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/vmalloc.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 ((__force __virtio32)3)
52 /* Lower device DMA done */
53 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
54 /* Lower device DMA in progress */
55 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
57 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
59 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)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) |
65 (1ULL << VIRTIO_F_VERSION_1),
74 struct vhost_net_ubuf_ref {
75 /* refcount follows semantics similar to kref:
76 * 0: object is released
77 * 1: no outstanding ubufs
78 * >1: outstanding ubufs
81 wait_queue_head_t wait;
82 struct vhost_virtqueue *vq;
85 struct vhost_net_virtqueue {
86 struct vhost_virtqueue vq;
89 /* vhost zerocopy support fields below: */
90 /* last used idx for outstanding DMA zerocopy buffers */
92 /* first used idx for DMA done zerocopy buffers */
94 /* an array of userspace buffers info */
95 struct ubuf_info *ubuf_info;
96 /* Reference counting for outstanding ubufs.
97 * Protected by vq mutex. Writers must also take device mutex. */
98 struct vhost_net_ubuf_ref *ubufs;
102 struct vhost_dev dev;
103 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
104 struct vhost_poll poll[VHOST_NET_VQ_MAX];
105 /* Number of TX recently submitted.
106 * Protected by tx vq lock. */
108 /* Number of times zerocopy TX recently failed.
109 * Protected by tx vq lock. */
110 unsigned tx_zcopy_err;
111 /* Flush in progress. Protected by tx vq lock. */
115 static unsigned vhost_net_zcopy_mask __read_mostly;
117 static void vhost_net_enable_zcopy(int vq)
119 vhost_net_zcopy_mask |= 0x1 << vq;
122 static struct vhost_net_ubuf_ref *
123 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
125 struct vhost_net_ubuf_ref *ubufs;
126 /* No zero copy backend? Nothing to count. */
129 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
131 return ERR_PTR(-ENOMEM);
132 atomic_set(&ubufs->refcount, 1);
133 init_waitqueue_head(&ubufs->wait);
138 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
140 int r = atomic_sub_return(1, &ubufs->refcount);
142 wake_up(&ubufs->wait);
146 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
148 vhost_net_ubuf_put(ubufs);
149 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
152 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
154 vhost_net_ubuf_put_and_wait(ubufs);
158 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
162 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
163 kfree(n->vqs[i].ubuf_info);
164 n->vqs[i].ubuf_info = NULL;
168 static int vhost_net_set_ubuf_info(struct vhost_net *n)
173 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
174 zcopy = vhost_net_zcopy_mask & (0x1 << i);
177 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
178 UIO_MAXIOV, GFP_KERNEL);
179 if (!n->vqs[i].ubuf_info)
185 vhost_net_clear_ubuf_info(n);
189 static void vhost_net_vq_reset(struct vhost_net *n)
193 vhost_net_clear_ubuf_info(n);
195 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
196 n->vqs[i].done_idx = 0;
197 n->vqs[i].upend_idx = 0;
198 n->vqs[i].ubufs = NULL;
199 n->vqs[i].vhost_hlen = 0;
200 n->vqs[i].sock_hlen = 0;
205 static void vhost_net_tx_packet(struct vhost_net *net)
208 if (net->tx_packets < 1024)
211 net->tx_zcopy_err = 0;
214 static void vhost_net_tx_err(struct vhost_net *net)
219 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
221 /* TX flush waits for outstanding DMAs to be done.
222 * Don't start new DMAs.
224 return !net->tx_flush &&
225 net->tx_packets / 64 >= net->tx_zcopy_err;
228 static bool vhost_sock_zcopy(struct socket *sock)
230 return unlikely(experimental_zcopytx) &&
231 sock_flag(sock->sk, SOCK_ZEROCOPY);
234 /* In case of DMA done not in order in lower device driver for some reason.
235 * upend_idx is used to track end of used idx, done_idx is used to track head
236 * of used idx. Once lower device DMA done contiguously, we will signal KVM
239 static void vhost_zerocopy_signal_used(struct vhost_net *net,
240 struct vhost_virtqueue *vq)
242 struct vhost_net_virtqueue *nvq =
243 container_of(vq, struct vhost_net_virtqueue, vq);
247 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
248 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
249 vhost_net_tx_err(net);
250 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
251 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
257 add = min(UIO_MAXIOV - nvq->done_idx, j);
258 vhost_add_used_and_signal_n(vq->dev, vq,
259 &vq->heads[nvq->done_idx], add);
260 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
265 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
267 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
268 struct vhost_virtqueue *vq = ubufs->vq;
273 /* set len to mark this desc buffers done DMA */
274 vq->heads[ubuf->desc].len = success ?
275 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
276 cnt = vhost_net_ubuf_put(ubufs);
279 * Trigger polling thread if guest stopped submitting new buffers:
280 * in this case, the refcount after decrement will eventually reach 1.
281 * We also trigger polling periodically after each 16 packets
282 * (the value 16 here is more or less arbitrary, it's tuned to trigger
283 * less than 10% of times).
285 if (cnt <= 1 || !(cnt % 16))
286 vhost_poll_queue(&vq->poll);
288 rcu_read_unlock_bh();
291 /* Expects to be always run from workqueue - which acts as
292 * read-size critical section for our kind of RCU. */
293 static void handle_tx(struct vhost_net *net)
295 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
296 struct vhost_virtqueue *vq = &nvq->vq;
299 struct msghdr msg = {
304 .msg_flags = MSG_DONTWAIT,
306 size_t len, total_len = 0;
310 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
311 bool zcopy, zcopy_used;
313 mutex_lock(&vq->mutex);
314 sock = vq->private_data;
318 vhost_disable_notify(&net->dev, vq);
320 hdr_size = nvq->vhost_hlen;
324 /* Release DMAs done buffers first */
326 vhost_zerocopy_signal_used(net, vq);
328 /* If more outstanding DMAs, queue the work.
329 * Handle upend_idx wrap around
331 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
332 % UIO_MAXIOV == nvq->done_idx))
335 head = vhost_get_vq_desc(vq, vq->iov,
339 /* On error, stop handling until the next kick. */
340 if (unlikely(head < 0))
342 /* Nothing new? Wait for eventfd to tell us they refilled. */
343 if (head == vq->num) {
344 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
345 vhost_disable_notify(&net->dev, vq);
351 vq_err(vq, "Unexpected descriptor format for TX: "
352 "out %d, int %d\n", out, in);
355 /* Skip header. TODO: support TSO. */
356 len = iov_length(vq->iov, out);
357 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
358 iov_iter_advance(&msg.msg_iter, hdr_size);
360 if (!iov_iter_count(&msg.msg_iter)) {
361 vq_err(vq, "Unexpected header len for TX: "
362 "%zd expected %zd\n",
366 len = iov_iter_count(&msg.msg_iter);
368 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
369 && (nvq->upend_idx + 1) % UIO_MAXIOV !=
371 && vhost_net_tx_select_zcopy(net);
373 /* use msg_control to pass vhost zerocopy ubuf info to skb */
375 struct ubuf_info *ubuf;
376 ubuf = nvq->ubuf_info + nvq->upend_idx;
378 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
379 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
380 ubuf->callback = vhost_zerocopy_callback;
381 ubuf->ctx = nvq->ubufs;
382 ubuf->desc = nvq->upend_idx;
383 msg.msg_control = ubuf;
384 msg.msg_controllen = sizeof(ubuf);
386 atomic_inc(&ubufs->refcount);
387 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
389 msg.msg_control = NULL;
392 /* TODO: Check specific error and bomb out unless ENOBUFS? */
393 err = sock->ops->sendmsg(NULL, sock, &msg, len);
394 if (unlikely(err < 0)) {
396 vhost_net_ubuf_put(ubufs);
397 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
400 vhost_discard_vq_desc(vq, 1);
404 pr_debug("Truncated TX packet: "
405 " len %d != %zd\n", err, len);
407 vhost_add_used_and_signal(&net->dev, vq, head, 0);
409 vhost_zerocopy_signal_used(net, vq);
411 vhost_net_tx_packet(net);
412 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
413 vhost_poll_queue(&vq->poll);
418 mutex_unlock(&vq->mutex);
421 static int peek_head_len(struct sock *sk)
423 struct sk_buff *head;
427 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
428 head = skb_peek(&sk->sk_receive_queue);
431 if (skb_vlan_tag_present(head))
435 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
439 /* This is a multi-buffer version of vhost_get_desc, that works if
440 * vq has read descriptors only.
441 * @vq - the relevant virtqueue
442 * @datalen - data length we'll be reading
443 * @iovcount - returned count of io vectors we fill
445 * @log_num - log offset
446 * @quota - headcount quota, 1 for big buffer
447 * returns number of buffer heads allocated, negative on error
449 static int get_rx_bufs(struct vhost_virtqueue *vq,
450 struct vring_used_elem *heads,
453 struct vhost_log *log,
457 unsigned int out, in;
462 /* len is always initialized before use since we are always called with
465 u32 uninitialized_var(len);
467 while (datalen > 0 && headcount < quota) {
468 if (unlikely(seg >= UIO_MAXIOV)) {
472 r = vhost_get_vq_desc(vq, vq->iov + seg,
473 ARRAY_SIZE(vq->iov) - seg, &out,
483 if (unlikely(out || in <= 0)) {
484 vq_err(vq, "unexpected descriptor format for RX: "
485 "out %d, in %d\n", out, in);
493 heads[headcount].id = cpu_to_vhost32(vq, d);
494 len = iov_length(vq->iov + seg, in);
495 heads[headcount].len = cpu_to_vhost32(vq, len);
500 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
506 if (unlikely(datalen > 0)) {
512 vhost_discard_vq_desc(vq, headcount);
516 /* Expects to be always run from workqueue - which acts as
517 * read-size critical section for our kind of RCU. */
518 static void handle_rx(struct vhost_net *net)
520 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
521 struct vhost_virtqueue *vq = &nvq->vq;
522 unsigned uninitialized_var(in), log;
523 struct vhost_log *vq_log;
524 struct msghdr msg = {
527 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
529 .msg_flags = MSG_DONTWAIT,
531 struct virtio_net_hdr hdr = {
533 .gso_type = VIRTIO_NET_HDR_GSO_NONE
535 size_t total_len = 0;
538 size_t vhost_hlen, sock_hlen;
539 size_t vhost_len, sock_len;
541 struct iov_iter fixup;
542 __virtio16 num_buffers;
544 mutex_lock(&vq->mutex);
545 sock = vq->private_data;
548 vhost_disable_notify(&net->dev, vq);
550 vhost_hlen = nvq->vhost_hlen;
551 sock_hlen = nvq->sock_hlen;
553 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
555 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
557 while ((sock_len = peek_head_len(sock->sk))) {
558 sock_len += sock_hlen;
559 vhost_len = sock_len + vhost_hlen;
560 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
562 likely(mergeable) ? UIO_MAXIOV : 1);
563 /* On error, stop handling until the next kick. */
564 if (unlikely(headcount < 0))
566 /* On overrun, truncate and discard */
567 if (unlikely(headcount > UIO_MAXIOV)) {
568 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
569 err = sock->ops->recvmsg(NULL, sock, &msg,
570 1, MSG_DONTWAIT | MSG_TRUNC);
571 pr_debug("Discarded rx packet: len %zd\n", sock_len);
574 /* OK, now we need to know about added descriptors. */
576 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
577 /* They have slipped one in as we were
578 * doing that: check again. */
579 vhost_disable_notify(&net->dev, vq);
582 /* Nothing new? Wait for eventfd to tell us
586 /* We don't need to be notified again. */
587 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
588 fixup = msg.msg_iter;
589 if (unlikely((vhost_hlen))) {
590 /* We will supply the header ourselves
593 iov_iter_advance(&msg.msg_iter, vhost_hlen);
595 /* It'll come from socket; we'll need to patch
596 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
598 iov_iter_advance(&fixup, sizeof(hdr));
600 err = sock->ops->recvmsg(NULL, sock, &msg,
601 sock_len, MSG_DONTWAIT | MSG_TRUNC);
602 /* Userspace might have consumed the packet meanwhile:
603 * it's not supposed to do this usually, but might be hard
604 * to prevent. Discard data we got (if any) and keep going. */
605 if (unlikely(err != sock_len)) {
606 pr_debug("Discarded rx packet: "
607 " len %d, expected %zd\n", err, sock_len);
608 vhost_discard_vq_desc(vq, headcount);
611 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
612 if (unlikely(vhost_hlen) &&
613 copy_to_iter(&hdr, sizeof(hdr), &fixup) != sizeof(hdr)) {
614 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
618 /* TODO: Should check and handle checksum. */
620 num_buffers = cpu_to_vhost16(vq, headcount);
621 if (likely(mergeable) &&
622 copy_to_iter(&num_buffers, 2, &fixup) != 2) {
623 vq_err(vq, "Failed num_buffers write");
624 vhost_discard_vq_desc(vq, headcount);
627 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
629 if (unlikely(vq_log))
630 vhost_log_write(vq, vq_log, log, vhost_len);
631 total_len += vhost_len;
632 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
633 vhost_poll_queue(&vq->poll);
638 mutex_unlock(&vq->mutex);
641 static void handle_tx_kick(struct vhost_work *work)
643 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
645 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
650 static void handle_rx_kick(struct vhost_work *work)
652 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
654 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
659 static void handle_tx_net(struct vhost_work *work)
661 struct vhost_net *net = container_of(work, struct vhost_net,
662 poll[VHOST_NET_VQ_TX].work);
666 static void handle_rx_net(struct vhost_work *work)
668 struct vhost_net *net = container_of(work, struct vhost_net,
669 poll[VHOST_NET_VQ_RX].work);
673 static int vhost_net_open(struct inode *inode, struct file *f)
676 struct vhost_dev *dev;
677 struct vhost_virtqueue **vqs;
680 n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
682 n = vmalloc(sizeof *n);
686 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
693 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
694 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
695 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
696 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
697 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
698 n->vqs[i].ubufs = NULL;
699 n->vqs[i].ubuf_info = NULL;
700 n->vqs[i].upend_idx = 0;
701 n->vqs[i].done_idx = 0;
702 n->vqs[i].vhost_hlen = 0;
703 n->vqs[i].sock_hlen = 0;
705 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
707 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
708 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
715 static void vhost_net_disable_vq(struct vhost_net *n,
716 struct vhost_virtqueue *vq)
718 struct vhost_net_virtqueue *nvq =
719 container_of(vq, struct vhost_net_virtqueue, vq);
720 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
721 if (!vq->private_data)
723 vhost_poll_stop(poll);
726 static int vhost_net_enable_vq(struct vhost_net *n,
727 struct vhost_virtqueue *vq)
729 struct vhost_net_virtqueue *nvq =
730 container_of(vq, struct vhost_net_virtqueue, vq);
731 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
734 sock = vq->private_data;
738 return vhost_poll_start(poll, sock->file);
741 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
742 struct vhost_virtqueue *vq)
746 mutex_lock(&vq->mutex);
747 sock = vq->private_data;
748 vhost_net_disable_vq(n, vq);
749 vq->private_data = NULL;
750 mutex_unlock(&vq->mutex);
754 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
755 struct socket **rx_sock)
757 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
758 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
761 static void vhost_net_flush_vq(struct vhost_net *n, int index)
763 vhost_poll_flush(n->poll + index);
764 vhost_poll_flush(&n->vqs[index].vq.poll);
767 static void vhost_net_flush(struct vhost_net *n)
769 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
770 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
771 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
772 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
774 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
775 /* Wait for all lower device DMAs done. */
776 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
777 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
779 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
780 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
784 static int vhost_net_release(struct inode *inode, struct file *f)
786 struct vhost_net *n = f->private_data;
787 struct socket *tx_sock;
788 struct socket *rx_sock;
790 vhost_net_stop(n, &tx_sock, &rx_sock);
792 vhost_dev_stop(&n->dev);
793 vhost_dev_cleanup(&n->dev, false);
794 vhost_net_vq_reset(n);
799 /* Make sure no callbacks are outstanding */
800 synchronize_rcu_bh();
801 /* We do an extra flush before freeing memory,
802 * since jobs can re-queue themselves. */
809 static struct socket *get_raw_socket(int fd)
812 struct sockaddr_ll sa;
813 char buf[MAX_ADDR_LEN];
815 int uaddr_len = sizeof uaddr, r;
816 struct socket *sock = sockfd_lookup(fd, &r);
819 return ERR_PTR(-ENOTSOCK);
821 /* Parameter checking */
822 if (sock->sk->sk_type != SOCK_RAW) {
823 r = -ESOCKTNOSUPPORT;
827 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
832 if (uaddr.sa.sll_family != AF_PACKET) {
842 static struct socket *get_tap_socket(int fd)
844 struct file *file = fget(fd);
848 return ERR_PTR(-EBADF);
849 sock = tun_get_socket(file);
852 sock = macvtap_get_socket(file);
858 static struct socket *get_socket(int fd)
862 /* special case to disable backend */
865 sock = get_raw_socket(fd);
868 sock = get_tap_socket(fd);
871 return ERR_PTR(-ENOTSOCK);
874 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
876 struct socket *sock, *oldsock;
877 struct vhost_virtqueue *vq;
878 struct vhost_net_virtqueue *nvq;
879 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
882 mutex_lock(&n->dev.mutex);
883 r = vhost_dev_check_owner(&n->dev);
887 if (index >= VHOST_NET_VQ_MAX) {
891 vq = &n->vqs[index].vq;
892 nvq = &n->vqs[index];
893 mutex_lock(&vq->mutex);
895 /* Verify that ring has been setup correctly. */
896 if (!vhost_vq_access_ok(vq)) {
900 sock = get_socket(fd);
906 /* start polling new socket */
907 oldsock = vq->private_data;
908 if (sock != oldsock) {
909 ubufs = vhost_net_ubuf_alloc(vq,
910 sock && vhost_sock_zcopy(sock));
916 vhost_net_disable_vq(n, vq);
917 vq->private_data = sock;
918 r = vhost_init_used(vq);
921 r = vhost_net_enable_vq(n, vq);
925 oldubufs = nvq->ubufs;
933 mutex_unlock(&vq->mutex);
936 vhost_net_ubuf_put_wait_and_free(oldubufs);
937 mutex_lock(&vq->mutex);
938 vhost_zerocopy_signal_used(n, vq);
939 mutex_unlock(&vq->mutex);
943 vhost_net_flush_vq(n, index);
947 mutex_unlock(&n->dev.mutex);
951 vq->private_data = oldsock;
952 vhost_net_enable_vq(n, vq);
954 vhost_net_ubuf_put_wait_and_free(ubufs);
958 mutex_unlock(&vq->mutex);
960 mutex_unlock(&n->dev.mutex);
964 static long vhost_net_reset_owner(struct vhost_net *n)
966 struct socket *tx_sock = NULL;
967 struct socket *rx_sock = NULL;
969 struct vhost_memory *memory;
971 mutex_lock(&n->dev.mutex);
972 err = vhost_dev_check_owner(&n->dev);
975 memory = vhost_dev_reset_owner_prepare();
980 vhost_net_stop(n, &tx_sock, &rx_sock);
982 vhost_dev_reset_owner(&n->dev, memory);
983 vhost_net_vq_reset(n);
985 mutex_unlock(&n->dev.mutex);
993 static int vhost_net_set_features(struct vhost_net *n, u64 features)
995 size_t vhost_hlen, sock_hlen, hdr_len;
998 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
999 (1ULL << VIRTIO_F_VERSION_1))) ?
1000 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1001 sizeof(struct virtio_net_hdr);
1002 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1003 /* vhost provides vnet_hdr */
1004 vhost_hlen = hdr_len;
1007 /* socket provides vnet_hdr */
1009 sock_hlen = hdr_len;
1011 mutex_lock(&n->dev.mutex);
1012 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1013 !vhost_log_access_ok(&n->dev)) {
1014 mutex_unlock(&n->dev.mutex);
1017 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1018 mutex_lock(&n->vqs[i].vq.mutex);
1019 n->vqs[i].vq.acked_features = features;
1020 n->vqs[i].vhost_hlen = vhost_hlen;
1021 n->vqs[i].sock_hlen = sock_hlen;
1022 mutex_unlock(&n->vqs[i].vq.mutex);
1024 mutex_unlock(&n->dev.mutex);
1028 static long vhost_net_set_owner(struct vhost_net *n)
1032 mutex_lock(&n->dev.mutex);
1033 if (vhost_dev_has_owner(&n->dev)) {
1037 r = vhost_net_set_ubuf_info(n);
1040 r = vhost_dev_set_owner(&n->dev);
1042 vhost_net_clear_ubuf_info(n);
1045 mutex_unlock(&n->dev.mutex);
1049 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1052 struct vhost_net *n = f->private_data;
1053 void __user *argp = (void __user *)arg;
1054 u64 __user *featurep = argp;
1055 struct vhost_vring_file backend;
1060 case VHOST_NET_SET_BACKEND:
1061 if (copy_from_user(&backend, argp, sizeof backend))
1063 return vhost_net_set_backend(n, backend.index, backend.fd);
1064 case VHOST_GET_FEATURES:
1065 features = VHOST_NET_FEATURES;
1066 if (copy_to_user(featurep, &features, sizeof features))
1069 case VHOST_SET_FEATURES:
1070 if (copy_from_user(&features, featurep, sizeof features))
1072 if (features & ~VHOST_NET_FEATURES)
1074 return vhost_net_set_features(n, features);
1075 case VHOST_RESET_OWNER:
1076 return vhost_net_reset_owner(n);
1077 case VHOST_SET_OWNER:
1078 return vhost_net_set_owner(n);
1080 mutex_lock(&n->dev.mutex);
1081 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1082 if (r == -ENOIOCTLCMD)
1083 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1086 mutex_unlock(&n->dev.mutex);
1091 #ifdef CONFIG_COMPAT
1092 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1095 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1099 static const struct file_operations vhost_net_fops = {
1100 .owner = THIS_MODULE,
1101 .release = vhost_net_release,
1102 .unlocked_ioctl = vhost_net_ioctl,
1103 #ifdef CONFIG_COMPAT
1104 .compat_ioctl = vhost_net_compat_ioctl,
1106 .open = vhost_net_open,
1107 .llseek = noop_llseek,
1110 static struct miscdevice vhost_net_misc = {
1111 .minor = VHOST_NET_MINOR,
1112 .name = "vhost-net",
1113 .fops = &vhost_net_fops,
1116 static int vhost_net_init(void)
1118 if (experimental_zcopytx)
1119 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1120 return misc_register(&vhost_net_misc);
1122 module_init(vhost_net_init);
1124 static void vhost_net_exit(void)
1126 misc_deregister(&vhost_net_misc);
1128 module_exit(vhost_net_exit);
1130 MODULE_VERSION("0.0.1");
1131 MODULE_LICENSE("GPL v2");
1132 MODULE_AUTHOR("Michael S. Tsirkin");
1133 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1134 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1135 MODULE_ALIAS("devname:vhost-net");