded rx packet: "
837 " len %d, expected %zd\n", err
, sock_len
);
838 vhost_discard_vq_desc(vq
, headcount
);
841 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
842 if (unlikely(vhost_hlen
)) {
843 if (copy_to_iter(&hdr
, sizeof(hdr
),
844 &fixup
) != sizeof(hdr
)) {
845 vq_err(vq
, "Unable to write vnet_hdr "
846 "at addr %p\n", vq
->iov
->iov_base
);
850 /* Header came from socket; we'll need to patch
851 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
853 iov_iter_advance(&fixup
, sizeof(hdr
));
855 /* TODO: Should check and handle checksum. */
857 num_buffers
= cpu_to_vhost16(vq
, headcount
);
858 if (likely(mergeable
) &&
859 copy_to_iter(&num_buffers
, sizeof num_buffers
,
860 &fixup
) != sizeof num_buffers
) {
861 vq_err(vq
, "Failed num_buffers write");
862 vhost_discard_vq_desc(vq
, headcount
);
865 vhost_add_used_and_signal_n(&net
->dev
, vq
, vq
->heads
,
867 if (unlikely(vq_log
))
868 vhost_log_write(vq
, vq_log
, log
, vhost_len
,
870 total_len
+= vhost_len
;
871 if (unlikely(total_len
>= VHOST_NET_WEIGHT
) ||
872 unlikely(++recv_pkts
>= VHOST_NET_PKT_WEIGHT
)) {
873 vhost_poll_queue(&vq
->poll
);
877 vhost_net_enable_vq(net
, vq
);
879 mutex_unlock(&vq
->mutex
);
882 static void handle_tx_kick(struct vhost_work
*work
)
884 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
886 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
891 static void handle_rx_kick(struct vhost_work
*work
)
893 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
895 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
900 static void handle_tx_net(struct vhost_work
*work
)
902 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
903 poll
[VHOST_NET_VQ_TX
].work
);
907 static void handle_rx_net(struct vhost_work
*work
)
909 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
910 poll
[VHOST_NET_VQ_RX
].work
);
914 static int vhost_net_open(struct inode
*inode
, struct file
*f
)
917 struct vhost_dev
*dev
;
918 struct vhost_virtqueue
**vqs
;
919 struct sk_buff
**queue
;
922 n
= kvmalloc(sizeof *n
, GFP_KERNEL
| __GFP_RETRY_MAYFAIL
);
925 vqs
= kmalloc(VHOST_NET_VQ_MAX
* sizeof(*vqs
), GFP_KERNEL
);
931 queue
= kmalloc_array(VHOST_RX_BATCH
, sizeof(struct sk_buff
*),
938 n
->vqs
[VHOST_NET_VQ_RX
].rxq
.queue
= queue
;
941 vqs
[VHOST_NET_VQ_TX
] = &n
->vqs
[VHOST_NET_VQ_TX
].vq
;
942 vqs
[VHOST_NET_VQ_RX
] = &n
->vqs
[VHOST_NET_VQ_RX
].vq
;
943 n
->vqs
[VHOST_NET_VQ_TX
].vq
.handle_kick
= handle_tx_kick
;
944 n
->vqs
[VHOST_NET_VQ_RX
].vq
.handle_kick
= handle_rx_kick
;
945 for (i
= 0; i
< VHOST_NET_VQ_MAX
; i
++) {
946 n
->vqs
[i
].ubufs
= NULL
;
947 n
->vqs
[i
].ubuf_info
= NULL
;
948 n
->vqs
[i
].upend_idx
= 0;
949 n
->vqs
[i
].done_idx
= 0;
950 n
->vqs
[i
].vhost_hlen
= 0;
951 n
->vqs
[i
].sock_hlen
= 0;
952 vhost_net_buf_init(&n
->vqs
[i
].rxq
);
954 vhost_dev_init(dev
, vqs
, VHOST_NET_VQ_MAX
);
956 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_TX
, handle_tx_net
, POLLOUT
, dev
);
957 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_RX
, handle_rx_net
, POLLIN
, dev
);
964 static struct socket
*vhost_net_stop_vq(struct vhost_net
*n
,
965 struct vhost_virtqueue
*vq
)
968 struct vhost_net_virtqueue
*nvq
=
969 container_of(vq
, struct vhost_net_virtqueue
, vq
);
971 mutex_lock(&vq
->mutex
);
972 sock
= vq
->private_data
;
973 vhost_net_disable_vq(n
, vq
);
974 vq
->private_data
= NULL
;
975 vhost_net_buf_unproduce(nvq
);
976 mutex_unlock(&vq
->mutex
);
980 static void vhost_net_stop(struct vhost_net
*n
, struct socket
**tx_sock
,
981 struct socket
**rx_sock
)
983 *tx_sock
= vhost_net_stop_vq(n
, &n
->vqs
[VHOST_NET_VQ_TX
].vq
);
984 *rx_sock
= vhost_net_stop_vq(n
, &n
->vqs
[VHOST_NET_VQ_RX
].vq
);
987 static void vhost_net_flush_vq(struct vhost_net
*n
, int index
)
989 vhost_poll_flush(n
->poll
+ index
);
990 vhost_poll_flush(&n
->vqs
[index
].vq
.poll
);
993 static void vhost_net_flush(struct vhost_net
*n
)
995 vhost_net_flush_vq(n
, VHOST_NET_VQ_TX
);
996 vhost_net_flush_vq(n
, VHOST_NET_VQ_RX
);
997 if (n
->vqs
[VHOST_NET_VQ_TX
].ubufs
) {
998 mutex_lock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
1000 mutex_unlock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
1001 /* Wait for all lower device DMAs done. */
1002 vhost_net_ubuf_put_and_wait(n
->vqs
[VHOST_NET_VQ_TX
].ubufs
);
1003 mutex_lock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
1004 n
->tx_flush
= false;
1005 atomic_set(&n
->vqs
[VHOST_NET_VQ_TX
].ubufs
->refcount
, 1);
1006 mutex_unlock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
1010 static int vhost_net_release(struct inode
*inode
, struct file
*f
)
1012 struct vhost_net
*n
= f
->private_data
;
1013 struct socket
*tx_sock
;
1014 struct socket
*rx_sock
;
1016 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
1018 vhost_dev_stop(&n
->dev
);
1019 vhost_dev_cleanup(&n
->dev
, false);
1020 vhost_net_vq_reset(n
);
1022 sockfd_put(tx_sock
);
1024 sockfd_put(rx_sock
);
1025 /* Make sure no callbacks are outstanding */
1026 synchronize_rcu_bh();
1027 /* We do an extra flush before freeing memory,
1028 * since jobs can re-queue themselves. */
1030 kfree(n
->vqs
[VHOST_NET_VQ_RX
].rxq
.queue
);
1036 static struct socket
*get_raw_socket(int fd
)
1039 struct sockaddr_ll sa
;
1040 char buf
[MAX_ADDR_LEN
];
1042 int uaddr_len
= sizeof uaddr
, r
;
1043 struct socket
*sock
= sockfd_lookup(fd
, &r
);
1046 return ERR_PTR(-ENOTSOCK
);
1048 /* Parameter checking */
1049 if (sock
->sk
->sk_type
!= SOCK_RAW
) {
1050 r
= -ESOCKTNOSUPPORT
;
1054 r
= sock
->ops
->getname(sock
, (struct sockaddr
*)&uaddr
.sa
,
1059 if (uaddr
.sa
.sll_family
!= AF_PACKET
) {
1069 static struct skb_array
*get_tap_skb_array(int fd
)
1071 struct skb_array
*array
;
1072 struct file
*file
= fget(fd
);
1076 array
= tun_get_skb_array(file
);
1079 array
= tap_get_skb_array(file
);
1088 static struct socket
*get_tap_socket(int fd
)
1090 struct file
*file
= fget(fd
);
1091 struct socket
*sock
;
1094 return ERR_PTR(-EBADF
);
1095 sock
= tun_get_socket(file
);
1098 sock
= tap_get_socket(file
);
1104 static struct socket
*get_socket(int fd
)
1106 struct socket
*sock
;
1108 /* special case to disable backend */
1111 sock
= get_raw_socket(fd
);
1114 sock
= get_tap_socket(fd
);
1117 return ERR_PTR(-ENOTSOCK
);
1120 static long vhost_net_set_backend(struct vhost_net
*n
, unsigned index
, int fd
)
1122 struct socket
*sock
, *oldsock
;
1123 struct vhost_virtqueue
*vq
;
1124 struct vhost_net_virtqueue
*nvq
;
1125 struct vhost_net_ubuf_ref
*ubufs
, *oldubufs
= NULL
;
1128 mutex_lock(&n
->dev
.mutex
);
1129 r
= vhost_dev_check_owner(&n
->dev
);
1133 if (index
>= VHOST_NET_VQ_MAX
) {
1137 vq
= &n
->vqs
[index
].vq
;
1138 nvq
= &n
->vqs
[index
];
1139 mutex_lock(&vq
->mutex
);
1141 /* Verify that ring has been setup correctly. */
1142 if (!vhost_vq_access_ok(vq
)) {
1146 sock
= get_socket(fd
);
1152 /* start polling new socket */
1153 oldsock
= vq
->private_data
;
1154 if (sock
!= oldsock
) {
1155 ubufs
= vhost_net_ubuf_alloc(vq
,
1156 sock
&& vhost_sock_zcopy(sock
));
1157 if (IS_ERR(ubufs
)) {
1162 vhost_net_disable_vq(n
, vq
);
1163 vq
->private_data
= sock
;
1164 vhost_net_buf_unproduce(nvq
);
1165 if (index
== VHOST_NET_VQ_RX
)
1166 nvq
->rx_array
= get_tap_skb_array(fd
);
1167 r
= vhost_vq_init_access(vq
);
1170 r
= vhost_net_enable_vq(n
, vq
);
1174 oldubufs
= nvq
->ubufs
;
1178 n
->tx_zcopy_err
= 0;
1179 n
->tx_flush
= false;
1182 mutex_unlock(&vq
->mutex
);
1185 vhost_net_ubuf_put_wait_and_free(oldubufs
);
1186 mutex_lock(&vq
->mutex
);
1187 vhost_zerocopy_signal_used(n
, vq
);
1188 mutex_unlock(&vq
->mutex
);
1192 vhost_net_flush_vq(n
, index
);
1193 sockfd_put(oldsock
);
1196 mutex_unlock(&n
->dev
.mutex
);
1200 vq
->private_data
= oldsock
;
1201 vhost_net_enable_vq(n
, vq
);
1203 vhost_net_ubuf_put_wait_and_free(ubufs
);
1208 mutex_unlock(&vq
->mutex
);
1210 mutex_unlock(&n
->dev
.mutex
);
1214 static long vhost_net_reset_owner(struct vhost_net
*n
)
1216 struct socket
*tx_sock
= NULL
;
1217 struct socket
*rx_sock
= NULL
;
1219 struct vhost_umem
*umem
;
1221 mutex_lock(&n
->dev
.mutex
);
1222 err
= vhost_dev_check_owner(&n
->dev
);
1225 umem
= vhost_dev_reset_owner_prepare();
1230 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
1232 vhost_dev_stop(&n
->dev
);
1233 vhost_dev_reset_owner(&n
->dev
, umem
);
1234 vhost_net_vq_reset(n
);
1236 mutex_unlock(&n
->dev
.mutex
);
1238 sockfd_put(tx_sock
);
1240 sockfd_put(rx_sock
);
1244 static int vhost_net_set_features(struct vhost_net
*n
, u64 features
)
1246 size_t vhost_hlen
, sock_hlen
, hdr_len
;
1249 hdr_len
= (features
& ((1ULL << VIRTIO_NET_F_MRG_RXBUF
) |
1250 (1ULL << VIRTIO_F_VERSION_1
))) ?
1251 sizeof(struct virtio_net_hdr_mrg_rxbuf
) :
1252 sizeof(struct virtio_net_hdr
);
1253 if (features
& (1 << VHOST_NET_F_VIRTIO_NET_HDR
)) {
1254 /* vhost provides vnet_hdr */
1255 vhost_hlen
= hdr_len
;
1258 /* socket provides vnet_hdr */
1260 sock_hlen
= hdr_len
;
1262 mutex_lock(&n
->dev
.mutex
);
1263 if ((features
& (1 << VHOST_F_LOG_ALL
)) &&
1264 !vhost_log_access_ok(&n
->dev
))
1267 if ((features
& (1ULL << VIRTIO_F_IOMMU_PLATFORM
))) {
1268 if (vhost_init_device_iotlb(&n
->dev
, true))
1272 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
1273 mutex_lock(&n
->vqs
[i
].vq
.mutex
);
1274 n
->vqs
[i
].vq
.acked_features
= features
;
1275 n
->vqs
[i
].vhost_hlen
= vhost_hlen
;
1276 n
->vqs
[i
].sock_hlen
= sock_hlen
;
1277 mutex_unlock(&n
->vqs
[i
].vq
.mutex
);
1279 mutex_unlock(&n
->dev
.mutex
);
1283 mutex_unlock(&n
->dev
.mutex
);
1287 static long vhost_net_set_owner(struct vhost_net
*n
)
1291 mutex_lock(&n
->dev
.mutex
);
1292 if (vhost_dev_has_owner(&n
->dev
)) {
1296 r
= vhost_net_set_ubuf_info(n
);
1299 r
= vhost_dev_set_owner(&n
->dev
);
1301 vhost_net_clear_ubuf_info(n
);
1304 mutex_unlock(&n
->dev
.mutex
);
1308 static long vhost_net_ioctl(struct file
*f
, unsigned int ioctl
,
1311 struct vhost_net
*n
= f
->private_data
;
1312 void __user
*argp
= (void __user
*)arg
;
1313 u64 __user
*featurep
= argp
;
1314 struct vhost_vring_file backend
;
1319 case VHOST_NET_SET_BACKEND
:
1320 if (copy_from_user(&backend
, argp
, sizeof backend
))
1322 return vhost_net_set_backend(n
, backend
.index
, backend
.fd
);
1323 case VHOST_GET_FEATURES
:
1324 features
= VHOST_NET_FEATURES
;
1325 if (copy_to_user(featurep
, &features
, sizeof features
))
1328 case VHOST_SET_FEATURES
:
1329 if (copy_from_user(&features
, featurep
, sizeof features
))
1331 if (features
& ~VHOST_NET_FEATURES
)
1333 return vhost_net_set_features(n
, features
);
1334 case VHOST_RESET_OWNER
:
1335 return vhost_net_reset_owner(n
);
1336 case VHOST_SET_OWNER
:
1337 return vhost_net_set_owner(n
);
1339 mutex_lock(&n
->dev
.mutex
);
1340 r
= vhost_dev_ioctl(&n
->dev
, ioctl
, argp
);
1341 if (r
== -ENOIOCTLCMD
)
1342 r
= vhost_vring_ioctl(&n
->dev
, ioctl
, argp
);
1345 mutex_unlock(&n
->dev
.mutex
);
1350 #ifdef CONFIG_COMPAT
1351 static long vhost_net_compat_ioctl(struct file
*f
, unsigned int ioctl
,
1354 return vhost_net_ioctl(f
, ioctl
, (unsigned long)compat_ptr(arg
));
1358 static ssize_t
vhost_net_chr_read_iter(struct kiocb
*iocb
, struct iov_iter
*to
)
1360 struct file
*file
= iocb
->ki_filp
;
1361 struct vhost_net
*n
= file
->private_data
;
1362 struct vhost_dev
*dev
= &n
->dev
;
1363 int noblock
= file
->f_flags
& O_NONBLOCK
;
1365 return vhost_chr_read_iter(dev
, to
, noblock
);
1368 static ssize_t
vhost_net_chr_write_iter(struct kiocb
*iocb
,
1369 struct iov_iter
*from
)
1371 struct file
*file
= iocb
->ki_filp
;
1372 struct vhost_net
*n
= file
->private_data
;
1373 struct vhost_dev
*dev
= &n
->dev
;
1375 return vhost_chr_write_iter(dev
, from
);
1378 static unsigned int vhost_net_chr_poll(struct file
*file
, poll_table
*wait
)
1380 struct vhost_net
*n
= file
->private_data
;
1381 struct vhost_dev
*dev
= &n
->dev
;
1383 return vhost_chr_poll(file
, dev
, wait
);
1386 static const struct file_operations vhost_net_fops
= {
1387 .owner
= THIS_MODULE
,
1388 .release
= vhost_net_release
,
1389 .read_iter
= vhost_net_chr_read_iter
,
1390 .write_iter
= vhost_net_chr_write_iter
,
1391 .poll
= vhost_net_chr_poll
,
1392 .unlocked_ioctl
= vhost_net_ioctl
,
1393 #ifdef CONFIG_COMPAT
1394 .compat_ioctl
= vhost_net_compat_ioctl
,
1396 .open
= vhost_net_open
,
1397 .llseek
= noop_llseek
,
1400 static struct miscdevice vhost_net_misc
= {
1401 .minor
= VHOST_NET_MINOR
,
1402 .name
= "vhost-net",
1403 .fops
= &vhost_net_fops
,
1406 static int vhost_net_init(void)
1408 if (experimental_zcopytx
)
1409 vhost_net_enable_zcopy(VHOST_NET_VQ_TX
);
1410 return misc_register(&vhost_net_misc
);
1412 module_init(vhost_net_init
);
1414 static void vhost_net_exit(void)
1416 misc_deregister(&vhost_net_misc
);
1418 module_exit(vhost_net_exit
);
1420 MODULE_VERSION("0.0.1");
1421 MODULE_LICENSE("GPL v2");
1422 MODULE_AUTHOR("Michael S. Tsirkin");
1423 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1424 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR
);
1425 MODULE_ALIAS("devname:vhost-net");