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 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 {
74 /* refcount follows semantics similar to kref:
75 * 0: object is released
76 * 1: no outstanding ubufs
77 * >1: outstanding ubufs
80 wait_queue_head_t wait;
81 struct vhost_virtqueue *vq;
84 struct vhost_net_virtqueue {
85 struct vhost_virtqueue vq;
86 /* hdr is used to store the virtio header.
87 * Since each iovec has >= 1 byte length, we never need more than
88 * header length entries to store the header. */
89 struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
92 /* vhost zerocopy support fields below: */
93 /* last used idx for outstanding DMA zerocopy buffers */
95 /* first used idx for DMA done zerocopy buffers */
97 /* an array of userspace buffers info */
98 struct ubuf_info *ubuf_info;
99 /* Reference counting for outstanding ubufs.
100 * Protected by vq mutex. Writers must also take device mutex. */
101 struct vhost_net_ubuf_ref *ubufs;
105 struct vhost_dev dev;
106 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
107 struct vhost_poll poll[VHOST_NET_VQ_MAX];
108 /* Number of TX recently submitted.
109 * Protected by tx vq lock. */
111 /* Number of times zerocopy TX recently failed.
112 * Protected by tx vq lock. */
113 unsigned tx_zcopy_err;
114 /* Flush in progress. Protected by tx vq lock. */
118 static unsigned vhost_net_zcopy_mask __read_mostly;
120 static void vhost_net_enable_zcopy(int vq)
122 vhost_net_zcopy_mask |= 0x1 << vq;
125 static struct vhost_net_ubuf_ref *
126 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
128 struct vhost_net_ubuf_ref *ubufs;
129 /* No zero copy backend? Nothing to count. */
132 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
134 return ERR_PTR(-ENOMEM);
135 atomic_set(&ubufs->refcount, 1);
136 init_waitqueue_head(&ubufs->wait);
141 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
143 int r = atomic_sub_return(1, &ubufs->refcount);
145 wake_up(&ubufs->wait);
149 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
151 vhost_net_ubuf_put(ubufs);
152 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
155 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
157 vhost_net_ubuf_put_and_wait(ubufs);
161 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
165 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
166 kfree(n->vqs[i].ubuf_info);
167 n->vqs[i].ubuf_info = NULL;
171 static int vhost_net_set_ubuf_info(struct vhost_net *n)
176 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
177 zcopy = vhost_net_zcopy_mask & (0x1 << i);
180 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
181 UIO_MAXIOV, GFP_KERNEL);
182 if (!n->vqs[i].ubuf_info)
188 vhost_net_clear_ubuf_info(n);
192 static void vhost_net_vq_reset(struct vhost_net *n)
196 vhost_net_clear_ubuf_info(n);
198 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
199 n->vqs[i].done_idx = 0;
200 n->vqs[i].upend_idx = 0;
201 n->vqs[i].ubufs = NULL;
202 n->vqs[i].vhost_hlen = 0;
203 n->vqs[i].sock_hlen = 0;
208 static void vhost_net_tx_packet(struct vhost_net *net)
211 if (net->tx_packets < 1024)
214 net->tx_zcopy_err = 0;
217 static void vhost_net_tx_err(struct vhost_net *net)
222 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
224 /* TX flush waits for outstanding DMAs to be done.
225 * Don't start new DMAs.
227 return !net->tx_flush &&
228 net->tx_packets / 64 >= net->tx_zcopy_err;
231 static bool vhost_sock_zcopy(struct socket *sock)
233 return unlikely(experimental_zcopytx) &&
234 sock_flag(sock->sk, SOCK_ZEROCOPY);
237 /* Pop first len bytes from iovec. Return number of segments used. */
238 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
239 size_t len, int iov_count)
244 while (len && seg < iov_count) {
245 size = min(from->iov_len, len);
246 to->iov_base = from->iov_base;
248 from->iov_len -= size;
249 from->iov_base += size;
257 /* Copy iovec entries for len bytes from iovec. */
258 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
259 size_t len, int iovcount)
264 while (len && seg < iovcount) {
265 size = min(from->iov_len, len);
266 to->iov_base = from->iov_base;
275 /* In case of DMA done not in order in lower device driver for some reason.
276 * upend_idx is used to track end of used idx, done_idx is used to track head
277 * of used idx. Once lower device DMA done contiguously, we will signal KVM
280 static void vhost_zerocopy_signal_used(struct vhost_net *net,
281 struct vhost_virtqueue *vq)
283 struct vhost_net_virtqueue *nvq =
284 container_of(vq, struct vhost_net_virtqueue, vq);
288 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
289 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
290 vhost_net_tx_err(net);
291 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
292 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
298 add = min(UIO_MAXIOV - nvq->done_idx, j);
299 vhost_add_used_and_signal_n(vq->dev, vq,
300 &vq->heads[nvq->done_idx], add);
301 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
306 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
308 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
309 struct vhost_virtqueue *vq = ubufs->vq;
314 /* set len to mark this desc buffers done DMA */
315 vq->heads[ubuf->desc].len = success ?
316 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
317 cnt = vhost_net_ubuf_put(ubufs);
320 * Trigger polling thread if guest stopped submitting new buffers:
321 * in this case, the refcount after decrement will eventually reach 1.
322 * We also trigger polling periodically after each 16 packets
323 * (the value 16 here is more or less arbitrary, it's tuned to trigger
324 * less than 10% of times).
326 if (cnt <= 1 || !(cnt % 16))
327 vhost_poll_queue(&vq->poll);
329 rcu_read_unlock_bh();
332 /* Expects to be always run from workqueue - which acts as
333 * read-size critical section for our kind of RCU. */
334 static void handle_tx(struct vhost_net *net)
336 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
337 struct vhost_virtqueue *vq = &nvq->vq;
340 struct msghdr msg = {
346 .msg_flags = MSG_DONTWAIT,
348 size_t len, total_len = 0;
352 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
353 bool zcopy, zcopy_used;
355 mutex_lock(&vq->mutex);
356 sock = vq->private_data;
360 vhost_disable_notify(&net->dev, vq);
362 hdr_size = nvq->vhost_hlen;
366 /* Release DMAs done buffers first */
368 vhost_zerocopy_signal_used(net, vq);
370 /* If more outstanding DMAs, queue the work.
371 * Handle upend_idx wrap around
373 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
374 % UIO_MAXIOV == nvq->done_idx))
377 head = vhost_get_vq_desc(vq, vq->iov,
381 /* On error, stop handling until the next kick. */
382 if (unlikely(head < 0))
384 /* Nothing new? Wait for eventfd to tell us they refilled. */
385 if (head == vq->num) {
386 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
387 vhost_disable_notify(&net->dev, vq);
393 vq_err(vq, "Unexpected descriptor format for TX: "
394 "out %d, int %d\n", out, in);
397 /* Skip header. TODO: support TSO. */
398 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
399 msg.msg_iovlen = out;
400 len = iov_length(vq->iov, out);
403 vq_err(vq, "Unexpected header len for TX: "
404 "%zd expected %zd\n",
405 iov_length(nvq->hdr, s), hdr_size);
409 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
410 && (nvq->upend_idx + 1) % UIO_MAXIOV !=
412 && vhost_net_tx_select_zcopy(net);
414 /* use msg_control to pass vhost zerocopy ubuf info to skb */
416 struct ubuf_info *ubuf;
417 ubuf = nvq->ubuf_info + nvq->upend_idx;
419 vq->heads[nvq->upend_idx].id = head;
420 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
421 ubuf->callback = vhost_zerocopy_callback;
422 ubuf->ctx = nvq->ubufs;
423 ubuf->desc = nvq->upend_idx;
424 msg.msg_control = ubuf;
425 msg.msg_controllen = sizeof(ubuf);
427 atomic_inc(&ubufs->refcount);
428 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
430 msg.msg_control = NULL;
433 /* TODO: Check specific error and bomb out unless ENOBUFS? */
434 err = sock->ops->sendmsg(NULL, sock, &msg, len);
435 if (unlikely(err < 0)) {
437 vhost_net_ubuf_put(ubufs);
438 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
441 vhost_discard_vq_desc(vq, 1);
445 pr_debug("Truncated TX packet: "
446 " len %d != %zd\n", err, len);
448 vhost_add_used_and_signal(&net->dev, vq, head, 0);
450 vhost_zerocopy_signal_used(net, vq);
452 vhost_net_tx_packet(net);
453 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
454 vhost_poll_queue(&vq->poll);
459 mutex_unlock(&vq->mutex);
462 static int peek_head_len(struct sock *sk)
464 struct sk_buff *head;
468 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
469 head = skb_peek(&sk->sk_receive_queue);
472 if (vlan_tx_tag_present(head))
476 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
480 /* This is a multi-buffer version of vhost_get_desc, that works if
481 * vq has read descriptors only.
482 * @vq - the relevant virtqueue
483 * @datalen - data length we'll be reading
484 * @iovcount - returned count of io vectors we fill
486 * @log_num - log offset
487 * @quota - headcount quota, 1 for big buffer
488 * returns number of buffer heads allocated, negative on error
490 static int get_rx_bufs(struct vhost_virtqueue *vq,
491 struct vring_used_elem *heads,
494 struct vhost_log *log,
498 unsigned int out, in;
504 while (datalen > 0 && headcount < quota) {
505 if (unlikely(seg >= UIO_MAXIOV)) {
509 r = vhost_get_vq_desc(vq, vq->iov + seg,
510 ARRAY_SIZE(vq->iov) - seg, &out,
520 if (unlikely(out || in <= 0)) {
521 vq_err(vq, "unexpected descriptor format for RX: "
522 "out %d, in %d\n", out, in);
530 heads[headcount].id = d;
531 heads[headcount].len = iov_length(vq->iov + seg, in);
532 datalen -= heads[headcount].len;
536 heads[headcount - 1].len += datalen;
542 if (unlikely(datalen > 0)) {
548 vhost_discard_vq_desc(vq, headcount);
552 /* Expects to be always run from workqueue - which acts as
553 * read-size critical section for our kind of RCU. */
554 static void handle_rx(struct vhost_net *net)
556 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
557 struct vhost_virtqueue *vq = &nvq->vq;
558 unsigned uninitialized_var(in), log;
559 struct vhost_log *vq_log;
560 struct msghdr msg = {
563 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
566 .msg_flags = MSG_DONTWAIT,
568 struct virtio_net_hdr_mrg_rxbuf hdr = {
570 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
572 size_t total_len = 0;
575 size_t vhost_hlen, sock_hlen;
576 size_t vhost_len, sock_len;
579 mutex_lock(&vq->mutex);
580 sock = vq->private_data;
583 vhost_disable_notify(&net->dev, vq);
585 vhost_hlen = nvq->vhost_hlen;
586 sock_hlen = nvq->sock_hlen;
588 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
590 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
592 while ((sock_len = peek_head_len(sock->sk))) {
593 sock_len += sock_hlen;
594 vhost_len = sock_len + vhost_hlen;
595 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
597 likely(mergeable) ? UIO_MAXIOV : 1);
598 /* On error, stop handling until the next kick. */
599 if (unlikely(headcount < 0))
601 /* On overrun, truncate and discard */
602 if (unlikely(headcount > UIO_MAXIOV)) {
604 err = sock->ops->recvmsg(NULL, sock, &msg,
605 1, MSG_DONTWAIT | MSG_TRUNC);
606 pr_debug("Discarded rx packet: len %zd\n", sock_len);
609 /* OK, now we need to know about added descriptors. */
611 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
612 /* They have slipped one in as we were
613 * doing that: check again. */
614 vhost_disable_notify(&net->dev, vq);
617 /* Nothing new? Wait for eventfd to tell us
621 /* We don't need to be notified again. */
622 if (unlikely((vhost_hlen)))
623 /* Skip header. TODO: support TSO. */
624 move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
626 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
627 * needed because recvmsg can modify msg_iov. */
628 copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
630 err = sock->ops->recvmsg(NULL, sock, &msg,
631 sock_len, MSG_DONTWAIT | MSG_TRUNC);
632 /* Userspace might have consumed the packet meanwhile:
633 * it's not supposed to do this usually, but might be hard
634 * to prevent. Discard data we got (if any) and keep going. */
635 if (unlikely(err != sock_len)) {
636 pr_debug("Discarded rx packet: "
637 " len %d, expected %zd\n", err, sock_len);
638 vhost_discard_vq_desc(vq, headcount);
641 if (unlikely(vhost_hlen) &&
642 memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
644 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
648 /* TODO: Should check and handle checksum. */
649 if (likely(mergeable) &&
650 memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
651 offsetof(typeof(hdr), num_buffers),
652 sizeof hdr.num_buffers)) {
653 vq_err(vq, "Failed num_buffers write");
654 vhost_discard_vq_desc(vq, headcount);
657 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
659 if (unlikely(vq_log))
660 vhost_log_write(vq, vq_log, log, vhost_len);
661 total_len += vhost_len;
662 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
663 vhost_poll_queue(&vq->poll);
668 mutex_unlock(&vq->mutex);
671 static void handle_tx_kick(struct vhost_work *work)
673 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
675 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
680 static void handle_rx_kick(struct vhost_work *work)
682 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
684 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
689 static void handle_tx_net(struct vhost_work *work)
691 struct vhost_net *net = container_of(work, struct vhost_net,
692 poll[VHOST_NET_VQ_TX].work);
696 static void handle_rx_net(struct vhost_work *work)
698 struct vhost_net *net = container_of(work, struct vhost_net,
699 poll[VHOST_NET_VQ_RX].work);
703 static int vhost_net_open(struct inode *inode, struct file *f)
706 struct vhost_dev *dev;
707 struct vhost_virtqueue **vqs;
710 n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
712 n = vmalloc(sizeof *n);
716 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
723 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
724 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
725 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
726 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
727 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
728 n->vqs[i].ubufs = NULL;
729 n->vqs[i].ubuf_info = NULL;
730 n->vqs[i].upend_idx = 0;
731 n->vqs[i].done_idx = 0;
732 n->vqs[i].vhost_hlen = 0;
733 n->vqs[i].sock_hlen = 0;
735 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
737 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
738 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
745 static void vhost_net_disable_vq(struct vhost_net *n,
746 struct vhost_virtqueue *vq)
748 struct vhost_net_virtqueue *nvq =
749 container_of(vq, struct vhost_net_virtqueue, vq);
750 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
751 if (!vq->private_data)
753 vhost_poll_stop(poll);
756 static int vhost_net_enable_vq(struct vhost_net *n,
757 struct vhost_virtqueue *vq)
759 struct vhost_net_virtqueue *nvq =
760 container_of(vq, struct vhost_net_virtqueue, vq);
761 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
764 sock = vq->private_data;
768 return vhost_poll_start(poll, sock->file);
771 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
772 struct vhost_virtqueue *vq)
776 mutex_lock(&vq->mutex);
777 sock = vq->private_data;
778 vhost_net_disable_vq(n, vq);
779 vq->private_data = NULL;
780 mutex_unlock(&vq->mutex);
784 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
785 struct socket **rx_sock)
787 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
788 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
791 static void vhost_net_flush_vq(struct vhost_net *n, int index)
793 vhost_poll_flush(n->poll + index);
794 vhost_poll_flush(&n->vqs[index].vq.poll);
797 static void vhost_net_flush(struct vhost_net *n)
799 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
800 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
801 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
802 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
804 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
805 /* Wait for all lower device DMAs done. */
806 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
807 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
809 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
810 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
814 static int vhost_net_release(struct inode *inode, struct file *f)
816 struct vhost_net *n = f->private_data;
817 struct socket *tx_sock;
818 struct socket *rx_sock;
820 vhost_net_stop(n, &tx_sock, &rx_sock);
822 vhost_dev_stop(&n->dev);
823 vhost_dev_cleanup(&n->dev, false);
824 vhost_net_vq_reset(n);
829 /* Make sure no callbacks are outstanding */
830 synchronize_rcu_bh();
831 /* We do an extra flush before freeing memory,
832 * since jobs can re-queue themselves. */
839 static struct socket *get_raw_socket(int fd)
842 struct sockaddr_ll sa;
843 char buf[MAX_ADDR_LEN];
845 int uaddr_len = sizeof uaddr, r;
846 struct socket *sock = sockfd_lookup(fd, &r);
849 return ERR_PTR(-ENOTSOCK);
851 /* Parameter checking */
852 if (sock->sk->sk_type != SOCK_RAW) {
853 r = -ESOCKTNOSUPPORT;
857 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
862 if (uaddr.sa.sll_family != AF_PACKET) {
872 static struct socket *get_tap_socket(int fd)
874 struct file *file = fget(fd);
878 return ERR_PTR(-EBADF);
879 sock = tun_get_socket(file);
882 sock = macvtap_get_socket(file);
888 static struct socket *get_socket(int fd)
892 /* special case to disable backend */
895 sock = get_raw_socket(fd);
898 sock = get_tap_socket(fd);
901 return ERR_PTR(-ENOTSOCK);
904 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
906 struct socket *sock, *oldsock;
907 struct vhost_virtqueue *vq;
908 struct vhost_net_virtqueue *nvq;
909 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
912 mutex_lock(&n->dev.mutex);
913 r = vhost_dev_check_owner(&n->dev);
917 if (index >= VHOST_NET_VQ_MAX) {
921 vq = &n->vqs[index].vq;
922 nvq = &n->vqs[index];
923 mutex_lock(&vq->mutex);
925 /* Verify that ring has been setup correctly. */
926 if (!vhost_vq_access_ok(vq)) {
930 sock = get_socket(fd);
936 /* start polling new socket */
937 oldsock = vq->private_data;
938 if (sock != oldsock) {
939 ubufs = vhost_net_ubuf_alloc(vq,
940 sock && vhost_sock_zcopy(sock));
946 vhost_net_disable_vq(n, vq);
947 vq->private_data = sock;
948 r = vhost_init_used(vq);
951 r = vhost_net_enable_vq(n, vq);
955 oldubufs = nvq->ubufs;
963 mutex_unlock(&vq->mutex);
966 vhost_net_ubuf_put_wait_and_free(oldubufs);
967 mutex_lock(&vq->mutex);
968 vhost_zerocopy_signal_used(n, vq);
969 mutex_unlock(&vq->mutex);
973 vhost_net_flush_vq(n, index);
977 mutex_unlock(&n->dev.mutex);
981 vq->private_data = oldsock;
982 vhost_net_enable_vq(n, vq);
984 vhost_net_ubuf_put_wait_and_free(ubufs);
988 mutex_unlock(&vq->mutex);
990 mutex_unlock(&n->dev.mutex);
994 static long vhost_net_reset_owner(struct vhost_net *n)
996 struct socket *tx_sock = NULL;
997 struct socket *rx_sock = NULL;
999 struct vhost_memory *memory;
1001 mutex_lock(&n->dev.mutex);
1002 err = vhost_dev_check_owner(&n->dev);
1005 memory = vhost_dev_reset_owner_prepare();
1010 vhost_net_stop(n, &tx_sock, &rx_sock);
1012 vhost_dev_reset_owner(&n->dev, memory);
1013 vhost_net_vq_reset(n);
1015 mutex_unlock(&n->dev.mutex);
1017 sockfd_put(tx_sock);
1019 sockfd_put(rx_sock);
1023 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1025 size_t vhost_hlen, sock_hlen, hdr_len;
1028 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1029 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1030 sizeof(struct virtio_net_hdr);
1031 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1032 /* vhost provides vnet_hdr */
1033 vhost_hlen = hdr_len;
1036 /* socket provides vnet_hdr */
1038 sock_hlen = hdr_len;
1040 mutex_lock(&n->dev.mutex);
1041 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1042 !vhost_log_access_ok(&n->dev)) {
1043 mutex_unlock(&n->dev.mutex);
1046 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1047 mutex_lock(&n->vqs[i].vq.mutex);
1048 n->vqs[i].vq.acked_features = features;
1049 n->vqs[i].vhost_hlen = vhost_hlen;
1050 n->vqs[i].sock_hlen = sock_hlen;
1051 mutex_unlock(&n->vqs[i].vq.mutex);
1053 mutex_unlock(&n->dev.mutex);
1057 static long vhost_net_set_owner(struct vhost_net *n)
1061 mutex_lock(&n->dev.mutex);
1062 if (vhost_dev_has_owner(&n->dev)) {
1066 r = vhost_net_set_ubuf_info(n);
1069 r = vhost_dev_set_owner(&n->dev);
1071 vhost_net_clear_ubuf_info(n);
1074 mutex_unlock(&n->dev.mutex);
1078 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1081 struct vhost_net *n = f->private_data;
1082 void __user *argp = (void __user *)arg;
1083 u64 __user *featurep = argp;
1084 struct vhost_vring_file backend;
1089 case VHOST_NET_SET_BACKEND:
1090 if (copy_from_user(&backend, argp, sizeof backend))
1092 return vhost_net_set_backend(n, backend.index, backend.fd);
1093 case VHOST_GET_FEATURES:
1094 features = VHOST_NET_FEATURES;
1095 if (copy_to_user(featurep, &features, sizeof features))
1098 case VHOST_SET_FEATURES:
1099 if (copy_from_user(&features, featurep, sizeof features))
1101 if (features & ~VHOST_NET_FEATURES)
1103 return vhost_net_set_features(n, features);
1104 case VHOST_RESET_OWNER:
1105 return vhost_net_reset_owner(n);
1106 case VHOST_SET_OWNER:
1107 return vhost_net_set_owner(n);
1109 mutex_lock(&n->dev.mutex);
1110 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1111 if (r == -ENOIOCTLCMD)
1112 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1115 mutex_unlock(&n->dev.mutex);
1120 #ifdef CONFIG_COMPAT
1121 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1124 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1128 static const struct file_operations vhost_net_fops = {
1129 .owner = THIS_MODULE,
1130 .release = vhost_net_release,
1131 .unlocked_ioctl = vhost_net_ioctl,
1132 #ifdef CONFIG_COMPAT
1133 .compat_ioctl = vhost_net_compat_ioctl,
1135 .open = vhost_net_open,
1136 .llseek = noop_llseek,
1139 static struct miscdevice vhost_net_misc = {
1140 .minor = VHOST_NET_MINOR,
1141 .name = "vhost-net",
1142 .fops = &vhost_net_fops,
1145 static int vhost_net_init(void)
1147 if (experimental_zcopytx)
1148 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1149 return misc_register(&vhost_net_misc);
1151 module_init(vhost_net_init);
1153 static void vhost_net_exit(void)
1155 misc_deregister(&vhost_net_misc);
1157 module_exit(vhost_net_exit);
1159 MODULE_VERSION("0.0.1");
1160 MODULE_LICENSE("GPL v2");
1161 MODULE_AUTHOR("Michael S. Tsirkin");
1162 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1163 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1164 MODULE_ALIAS("devname:vhost-net");