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));
156 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
162 for (i = 0; i < n->dev.nvqs; ++i) {
163 zcopy = vhost_net_zcopy_mask & (0x1 << i);
165 kfree(n->vqs[i].ubuf_info);
169 int vhost_net_set_ubuf_info(struct vhost_net *n)
174 for (i = 0; i < n->dev.nvqs; ++i) {
175 zcopy = vhost_net_zcopy_mask & (0x1 << i);
178 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
179 UIO_MAXIOV, GFP_KERNEL);
180 if (!n->vqs[i].ubuf_info)
187 zcopy = vhost_net_zcopy_mask & (0x1 << i);
190 kfree(n->vqs[i].ubuf_info);
195 void vhost_net_vq_reset(struct vhost_net *n)
199 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
200 n->vqs[i].done_idx = 0;
201 n->vqs[i].upend_idx = 0;
202 n->vqs[i].ubufs = NULL;
203 kfree(n->vqs[i].ubuf_info);
204 n->vqs[i].ubuf_info = NULL;
205 n->vqs[i].vhost_hlen = 0;
206 n->vqs[i].sock_hlen = 0;
211 static void vhost_net_tx_packet(struct vhost_net *net)
214 if (net->tx_packets < 1024)
217 net->tx_zcopy_err = 0;
220 static void vhost_net_tx_err(struct vhost_net *net)
225 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
227 /* TX flush waits for outstanding DMAs to be done.
228 * Don't start new DMAs.
230 return !net->tx_flush &&
231 net->tx_packets / 64 >= net->tx_zcopy_err;
234 static bool vhost_sock_zcopy(struct socket *sock)
236 return unlikely(experimental_zcopytx) &&
237 sock_flag(sock->sk, SOCK_ZEROCOPY);
240 /* Pop first len bytes from iovec. Return number of segments used. */
241 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
242 size_t len, int iov_count)
247 while (len && seg < iov_count) {
248 size = min(from->iov_len, len);
249 to->iov_base = from->iov_base;
251 from->iov_len -= size;
252 from->iov_base += size;
260 /* Copy iovec entries for len bytes from iovec. */
261 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
262 size_t len, int iovcount)
267 while (len && seg < iovcount) {
268 size = min(from->iov_len, len);
269 to->iov_base = from->iov_base;
278 /* In case of DMA done not in order in lower device driver for some reason.
279 * upend_idx is used to track end of used idx, done_idx is used to track head
280 * of used idx. Once lower device DMA done contiguously, we will signal KVM
283 static int vhost_zerocopy_signal_used(struct vhost_net *net,
284 struct vhost_virtqueue *vq)
286 struct vhost_net_virtqueue *nvq =
287 container_of(vq, struct vhost_net_virtqueue, vq);
291 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
292 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
293 vhost_net_tx_err(net);
294 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
295 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
296 vhost_add_used_and_signal(vq->dev, vq,
307 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
309 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
310 struct vhost_virtqueue *vq = ubufs->vq;
311 int cnt = atomic_read(&ubufs->kref.refcount);
314 * Trigger polling thread if guest stopped submitting new buffers:
315 * in this case, the refcount after decrement will eventually reach 1
317 * We also trigger polling periodically after each 16 packets
318 * (the value 16 here is more or less arbitrary, it's tuned to trigger
319 * less than 10% of times).
321 if (cnt <= 2 || !(cnt % 16))
322 vhost_poll_queue(&vq->poll);
323 /* set len to mark this desc buffers done DMA */
324 vq->heads[ubuf->desc].len = success ?
325 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
326 vhost_net_ubuf_put(ubufs);
329 /* Expects to be always run from workqueue - which acts as
330 * read-size critical section for our kind of RCU. */
331 static void handle_tx(struct vhost_net *net)
333 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
334 struct vhost_virtqueue *vq = &nvq->vq;
337 struct msghdr msg = {
343 .msg_flags = MSG_DONTWAIT,
345 size_t len, total_len = 0;
349 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
350 bool zcopy, zcopy_used;
352 /* TODO: check that we are running from vhost_worker? */
353 sock = rcu_dereference_check(vq->private_data, 1);
357 mutex_lock(&vq->mutex);
358 vhost_disable_notify(&net->dev, vq);
360 hdr_size = nvq->vhost_hlen;
364 /* Release DMAs done buffers first */
366 vhost_zerocopy_signal_used(net, vq);
368 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
372 /* On error, stop handling until the next kick. */
373 if (unlikely(head < 0))
375 /* Nothing new? Wait for eventfd to tell us they refilled. */
376 if (head == vq->num) {
379 /* If more outstanding DMAs, queue the work.
380 * Handle upend_idx wrap around
382 num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
383 (nvq->upend_idx - nvq->done_idx) :
384 (nvq->upend_idx + UIO_MAXIOV -
386 if (unlikely(num_pends > VHOST_MAX_PEND))
388 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
389 vhost_disable_notify(&net->dev, vq);
395 vq_err(vq, "Unexpected descriptor format for TX: "
396 "out %d, int %d\n", out, in);
399 /* Skip header. TODO: support TSO. */
400 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
401 msg.msg_iovlen = out;
402 len = iov_length(vq->iov, out);
405 vq_err(vq, "Unexpected header len for TX: "
406 "%zd expected %zd\n",
407 iov_length(nvq->hdr, s), hdr_size);
410 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
411 nvq->upend_idx != nvq->done_idx);
413 /* use msg_control to pass vhost zerocopy ubuf info to skb */
415 vq->heads[nvq->upend_idx].id = head;
416 if (!vhost_net_tx_select_zcopy(net) ||
417 len < VHOST_GOODCOPY_LEN) {
418 /* copy don't need to wait for DMA done */
419 vq->heads[nvq->upend_idx].len =
421 msg.msg_control = NULL;
422 msg.msg_controllen = 0;
425 struct ubuf_info *ubuf;
426 ubuf = nvq->ubuf_info + nvq->upend_idx;
428 vq->heads[nvq->upend_idx].len =
429 VHOST_DMA_IN_PROGRESS;
430 ubuf->callback = vhost_zerocopy_callback;
431 ubuf->ctx = nvq->ubufs;
432 ubuf->desc = nvq->upend_idx;
433 msg.msg_control = ubuf;
434 msg.msg_controllen = sizeof(ubuf);
436 kref_get(&ubufs->kref);
438 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
440 /* TODO: Check specific error and bomb out unless ENOBUFS? */
441 err = sock->ops->sendmsg(NULL, sock, &msg, len);
442 if (unlikely(err < 0)) {
445 vhost_net_ubuf_put(ubufs);
446 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
449 vhost_discard_vq_desc(vq, 1);
453 pr_debug("Truncated TX packet: "
454 " len %d != %zd\n", err, len);
456 vhost_add_used_and_signal(&net->dev, vq, head, 0);
458 vhost_zerocopy_signal_used(net, vq);
460 vhost_net_tx_packet(net);
461 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
462 vhost_poll_queue(&vq->poll);
467 mutex_unlock(&vq->mutex);
470 static int peek_head_len(struct sock *sk)
472 struct sk_buff *head;
476 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
477 head = skb_peek(&sk->sk_receive_queue);
480 if (vlan_tx_tag_present(head))
484 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
488 /* This is a multi-buffer version of vhost_get_desc, that works if
489 * vq has read descriptors only.
490 * @vq - the relevant virtqueue
491 * @datalen - data length we'll be reading
492 * @iovcount - returned count of io vectors we fill
494 * @log_num - log offset
495 * @quota - headcount quota, 1 for big buffer
496 * returns number of buffer heads allocated, negative on error
498 static int get_rx_bufs(struct vhost_virtqueue *vq,
499 struct vring_used_elem *heads,
502 struct vhost_log *log,
506 unsigned int out, in;
512 while (datalen > 0 && headcount < quota) {
513 if (unlikely(seg >= UIO_MAXIOV)) {
517 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
518 ARRAY_SIZE(vq->iov) - seg, &out,
524 if (unlikely(out || in <= 0)) {
525 vq_err(vq, "unexpected descriptor format for RX: "
526 "out %d, in %d\n", out, in);
534 heads[headcount].id = d;
535 heads[headcount].len = iov_length(vq->iov + seg, in);
536 datalen -= heads[headcount].len;
540 heads[headcount - 1].len += datalen;
546 vhost_discard_vq_desc(vq, headcount);
550 /* Expects to be always run from workqueue - which acts as
551 * read-size critical section for our kind of RCU. */
552 static void handle_rx(struct vhost_net *net)
554 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
555 struct vhost_virtqueue *vq = &nvq->vq;
556 unsigned uninitialized_var(in), log;
557 struct vhost_log *vq_log;
558 struct msghdr msg = {
561 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
564 .msg_flags = MSG_DONTWAIT,
566 struct virtio_net_hdr_mrg_rxbuf hdr = {
568 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
570 size_t total_len = 0;
573 size_t vhost_hlen, sock_hlen;
574 size_t vhost_len, sock_len;
575 /* TODO: check that we are running from vhost_worker? */
576 struct socket *sock = rcu_dereference_check(vq->private_data, 1);
581 mutex_lock(&vq->mutex);
582 vhost_disable_notify(&net->dev, vq);
583 vhost_hlen = nvq->vhost_hlen;
584 sock_hlen = nvq->sock_hlen;
586 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
588 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
590 while ((sock_len = peek_head_len(sock->sk))) {
591 sock_len += sock_hlen;
592 vhost_len = sock_len + vhost_hlen;
593 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
595 likely(mergeable) ? UIO_MAXIOV : 1);
596 /* On error, stop handling until the next kick. */
597 if (unlikely(headcount < 0))
599 /* OK, now we need to know about added descriptors. */
601 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
602 /* They have slipped one in as we were
603 * doing that: check again. */
604 vhost_disable_notify(&net->dev, vq);
607 /* Nothing new? Wait for eventfd to tell us
611 /* We don't need to be notified again. */
612 if (unlikely((vhost_hlen)))
613 /* Skip header. TODO: support TSO. */
614 move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
616 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
617 * needed because recvmsg can modify msg_iov. */
618 copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
620 err = sock->ops->recvmsg(NULL, sock, &msg,
621 sock_len, MSG_DONTWAIT | MSG_TRUNC);
622 /* Userspace might have consumed the packet meanwhile:
623 * it's not supposed to do this usually, but might be hard
624 * to prevent. Discard data we got (if any) and keep going. */
625 if (unlikely(err != sock_len)) {
626 pr_debug("Discarded rx packet: "
627 " len %d, expected %zd\n", err, sock_len);
628 vhost_discard_vq_desc(vq, headcount);
631 if (unlikely(vhost_hlen) &&
632 memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
634 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
638 /* TODO: Should check and handle checksum. */
639 if (likely(mergeable) &&
640 memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
641 offsetof(typeof(hdr), num_buffers),
642 sizeof hdr.num_buffers)) {
643 vq_err(vq, "Failed num_buffers write");
644 vhost_discard_vq_desc(vq, headcount);
647 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
649 if (unlikely(vq_log))
650 vhost_log_write(vq, vq_log, log, vhost_len);
651 total_len += vhost_len;
652 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
653 vhost_poll_queue(&vq->poll);
658 mutex_unlock(&vq->mutex);
661 static void handle_tx_kick(struct vhost_work *work)
663 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
665 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
670 static void handle_rx_kick(struct vhost_work *work)
672 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
674 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
679 static void handle_tx_net(struct vhost_work *work)
681 struct vhost_net *net = container_of(work, struct vhost_net,
682 poll[VHOST_NET_VQ_TX].work);
686 static void handle_rx_net(struct vhost_work *work)
688 struct vhost_net *net = container_of(work, struct vhost_net,
689 poll[VHOST_NET_VQ_RX].work);
693 static int vhost_net_open(struct inode *inode, struct file *f)
695 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
696 struct vhost_dev *dev;
697 struct vhost_virtqueue **vqs;
702 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
709 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
710 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
711 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
712 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
713 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
714 n->vqs[i].ubufs = NULL;
715 n->vqs[i].ubuf_info = NULL;
716 n->vqs[i].upend_idx = 0;
717 n->vqs[i].done_idx = 0;
718 n->vqs[i].vhost_hlen = 0;
719 n->vqs[i].sock_hlen = 0;
721 r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
728 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
729 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
736 static void vhost_net_disable_vq(struct vhost_net *n,
737 struct vhost_virtqueue *vq)
739 struct vhost_net_virtqueue *nvq =
740 container_of(vq, struct vhost_net_virtqueue, vq);
741 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
742 if (!vq->private_data)
744 vhost_poll_stop(poll);
747 static int vhost_net_enable_vq(struct vhost_net *n,
748 struct vhost_virtqueue *vq)
750 struct vhost_net_virtqueue *nvq =
751 container_of(vq, struct vhost_net_virtqueue, vq);
752 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
755 sock = rcu_dereference_protected(vq->private_data,
756 lockdep_is_held(&vq->mutex));
760 return vhost_poll_start(poll, sock->file);
763 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
764 struct vhost_virtqueue *vq)
768 mutex_lock(&vq->mutex);
769 sock = rcu_dereference_protected(vq->private_data,
770 lockdep_is_held(&vq->mutex));
771 vhost_net_disable_vq(n, vq);
772 rcu_assign_pointer(vq->private_data, NULL);
773 mutex_unlock(&vq->mutex);
777 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
778 struct socket **rx_sock)
780 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
781 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
784 static void vhost_net_flush_vq(struct vhost_net *n, int index)
786 vhost_poll_flush(n->poll + index);
787 vhost_poll_flush(&n->vqs[index].vq.poll);
790 static void vhost_net_flush(struct vhost_net *n)
792 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
793 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
794 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
795 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
797 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
798 /* Wait for all lower device DMAs done. */
799 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
800 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
802 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
803 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
807 static int vhost_net_release(struct inode *inode, struct file *f)
809 struct vhost_net *n = f->private_data;
810 struct socket *tx_sock;
811 struct socket *rx_sock;
813 vhost_net_stop(n, &tx_sock, &rx_sock);
815 vhost_dev_stop(&n->dev);
816 vhost_dev_cleanup(&n->dev, false);
817 vhost_net_vq_reset(n);
822 /* We do an extra flush before freeing memory,
823 * since jobs can re-queue themselves. */
830 static struct socket *get_raw_socket(int fd)
833 struct sockaddr_ll sa;
834 char buf[MAX_ADDR_LEN];
836 int uaddr_len = sizeof uaddr, r;
837 struct socket *sock = sockfd_lookup(fd, &r);
840 return ERR_PTR(-ENOTSOCK);
842 /* Parameter checking */
843 if (sock->sk->sk_type != SOCK_RAW) {
844 r = -ESOCKTNOSUPPORT;
848 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
853 if (uaddr.sa.sll_family != AF_PACKET) {
863 static struct socket *get_tap_socket(int fd)
865 struct file *file = fget(fd);
869 return ERR_PTR(-EBADF);
870 sock = tun_get_socket(file);
873 sock = macvtap_get_socket(file);
879 static struct socket *get_socket(int fd)
883 /* special case to disable backend */
886 sock = get_raw_socket(fd);
889 sock = get_tap_socket(fd);
892 return ERR_PTR(-ENOTSOCK);
895 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
897 struct socket *sock, *oldsock;
898 struct vhost_virtqueue *vq;
899 struct vhost_net_virtqueue *nvq;
900 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
903 mutex_lock(&n->dev.mutex);
904 r = vhost_dev_check_owner(&n->dev);
908 if (index >= VHOST_NET_VQ_MAX) {
912 vq = &n->vqs[index].vq;
913 nvq = &n->vqs[index];
914 mutex_lock(&vq->mutex);
916 /* Verify that ring has been setup correctly. */
917 if (!vhost_vq_access_ok(vq)) {
921 sock = get_socket(fd);
927 /* start polling new socket */
928 oldsock = rcu_dereference_protected(vq->private_data,
929 lockdep_is_held(&vq->mutex));
930 if (sock != oldsock) {
931 ubufs = vhost_net_ubuf_alloc(vq,
932 sock && vhost_sock_zcopy(sock));
938 vhost_net_disable_vq(n, vq);
939 rcu_assign_pointer(vq->private_data, sock);
940 r = vhost_init_used(vq);
943 r = vhost_net_enable_vq(n, vq);
947 oldubufs = nvq->ubufs;
955 mutex_unlock(&vq->mutex);
958 vhost_net_ubuf_put_and_wait(oldubufs);
959 mutex_lock(&vq->mutex);
960 vhost_zerocopy_signal_used(n, vq);
961 mutex_unlock(&vq->mutex);
965 vhost_net_flush_vq(n, index);
969 mutex_unlock(&n->dev.mutex);
973 rcu_assign_pointer(vq->private_data, oldsock);
974 vhost_net_enable_vq(n, vq);
976 vhost_net_ubuf_put_and_wait(ubufs);
980 mutex_unlock(&vq->mutex);
982 mutex_unlock(&n->dev.mutex);
986 static long vhost_net_reset_owner(struct vhost_net *n)
988 struct socket *tx_sock = NULL;
989 struct socket *rx_sock = NULL;
991 struct vhost_memory *memory;
993 mutex_lock(&n->dev.mutex);
994 err = vhost_dev_check_owner(&n->dev);
997 memory = vhost_dev_reset_owner_prepare();
1002 vhost_net_stop(n, &tx_sock, &rx_sock);
1004 vhost_dev_reset_owner(&n->dev, memory);
1005 vhost_net_vq_reset(n);
1007 mutex_unlock(&n->dev.mutex);
1009 fput(tx_sock->file);
1011 fput(rx_sock->file);
1015 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1017 size_t vhost_hlen, sock_hlen, hdr_len;
1020 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1021 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1022 sizeof(struct virtio_net_hdr);
1023 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1024 /* vhost provides vnet_hdr */
1025 vhost_hlen = hdr_len;
1028 /* socket provides vnet_hdr */
1030 sock_hlen = hdr_len;
1032 mutex_lock(&n->dev.mutex);
1033 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1034 !vhost_log_access_ok(&n->dev)) {
1035 mutex_unlock(&n->dev.mutex);
1038 n->dev.acked_features = features;
1040 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1041 mutex_lock(&n->vqs[i].vq.mutex);
1042 n->vqs[i].vhost_hlen = vhost_hlen;
1043 n->vqs[i].sock_hlen = sock_hlen;
1044 mutex_unlock(&n->vqs[i].vq.mutex);
1047 mutex_unlock(&n->dev.mutex);
1051 static long vhost_net_set_owner(struct vhost_net *n)
1055 mutex_lock(&n->dev.mutex);
1056 r = vhost_net_set_ubuf_info(n);
1059 r = vhost_dev_set_owner(&n->dev);
1061 vhost_net_clear_ubuf_info(n);
1064 mutex_unlock(&n->dev.mutex);
1068 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1071 struct vhost_net *n = f->private_data;
1072 void __user *argp = (void __user *)arg;
1073 u64 __user *featurep = argp;
1074 struct vhost_vring_file backend;
1079 case VHOST_NET_SET_BACKEND:
1080 if (copy_from_user(&backend, argp, sizeof backend))
1082 return vhost_net_set_backend(n, backend.index, backend.fd);
1083 case VHOST_GET_FEATURES:
1084 features = VHOST_NET_FEATURES;
1085 if (copy_to_user(featurep, &features, sizeof features))
1088 case VHOST_SET_FEATURES:
1089 if (copy_from_user(&features, featurep, sizeof features))
1091 if (features & ~VHOST_NET_FEATURES)
1093 return vhost_net_set_features(n, features);
1094 case VHOST_RESET_OWNER:
1095 return vhost_net_reset_owner(n);
1096 case VHOST_SET_OWNER:
1097 return vhost_net_set_owner(n);
1099 mutex_lock(&n->dev.mutex);
1100 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1101 if (r == -ENOIOCTLCMD)
1102 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1105 mutex_unlock(&n->dev.mutex);
1110 #ifdef CONFIG_COMPAT
1111 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1114 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1118 static const struct file_operations vhost_net_fops = {
1119 .owner = THIS_MODULE,
1120 .release = vhost_net_release,
1121 .unlocked_ioctl = vhost_net_ioctl,
1122 #ifdef CONFIG_COMPAT
1123 .compat_ioctl = vhost_net_compat_ioctl,
1125 .open = vhost_net_open,
1126 .llseek = noop_llseek,
1129 static struct miscdevice vhost_net_misc = {
1130 .minor = VHOST_NET_MINOR,
1131 .name = "vhost-net",
1132 .fops = &vhost_net_fops,
1135 static int vhost_net_init(void)
1137 if (experimental_zcopytx)
1138 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1139 return misc_register(&vhost_net_misc);
1141 module_init(vhost_net_init);
1143 static void vhost_net_exit(void)
1145 misc_deregister(&vhost_net_misc);
1147 module_exit(vhost_net_exit);
1149 MODULE_VERSION("0.0.1");
1150 MODULE_LICENSE("GPL v2");
1151 MODULE_AUTHOR("Michael S. Tsirkin");
1152 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1153 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1154 MODULE_ALIAS("devname:vhost-net");