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 TX used buffers for outstanding zerocopy */
48 #define VHOST_MAX_PEND 128
49 #define VHOST_GOODCOPY_LEN 256
52 * For transmit, used buffer len is unused; we override it to track buffer
53 * status internally; used for zerocopy tx only.
55 /* Lower device DMA failed */
56 #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3)
57 /* Lower device DMA done */
58 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
59 /* Lower device DMA in progress */
60 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
62 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
64 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
67 VHOST_NET_FEATURES
= VHOST_FEATURES
|
68 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR
) |
69 (1ULL << VIRTIO_NET_F_MRG_RXBUF
) |
70 (1ULL << VIRTIO_F_IOMMU_PLATFORM
)
79 struct vhost_net_ubuf_ref
{
80 /* refcount follows semantics similar to kref:
81 * 0: object is released
82 * 1: no outstanding ubufs
83 * >1: outstanding ubufs
86 wait_queue_head_t wait
;
87 struct vhost_virtqueue
*vq
;
90 #define VHOST_RX_BATCH 64
91 struct vhost_net_buf
{
92 struct sk_buff
**queue
;
97 struct vhost_net_virtqueue
{
98 struct vhost_virtqueue vq
;
101 /* vhost zerocopy support fields below: */
102 /* last used idx for outstanding DMA zerocopy buffers */
104 /* first used idx for DMA done zerocopy buffers */
106 /* an array of userspace buffers info */
107 struct ubuf_info
*ubuf_info
;
108 /* Reference counting for outstanding ubufs.
109 * Protected by vq mutex. Writers must also take device mutex. */
110 struct vhost_net_ubuf_ref
*ubufs
;
111 struct skb_array
*rx_array
;
112 struct vhost_net_buf rxq
;
116 struct vhost_dev dev
;
117 struct vhost_net_virtqueue vqs
[VHOST_NET_VQ_MAX
];
118 struct vhost_poll poll
[VHOST_NET_VQ_MAX
];
119 /* Number of TX recently submitted.
120 * Protected by tx vq lock. */
122 /* Number of times zerocopy TX recently failed.
123 * Protected by tx vq lock. */
124 unsigned tx_zcopy_err
;
125 /* Flush in progress. Protected by tx vq lock. */
129 static unsigned vhost_net_zcopy_mask __read_mostly
;
131 static void *vhost_net_buf_get_ptr(struct vhost_net_buf
*rxq
)
133 if (rxq
->tail
!= rxq
->head
)
134 return rxq
->queue
[rxq
->head
];
139 static int vhost_net_buf_get_size(struct vhost_net_buf
*rxq
)
141 return rxq
->tail
- rxq
->head
;
144 static int vhost_net_buf_is_empty(struct vhost_net_buf
*rxq
)
146 return rxq
->tail
== rxq
->head
;
149 static void *vhost_net_buf_consume(struct vhost_net_buf
*rxq
)
151 void *ret
= vhost_net_buf_get_ptr(rxq
);
156 static int vhost_net_buf_produce(struct vhost_net_virtqueue
*nvq
)
158 struct vhost_net_buf
*rxq
= &nvq
->rxq
;
161 rxq
->tail
= skb_array_consume_batched(nvq
->rx_array
, rxq
->queue
,
166 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue
*nvq
)
168 struct vhost_net_buf
*rxq
= &nvq
->rxq
;
170 if (nvq
->rx_array
&& !vhost_net_buf_is_empty(rxq
)) {
171 skb_array_unconsume(nvq
->rx_array
, rxq
->queue
+ rxq
->head
,
172 vhost_net_buf_get_size(rxq
));
173 rxq
->head
= rxq
->tail
= 0;
177 static int vhost_net_buf_peek(struct vhost_net_virtqueue
*nvq
)
179 struct vhost_net_buf
*rxq
= &nvq
->rxq
;
181 if (!vhost_net_buf_is_empty(rxq
))
184 if (!vhost_net_buf_produce(nvq
))
188 return __skb_array_len_with_tag(vhost_net_buf_get_ptr(rxq
));
191 static void vhost_net_buf_init(struct vhost_net_buf
*rxq
)
193 rxq
->head
= rxq
->tail
= 0;
196 static void vhost_net_enable_zcopy(int vq
)
198 vhost_net_zcopy_mask
|= 0x1 << vq
;
201 static struct vhost_net_ubuf_ref
*
202 vhost_net_ubuf_alloc(struct vhost_virtqueue
*vq
, bool zcopy
)
204 struct vhost_net_ubuf_ref
*ubufs
;
205 /* No zero copy backend? Nothing to count. */
208 ubufs
= kmalloc(sizeof(*ubufs
), GFP_KERNEL
);
210 return ERR_PTR(-ENOMEM
);
211 atomic_set(&ubufs
->refcount
, 1);
212 init_waitqueue_head(&ubufs
->wait
);
217 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref
*ubufs
)
219 int r
= atomic_sub_return(1, &ubufs
->refcount
);
221 wake_up(&ubufs
->wait
);
225 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref
*ubufs
)
227 vhost_net_ubuf_put(ubufs
);
228 wait_event(ubufs
->wait
, !atomic_read(&ubufs
->refcount
));
231 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref
*ubufs
)
233 vhost_net_ubuf_put_and_wait(ubufs
);
237 static void vhost_net_clear_ubuf_info(struct vhost_net
*n
)
241 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
242 kfree(n
->vqs
[i
].ubuf_info
);
243 n
->vqs
[i
].ubuf_info
= NULL
;
247 static int vhost_net_set_ubuf_info(struct vhost_net
*n
)
252 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
253 zcopy
= vhost_net_zcopy_mask
& (0x1 << i
);
256 n
->vqs
[i
].ubuf_info
= kmalloc(sizeof(*n
->vqs
[i
].ubuf_info
) *
257 UIO_MAXIOV
, GFP_KERNEL
);
258 if (!n
->vqs
[i
].ubuf_info
)
264 vhost_net_clear_ubuf_info(n
);
268 static void vhost_net_vq_reset(struct vhost_net
*n
)
272 vhost_net_clear_ubuf_info(n
);
274 for (i
= 0; i
< VHOST_NET_VQ_MAX
; i
++) {
275 n
->vqs
[i
].done_idx
= 0;
276 n
->vqs
[i
].upend_idx
= 0;
277 n
->vqs
[i
].ubufs
= NULL
;
278 n
->vqs
[i
].vhost_hlen
= 0;
279 n
->vqs
[i
].sock_hlen
= 0;
280 vhost_net_buf_init(&n
->vqs
[i
].rxq
);
285 static void vhost_net_tx_packet(struct vhost_net
*net
)
288 if (net
->tx_packets
< 1024)
291 net
->tx_zcopy_err
= 0;
294 static void vhost_net_tx_err(struct vhost_net
*net
)
299 static bool vhost_net_tx_select_zcopy(struct vhost_net
*net
)
301 /* TX flush waits for outstanding DMAs to be done.
302 * Don't start new DMAs.
304 return !net
->tx_flush
&&
305 net
->tx_packets
/ 64 >= net
->tx_zcopy_err
;
308 static bool vhost_sock_zcopy(struct socket
*sock
)
310 return unlikely(experimental_zcopytx
) &&
311 sock_flag(sock
->sk
, SOCK_ZEROCOPY
);
314 /* In case of DMA done not in order in lower device driver for some reason.
315 * upend_idx is used to track end of used idx, done_idx is used to track head
316 * of used idx. Once lower device DMA done contiguously, we will signal KVM
319 static void vhost_zerocopy_signal_used(struct vhost_net
*net
,
320 struct vhost_virtqueue
*vq
)
322 struct vhost_net_virtqueue
*nvq
=
323 container_of(vq
, struct vhost_net_virtqueue
, vq
);
327 for (i
= nvq
->done_idx
; i
!= nvq
->upend_idx
; i
= (i
+ 1) % UIO_MAXIOV
) {
328 if (vq
->heads
[i
].len
== VHOST_DMA_FAILED_LEN
)
329 vhost_net_tx_err(net
);
330 if (VHOST_DMA_IS_DONE(vq
->heads
[i
].len
)) {
331 vq
->heads
[i
].len
= VHOST_DMA_CLEAR_LEN
;
337 add
= min(UIO_MAXIOV
- nvq
->done_idx
, j
);
338 vhost_add_used_and_signal_n(vq
->dev
, vq
,
339 &vq
->heads
[nvq
->done_idx
], add
);
340 nvq
->done_idx
= (nvq
->done_idx
+ add
) % UIO_MAXIOV
;
345 static void vhost_zerocopy_callback(struct ubuf_info
*ubuf
, bool success
)
347 struct vhost_net_ubuf_ref
*ubufs
= ubuf
->ctx
;
348 struct vhost_virtqueue
*vq
= ubufs
->vq
;
353 /* set len to mark this desc buffers done DMA */
354 vq
->heads
[ubuf
->desc
].len
= success
?
355 VHOST_DMA_DONE_LEN
: VHOST_DMA_FAILED_LEN
;
356 cnt
= vhost_net_ubuf_put(ubufs
);
359 * Trigger polling thread if guest stopped submitting new buffers:
360 * in this case, the refcount after decrement will eventually reach 1.
361 * We also trigger polling periodically after each 16 packets
362 * (the value 16 here is more or less arbitrary, it's tuned to trigger
363 * less than 10% of times).
365 if (cnt
<= 1 || !(cnt
% 16))
366 vhost_poll_queue(&vq
->poll
);
368 rcu_read_unlock_bh();
371 static inline unsigned long busy_clock(void)
373 return local_clock() >> 10;
376 static bool vhost_can_busy_poll(unsigned long endtime
)
378 return likely(!need_resched() && !time_after(busy_clock(), endtime
) &&
379 !signal_pending(current
));
382 static void vhost_net_disable_vq(struct vhost_net
*n
,
383 struct vhost_virtqueue
*vq
)
385 struct vhost_net_virtqueue
*nvq
=
386 container_of(vq
, struct vhost_net_virtqueue
, vq
);
387 struct vhost_poll
*poll
= n
->poll
+ (nvq
- n
->vqs
);
388 if (!vq
->private_data
)
390 vhost_poll_stop(poll
);
393 static int vhost_net_enable_vq(struct vhost_net
*n
,
394 struct vhost_virtqueue
*vq
)
396 struct vhost_net_virtqueue
*nvq
=
397 container_of(vq
, struct vhost_net_virtqueue
, vq
);
398 struct vhost_poll
*poll
= n
->poll
+ (nvq
- n
->vqs
);
401 sock
= vq
->private_data
;
405 return vhost_poll_start(poll
, sock
->file
);
408 static int vhost_net_tx_get_vq_desc(struct vhost_net
*net
,
409 struct vhost_virtqueue
*vq
,
410 struct iovec iov
[], unsigned int iov_size
,
411 unsigned int *out_num
, unsigned int *in_num
,
414 unsigned long uninitialized_var(endtime
);
415 int r
= vhost_get_vq_desc(vq
, vq
->iov
, ARRAY_SIZE(vq
->iov
),
416 out_num
, in_num
, NULL
, NULL
);
418 if (r
== vq
->num
&& vq
->busyloop_timeout
) {
420 endtime
= busy_clock() + vq
->busyloop_timeout
;
421 while (vhost_can_busy_poll(endtime
)) {
422 if (vhost_has_work(vq
->dev
)) {
423 *busyloop_intr
= true;
426 if (!vhost_vq_avail_empty(vq
->dev
, vq
))
431 r
= vhost_get_vq_desc(vq
, vq
->iov
, ARRAY_SIZE(vq
->iov
),
432 out_num
, in_num
, NULL
, NULL
);
438 static bool vhost_exceeds_maxpend(struct vhost_net
*net
)
440 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_TX
];
441 struct vhost_virtqueue
*vq
= &nvq
->vq
;
443 return (nvq
->upend_idx
+ UIO_MAXIOV
- nvq
->done_idx
) % UIO_MAXIOV
>
444 min_t(unsigned int, VHOST_MAX_PEND
, vq
->num
>> 2);
447 /* Expects to be always run from workqueue - which acts as
448 * read-size critical section for our kind of RCU. */
449 static void handle_tx(struct vhost_net
*net
)
451 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_TX
];
452 struct vhost_virtqueue
*vq
= &nvq
->vq
;
455 struct msghdr msg
= {
460 .msg_flags
= MSG_DONTWAIT
,
462 size_t len
, total_len
= 0;
466 struct vhost_net_ubuf_ref
*uninitialized_var(ubufs
);
467 bool zcopy
, zcopy_used
;
469 mutex_lock(&vq
->mutex
);
470 sock
= vq
->private_data
;
474 if (!vq_iotlb_prefetch(vq
))
477 vhost_disable_notify(&net
->dev
, vq
);
478 vhost_net_disable_vq(net
, vq
);
480 hdr_size
= nvq
->vhost_hlen
;
486 /* Release DMAs done buffers first */
488 vhost_zerocopy_signal_used(net
, vq
);
490 busyloop_intr
= false;
491 head
= vhost_net_tx_get_vq_desc(net
, vq
, vq
->iov
,
493 &out
, &in
, &busyloop_intr
);
494 /* On error, stop handling until the next kick. */
495 if (unlikely(head
< 0))
497 /* Nothing new? Wait for eventfd to tell us they refilled. */
498 if (head
== vq
->num
) {
499 if (unlikely(busyloop_intr
)) {
500 vhost_poll_queue(&vq
->poll
);
501 } else if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
502 vhost_disable_notify(&net
->dev
, vq
);
508 vq_err(vq
, "Unexpected descriptor format for TX: "
509 "out %d, int %d\n", out
, in
);
512 /* Skip header. TODO: support TSO. */
513 len
= iov_length(vq
->iov
, out
);
514 iov_iter_init(&msg
.msg_iter
, WRITE
, vq
->iov
, out
, len
);
515 iov_iter_advance(&msg
.msg_iter
, hdr_size
);
517 if (!msg_data_left(&msg
)) {
518 vq_err(vq
, "Unexpected header len for TX: "
519 "%zd expected %zd\n",
523 len
= msg_data_left(&msg
);
525 zcopy_used
= zcopy
&& len
>= VHOST_GOODCOPY_LEN
526 && !vhost_exceeds_maxpend(net
)
527 && vhost_net_tx_select_zcopy(net
);
529 /* use msg_control to pass vhost zerocopy ubuf info to skb */
531 struct ubuf_info
*ubuf
;
532 ubuf
= nvq
->ubuf_info
+ nvq
->upend_idx
;
534 vq
->heads
[nvq
->upend_idx
].id
= cpu_to_vhost32(vq
, head
);
535 vq
->heads
[nvq
->upend_idx
].len
= VHOST_DMA_IN_PROGRESS
;
536 ubuf
->callback
= vhost_zerocopy_callback
;
537 ubuf
->ctx
= nvq
->ubufs
;
538 ubuf
->desc
= nvq
->upend_idx
;
539 refcount_set(&ubuf
->refcnt
, 1);
540 msg
.msg_control
= ubuf
;
541 msg
.msg_controllen
= sizeof(ubuf
);
543 atomic_inc(&ubufs
->refcount
);
544 nvq
->upend_idx
= (nvq
->upend_idx
+ 1) % UIO_MAXIOV
;
546 msg
.msg_control
= NULL
;
551 if (total_len
< VHOST_NET_WEIGHT
&&
552 !vhost_vq_avail_empty(&net
->dev
, vq
) &&
553 likely(!vhost_exceeds_maxpend(net
))) {
554 msg
.msg_flags
|= MSG_MORE
;
556 msg
.msg_flags
&= ~MSG_MORE
;
559 /* TODO: Check specific error and bomb out unless ENOBUFS? */
560 err
= sock
->ops
->sendmsg(sock
, &msg
, len
);
561 if (unlikely(err
< 0)) {
563 vhost_net_ubuf_put(ubufs
);
564 nvq
->upend_idx
= ((unsigned)nvq
->upend_idx
- 1)
567 vhost_discard_vq_desc(vq
, 1);
568 vhost_net_enable_vq(net
, vq
);
572 pr_debug("Truncated TX packet: "
573 " len %d != %zd\n", err
, len
);
575 vhost_add_used_and_signal(&net
->dev
, vq
, head
, 0);
577 vhost_zerocopy_signal_used(net
, vq
);
578 vhost_net_tx_packet(net
);
579 if (unlikely(total_len
>= VHOST_NET_WEIGHT
)) {
580 vhost_poll_queue(&vq
->poll
);
585 mutex_unlock(&vq
->mutex
);
588 static int peek_head_len(struct vhost_net_virtqueue
*rvq
, struct sock
*sk
)
590 struct sk_buff
*head
;
595 return vhost_net_buf_peek(rvq
);
597 spin_lock_irqsave(&sk
->sk_receive_queue
.lock
, flags
);
598 head
= skb_peek(&sk
->sk_receive_queue
);
601 if (skb_vlan_tag_present(head
))
605 spin_unlock_irqrestore(&sk
->sk_receive_queue
.lock
, flags
);
609 static int sk_has_rx_data(struct sock
*sk
)
611 struct socket
*sock
= sk
->sk_socket
;
613 if (sock
->ops
->peek_len
)
614 return sock
->ops
->peek_len(sock
);
616 return skb_queue_empty(&sk
->sk_receive_queue
);
619 static int vhost_net_rx_peek_head_len(struct vhost_net
*net
, struct sock
*sk
)
621 struct vhost_net_virtqueue
*rvq
= &net
->vqs
[VHOST_NET_VQ_RX
];
622 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_TX
];
623 struct vhost_virtqueue
*vq
= &nvq
->vq
;
624 unsigned long uninitialized_var(endtime
);
625 int len
= peek_head_len(rvq
, sk
);
627 if (!len
&& vq
->busyloop_timeout
) {
628 /* Both tx vq and rx socket were polled here */
629 mutex_lock_nested(&vq
->mutex
, 1);
630 vhost_disable_notify(&net
->dev
, vq
);
633 endtime
= busy_clock() + vq
->busyloop_timeout
;
635 while (vhost_can_busy_poll(endtime
) &&
636 !vhost_has_work(&net
->dev
) &&
637 !sk_has_rx_data(sk
) &&
638 vhost_vq_avail_empty(&net
->dev
, vq
))
643 if (!vhost_vq_avail_empty(&net
->dev
, vq
))
644 vhost_poll_queue(&vq
->poll
);
645 else if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
646 vhost_disable_notify(&net
->dev
, vq
);
647 vhost_poll_queue(&vq
->poll
);
650 mutex_unlock(&vq
->mutex
);
652 len
= peek_head_len(rvq
, sk
);
658 /* This is a multi-buffer version of vhost_get_desc, that works if
659 * vq has read descriptors only.
660 * @vq - the relevant virtqueue
661 * @datalen - data length we'll be reading
662 * @iovcount - returned count of io vectors we fill
664 * @log_num - log offset
665 * @quota - headcount quota, 1 for big buffer
666 * returns number of buffer heads allocated, negative on error
668 static int get_rx_bufs(struct vhost_virtqueue
*vq
,
669 struct vring_used_elem
*heads
,
672 struct vhost_log
*log
,
676 unsigned int out
, in
;
681 /* len is always initialized before use since we are always called with
684 u32
uninitialized_var(len
);
686 while (datalen
> 0 && headcount
< quota
) {
687 if (unlikely(seg
>= UIO_MAXIOV
)) {
691 r
= vhost_get_vq_desc(vq
, vq
->iov
+ seg
,
692 ARRAY_SIZE(vq
->iov
) - seg
, &out
,
702 if (unlikely(out
|| in
<= 0)) {
703 vq_err(vq
, "unexpected descriptor format for RX: "
704 "out %d, in %d\n", out
, in
);
712 heads
[headcount
].id
= cpu_to_vhost32(vq
, d
);
713 len
= iov_length(vq
->iov
+ seg
, in
);
714 heads
[headcount
].len
= cpu_to_vhost32(vq
, len
);
719 heads
[headcount
- 1].len
= cpu_to_vhost32(vq
, len
+ datalen
);
725 if (unlikely(datalen
> 0)) {
731 vhost_discard_vq_desc(vq
, headcount
);
735 /* Expects to be always run from workqueue - which acts as
736 * read-size critical section for our kind of RCU. */
737 static void handle_rx(struct vhost_net
*net
)
739 struct vhost_net_virtqueue
*nvq
= &net
->vqs
[VHOST_NET_VQ_RX
];
740 struct vhost_virtqueue
*vq
= &nvq
->vq
;
741 unsigned uninitialized_var(in
), log
;
742 struct vhost_log
*vq_log
;
743 struct msghdr msg
= {
746 .msg_control
= NULL
, /* FIXME: get and handle RX aux data. */
748 .msg_flags
= MSG_DONTWAIT
,
750 struct virtio_net_hdr hdr
= {
752 .gso_type
= VIRTIO_NET_HDR_GSO_NONE
754 size_t total_len
= 0;
757 size_t vhost_hlen
, sock_hlen
;
758 size_t vhost_len
, sock_len
;
760 struct iov_iter fixup
;
761 __virtio16 num_buffers
;
763 mutex_lock_nested(&vq
->mutex
, 0);
764 sock
= vq
->private_data
;
768 if (!vq_iotlb_prefetch(vq
))
771 vhost_disable_notify(&net
->dev
, vq
);
772 vhost_net_disable_vq(net
, vq
);
774 vhost_hlen
= nvq
->vhost_hlen
;
775 sock_hlen
= nvq
->sock_hlen
;
777 vq_log
= unlikely(vhost_has_feature(vq
, VHOST_F_LOG_ALL
)) ?
779 mergeable
= vhost_has_feature(vq
, VIRTIO_NET_F_MRG_RXBUF
);
781 while ((sock_len
= vhost_net_rx_peek_head_len(net
, sock
->sk
))) {
782 sock_len
+= sock_hlen
;
783 vhost_len
= sock_len
+ vhost_hlen
;
784 headcount
= get_rx_bufs(vq
, vq
->heads
, vhost_len
,
786 likely(mergeable
) ? UIO_MAXIOV
: 1);
787 /* On error, stop handling until the next kick. */
788 if (unlikely(headcount
< 0))
790 /* OK, now we need to know about added descriptors. */
792 if (unlikely(vhost_enable_notify(&net
->dev
, vq
))) {
793 /* They have slipped one in as we were
794 * doing that: check again. */
795 vhost_disable_notify(&net
->dev
, vq
);
798 /* Nothing new? Wait for eventfd to tell us
803 msg
.msg_control
= vhost_net_buf_consume(&nvq
->rxq
);
804 /* On overrun, truncate and discard */
805 if (unlikely(headcount
> UIO_MAXIOV
)) {
806 iov_iter_init(&msg
.msg_iter
, READ
, vq
->iov
, 1, 1);
807 err
= sock
->ops
->recvmsg(sock
, &msg
,
808 1, MSG_DONTWAIT
| MSG_TRUNC
);
809 pr_debug("Discarded rx packet: len %zd\n", sock_len
);
812 /* We don't need to be notified again. */
813 iov_iter_init(&msg
.msg_iter
, READ
, vq
->iov
, in
, vhost_len
);
814 fixup
= msg
.msg_iter
;
815 if (unlikely((vhost_hlen
))) {
816 /* We will supply the header ourselves
819 iov_iter_advance(&msg
.msg_iter
, vhost_hlen
);
821 err
= sock
->ops
->recvmsg(sock
, &msg
,
822 sock_len
, MSG_DONTWAIT
| MSG_TRUNC
);
823 /* Userspace might have consumed the packet meanwhile:
824 * it's not supposed to do this usually, but might be hard
825 * to prevent. Discard data we got (if any) and keep going. */
826 if (unlikely(err
!= sock_len
)) {
827 pr_debug("Discarded rx packet: "
828 " len %d, expected %zd\n", err
, sock_len
);
829 vhost_discard_vq_desc(vq
, headcount
);
832 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
833 if (unlikely(vhost_hlen
)) {
834 if (copy_to_iter(&hdr
, sizeof(hdr
),
835 &fixup
) != sizeof(hdr
)) {
836 vq_err(vq
, "Unable to write vnet_hdr "
837 "at addr %p\n", vq
->iov
->iov_base
);
841 /* Header came from socket; we'll need to patch
842 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
844 iov_iter_advance(&fixup
, sizeof(hdr
));
846 /* TODO: Should check and handle checksum. */
848 num_buffers
= cpu_to_vhost16(vq
, headcount
);
849 if (likely(mergeable
) &&
850 copy_to_iter(&num_buffers
, sizeof num_buffers
,
851 &fixup
) != sizeof num_buffers
) {
852 vq_err(vq
, "Failed num_buffers write");
853 vhost_discard_vq_desc(vq
, headcount
);
856 vhost_add_used_and_signal_n(&net
->dev
, vq
, vq
->heads
,
858 if (unlikely(vq_log
))
859 vhost_log_write(vq
, vq_log
, log
, vhost_len
,
861 total_len
+= vhost_len
;
862 if (unlikely(total_len
>= VHOST_NET_WEIGHT
)) {
863 vhost_poll_queue(&vq
->poll
);
867 vhost_net_enable_vq(net
, vq
);
869 mutex_unlock(&vq
->mutex
);
872 static void handle_tx_kick(struct vhost_work
*work
)
874 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
876 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
881 static void handle_rx_kick(struct vhost_work
*work
)
883 struct vhost_virtqueue
*vq
= container_of(work
, struct vhost_virtqueue
,
885 struct vhost_net
*net
= container_of(vq
->dev
, struct vhost_net
, dev
);
890 static void handle_tx_net(struct vhost_work
*work
)
892 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
893 poll
[VHOST_NET_VQ_TX
].work
);
897 static void handle_rx_net(struct vhost_work
*work
)
899 struct vhost_net
*net
= container_of(work
, struct vhost_net
,
900 poll
[VHOST_NET_VQ_RX
].work
);
904 static int vhost_net_open(struct inode
*inode
, struct file
*f
)
907 struct vhost_dev
*dev
;
908 struct vhost_virtqueue
**vqs
;
909 struct sk_buff
**queue
;
912 n
= kvmalloc(sizeof *n
, GFP_KERNEL
| __GFP_RETRY_MAYFAIL
);
915 vqs
= kmalloc(VHOST_NET_VQ_MAX
* sizeof(*vqs
), GFP_KERNEL
);
921 queue
= kmalloc_array(VHOST_RX_BATCH
, sizeof(struct sk_buff
*),
928 n
->vqs
[VHOST_NET_VQ_RX
].rxq
.queue
= queue
;
931 vqs
[VHOST_NET_VQ_TX
] = &n
->vqs
[VHOST_NET_VQ_TX
].vq
;
932 vqs
[VHOST_NET_VQ_RX
] = &n
->vqs
[VHOST_NET_VQ_RX
].vq
;
933 n
->vqs
[VHOST_NET_VQ_TX
].vq
.handle_kick
= handle_tx_kick
;
934 n
->vqs
[VHOST_NET_VQ_RX
].vq
.handle_kick
= handle_rx_kick
;
935 for (i
= 0; i
< VHOST_NET_VQ_MAX
; i
++) {
936 n
->vqs
[i
].ubufs
= NULL
;
937 n
->vqs
[i
].ubuf_info
= NULL
;
938 n
->vqs
[i
].upend_idx
= 0;
939 n
->vqs
[i
].done_idx
= 0;
940 n
->vqs
[i
].vhost_hlen
= 0;
941 n
->vqs
[i
].sock_hlen
= 0;
942 vhost_net_buf_init(&n
->vqs
[i
].rxq
);
944 vhost_dev_init(dev
, vqs
, VHOST_NET_VQ_MAX
);
946 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_TX
, handle_tx_net
, POLLOUT
, dev
);
947 vhost_poll_init(n
->poll
+ VHOST_NET_VQ_RX
, handle_rx_net
, POLLIN
, dev
);
954 static struct socket
*vhost_net_stop_vq(struct vhost_net
*n
,
955 struct vhost_virtqueue
*vq
)
958 struct vhost_net_virtqueue
*nvq
=
959 container_of(vq
, struct vhost_net_virtqueue
, vq
);
961 mutex_lock(&vq
->mutex
);
962 sock
= vq
->private_data
;
963 vhost_net_disable_vq(n
, vq
);
964 vq
->private_data
= NULL
;
965 vhost_net_buf_unproduce(nvq
);
966 mutex_unlock(&vq
->mutex
);
970 static void vhost_net_stop(struct vhost_net
*n
, struct socket
**tx_sock
,
971 struct socket
**rx_sock
)
973 *tx_sock
= vhost_net_stop_vq(n
, &n
->vqs
[VHOST_NET_VQ_TX
].vq
);
974 *rx_sock
= vhost_net_stop_vq(n
, &n
->vqs
[VHOST_NET_VQ_RX
].vq
);
977 static void vhost_net_flush_vq(struct vhost_net
*n
, int index
)
979 vhost_poll_flush(n
->poll
+ index
);
980 vhost_poll_flush(&n
->vqs
[index
].vq
.poll
);
983 static void vhost_net_flush(struct vhost_net
*n
)
985 vhost_net_flush_vq(n
, VHOST_NET_VQ_TX
);
986 vhost_net_flush_vq(n
, VHOST_NET_VQ_RX
);
987 if (n
->vqs
[VHOST_NET_VQ_TX
].ubufs
) {
988 mutex_lock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
990 mutex_unlock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
991 /* Wait for all lower device DMAs done. */
992 vhost_net_ubuf_put_and_wait(n
->vqs
[VHOST_NET_VQ_TX
].ubufs
);
993 mutex_lock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
995 atomic_set(&n
->vqs
[VHOST_NET_VQ_TX
].ubufs
->refcount
, 1);
996 mutex_unlock(&n
->vqs
[VHOST_NET_VQ_TX
].vq
.mutex
);
1000 static int vhost_net_release(struct inode
*inode
, struct file
*f
)
1002 struct vhost_net
*n
= f
->private_data
;
1003 struct socket
*tx_sock
;
1004 struct socket
*rx_sock
;
1006 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
1008 vhost_dev_stop(&n
->dev
);
1009 vhost_dev_cleanup(&n
->dev
, false);
1010 vhost_net_vq_reset(n
);
1012 sockfd_put(tx_sock
);
1014 sockfd_put(rx_sock
);
1015 /* Make sure no callbacks are outstanding */
1016 synchronize_rcu_bh();
1017 /* We do an extra flush before freeing memory,
1018 * since jobs can re-queue themselves. */
1020 kfree(n
->vqs
[VHOST_NET_VQ_RX
].rxq
.queue
);
1026 static struct socket
*get_raw_socket(int fd
)
1029 struct sockaddr_ll sa
;
1030 char buf
[MAX_ADDR_LEN
];
1032 int uaddr_len
= sizeof uaddr
, r
;
1033 struct socket
*sock
= sockfd_lookup(fd
, &r
);
1036 return ERR_PTR(-ENOTSOCK
);
1038 /* Parameter checking */
1039 if (sock
->sk
->sk_type
!= SOCK_RAW
) {
1040 r
= -ESOCKTNOSUPPORT
;
1044 r
= sock
->ops
->getname(sock
, (struct sockaddr
*)&uaddr
.sa
,
1049 if (uaddr
.sa
.sll_family
!= AF_PACKET
) {
1059 static struct skb_array
*get_tap_skb_array(int fd
)
1061 struct skb_array
*array
;
1062 struct file
*file
= fget(fd
);
1066 array
= tun_get_skb_array(file
);
1069 array
= tap_get_skb_array(file
);
1078 static struct socket
*get_tap_socket(int fd
)
1080 struct file
*file
= fget(fd
);
1081 struct socket
*sock
;
1084 return ERR_PTR(-EBADF
);
1085 sock
= tun_get_socket(file
);
1088 sock
= tap_get_socket(file
);
1094 static struct socket
*get_socket(int fd
)
1096 struct socket
*sock
;
1098 /* special case to disable backend */
1101 sock
= get_raw_socket(fd
);
1104 sock
= get_tap_socket(fd
);
1107 return ERR_PTR(-ENOTSOCK
);
1110 static long vhost_net_set_backend(struct vhost_net
*n
, unsigned index
, int fd
)
1112 struct socket
*sock
, *oldsock
;
1113 struct vhost_virtqueue
*vq
;
1114 struct vhost_net_virtqueue
*nvq
;
1115 struct vhost_net_ubuf_ref
*ubufs
, *oldubufs
= NULL
;
1118 mutex_lock(&n
->dev
.mutex
);
1119 r
= vhost_dev_check_owner(&n
->dev
);
1123 if (index
>= VHOST_NET_VQ_MAX
) {
1127 vq
= &n
->vqs
[index
].vq
;
1128 nvq
= &n
->vqs
[index
];
1129 mutex_lock(&vq
->mutex
);
1131 /* Verify that ring has been setup correctly. */
1132 if (!vhost_vq_access_ok(vq
)) {
1136 sock
= get_socket(fd
);
1142 /* start polling new socket */
1143 oldsock
= vq
->private_data
;
1144 if (sock
!= oldsock
) {
1145 ubufs
= vhost_net_ubuf_alloc(vq
,
1146 sock
&& vhost_sock_zcopy(sock
));
1147 if (IS_ERR(ubufs
)) {
1152 vhost_net_disable_vq(n
, vq
);
1153 vq
->private_data
= sock
;
1154 vhost_net_buf_unproduce(nvq
);
1155 if (index
== VHOST_NET_VQ_RX
)
1156 nvq
->rx_array
= get_tap_skb_array(fd
);
1157 r
= vhost_vq_init_access(vq
);
1160 r
= vhost_net_enable_vq(n
, vq
);
1164 oldubufs
= nvq
->ubufs
;
1168 n
->tx_zcopy_err
= 0;
1169 n
->tx_flush
= false;
1172 mutex_unlock(&vq
->mutex
);
1175 vhost_net_ubuf_put_wait_and_free(oldubufs
);
1176 mutex_lock(&vq
->mutex
);
1177 vhost_zerocopy_signal_used(n
, vq
);
1178 mutex_unlock(&vq
->mutex
);
1182 vhost_net_flush_vq(n
, index
);
1183 sockfd_put(oldsock
);
1186 mutex_unlock(&n
->dev
.mutex
);
1190 vq
->private_data
= oldsock
;
1191 vhost_net_enable_vq(n
, vq
);
1193 vhost_net_ubuf_put_wait_and_free(ubufs
);
1198 mutex_unlock(&vq
->mutex
);
1200 mutex_unlock(&n
->dev
.mutex
);
1204 static long vhost_net_reset_owner(struct vhost_net
*n
)
1206 struct socket
*tx_sock
= NULL
;
1207 struct socket
*rx_sock
= NULL
;
1209 struct vhost_umem
*umem
;
1211 mutex_lock(&n
->dev
.mutex
);
1212 err
= vhost_dev_check_owner(&n
->dev
);
1215 umem
= vhost_dev_reset_owner_prepare();
1220 vhost_net_stop(n
, &tx_sock
, &rx_sock
);
1222 vhost_dev_stop(&n
->dev
);
1223 vhost_dev_reset_owner(&n
->dev
, umem
);
1224 vhost_net_vq_reset(n
);
1226 mutex_unlock(&n
->dev
.mutex
);
1228 sockfd_put(tx_sock
);
1230 sockfd_put(rx_sock
);
1234 static int vhost_net_set_features(struct vhost_net
*n
, u64 features
)
1236 size_t vhost_hlen
, sock_hlen
, hdr_len
;
1239 hdr_len
= (features
& ((1ULL << VIRTIO_NET_F_MRG_RXBUF
) |
1240 (1ULL << VIRTIO_F_VERSION_1
))) ?
1241 sizeof(struct virtio_net_hdr_mrg_rxbuf
) :
1242 sizeof(struct virtio_net_hdr
);
1243 if (features
& (1 << VHOST_NET_F_VIRTIO_NET_HDR
)) {
1244 /* vhost provides vnet_hdr */
1245 vhost_hlen
= hdr_len
;
1248 /* socket provides vnet_hdr */
1250 sock_hlen
= hdr_len
;
1252 mutex_lock(&n
->dev
.mutex
);
1253 if ((features
& (1 << VHOST_F_LOG_ALL
)) &&
1254 !vhost_log_access_ok(&n
->dev
))
1257 if ((features
& (1ULL << VIRTIO_F_IOMMU_PLATFORM
))) {
1258 if (vhost_init_device_iotlb(&n
->dev
, true))
1262 for (i
= 0; i
< VHOST_NET_VQ_MAX
; ++i
) {
1263 mutex_lock(&n
->vqs
[i
].vq
.mutex
);
1264 n
->vqs
[i
].vq
.acked_features
= features
;
1265 n
->vqs
[i
].vhost_hlen
= vhost_hlen
;
1266 n
->vqs
[i
].sock_hlen
= sock_hlen
;
1267 mutex_unlock(&n
->vqs
[i
].vq
.mutex
);
1269 mutex_unlock(&n
->dev
.mutex
);
1273 mutex_unlock(&n
->dev
.mutex
);
1277 static long vhost_net_set_owner(struct vhost_net
*n
)
1281 mutex_lock(&n
->dev
.mutex
);
1282 if (vhost_dev_has_owner(&n
->dev
)) {
1286 r
= vhost_net_set_ubuf_info(n
);
1289 r
= vhost_dev_set_owner(&n
->dev
);
1291 vhost_net_clear_ubuf_info(n
);
1294 mutex_unlock(&n
->dev
.mutex
);
1298 static long vhost_net_ioctl(struct file
*f
, unsigned int ioctl
,
1301 struct vhost_net
*n
= f
->private_data
;
1302 void __user
*argp
= (void __user
*)arg
;
1303 u64 __user
*featurep
= argp
;
1304 struct vhost_vring_file backend
;
1309 case VHOST_NET_SET_BACKEND
:
1310 if (copy_from_user(&backend
, argp
, sizeof backend
))
1312 return vhost_net_set_backend(n
, backend
.index
, backend
.fd
);
1313 case VHOST_GET_FEATURES
:
1314 features
= VHOST_NET_FEATURES
;
1315 if (copy_to_user(featurep
, &features
, sizeof features
))
1318 case VHOST_SET_FEATURES
:
1319 if (copy_from_user(&features
, featurep
, sizeof features
))
1321 if (features
& ~VHOST_NET_FEATURES
)
1323 return vhost_net_set_features(n
, features
);
1324 case VHOST_RESET_OWNER
:
1325 return vhost_net_reset_owner(n
);
1326 case VHOST_SET_OWNER
:
1327 return vhost_net_set_owner(n
);
1329 mutex_lock(&n
->dev
.mutex
);
1330 r
= vhost_dev_ioctl(&n
->dev
, ioctl
, argp
);
1331 if (r
== -ENOIOCTLCMD
)
1332 r
= vhost_vring_ioctl(&n
->dev
, ioctl
, argp
);
1335 mutex_unlock(&n
->dev
.mutex
);
1340 #ifdef CONFIG_COMPAT
1341 static long vhost_net_compat_ioctl(struct file
*f
, unsigned int ioctl
,
1344 return vhost_net_ioctl(f
, ioctl
, (unsigned long)compat_ptr(arg
));
1348 static ssize_t
vhost_net_chr_read_iter(struct kiocb
*iocb
, struct iov_iter
*to
)
1350 struct file
*file
= iocb
->ki_filp
;
1351 struct vhost_net
*n
= file
->private_data
;
1352 struct vhost_dev
*dev
= &n
->dev
;
1353 int noblock
= file
->f_flags
& O_NONBLOCK
;
1355 return vhost_chr_read_iter(dev
, to
, noblock
);
1358 static ssize_t
vhost_net_chr_write_iter(struct kiocb
*iocb
,
1359 struct iov_iter
*from
)
1361 struct file
*file
= iocb
->ki_filp
;
1362 struct vhost_net
*n
= file
->private_data
;
1363 struct vhost_dev
*dev
= &n
->dev
;
1365 return vhost_chr_write_iter(dev
, from
);
1368 static unsigned int vhost_net_chr_poll(struct file
*file
, poll_table
*wait
)
1370 struct vhost_net
*n
= file
->private_data
;
1371 struct vhost_dev
*dev
= &n
->dev
;
1373 return vhost_chr_poll(file
, dev
, wait
);
1376 static const struct file_operations vhost_net_fops
= {
1377 .owner
= THIS_MODULE
,
1378 .release
= vhost_net_release
,
1379 .read_iter
= vhost_net_chr_read_iter
,
1380 .write_iter
= vhost_net_chr_write_iter
,
1381 .poll
= vhost_net_chr_poll
,
1382 .unlocked_ioctl
= vhost_net_ioctl
,
1383 #ifdef CONFIG_COMPAT
1384 .compat_ioctl
= vhost_net_compat_ioctl
,
1386 .open
= vhost_net_open
,
1387 .llseek
= noop_llseek
,
1390 static struct miscdevice vhost_net_misc
= {
1391 .minor
= VHOST_NET_MINOR
,
1392 .name
= "vhost-net",
1393 .fops
= &vhost_net_fops
,
1396 static int vhost_net_init(void)
1398 if (experimental_zcopytx
)
1399 vhost_net_enable_zcopy(VHOST_NET_VQ_TX
);
1400 return misc_register(&vhost_net_misc
);
1402 module_init(vhost_net_init
);
1404 static void vhost_net_exit(void)
1406 misc_deregister(&vhost_net_misc
);
1408 module_exit(vhost_net_exit
);
1410 MODULE_VERSION("0.0.1");
1411 MODULE_LICENSE("GPL v2");
1412 MODULE_AUTHOR("Michael S. Tsirkin");
1413 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1414 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR
);
1415 MODULE_ALIAS("devname:vhost-net");