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)
67 enum vhost_net_poll_state {
68 VHOST_NET_POLL_DISABLED = 0,
69 VHOST_NET_POLL_STARTED = 1,
70 VHOST_NET_POLL_STOPPED = 2,
75 struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
76 struct vhost_poll poll[VHOST_NET_VQ_MAX];
77 /* Tells us whether we are polling a socket for TX.
78 * We only do this when socket buffer fills up.
79 * Protected by tx vq lock. */
80 enum vhost_net_poll_state tx_poll_state;
81 /* Number of TX recently submitted.
82 * Protected by tx vq lock. */
84 /* Number of times zerocopy TX recently failed.
85 * Protected by tx vq lock. */
86 unsigned tx_zcopy_err;
87 /* Flush in progress. Protected by tx vq lock. */
91 static void vhost_net_tx_packet(struct vhost_net *net)
94 if (net->tx_packets < 1024)
97 net->tx_zcopy_err = 0;
100 static void vhost_net_tx_err(struct vhost_net *net)
105 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
107 /* TX flush waits for outstanding DMAs to be done.
108 * Don't start new DMAs.
110 return !net->tx_flush &&
111 net->tx_packets / 64 >= net->tx_zcopy_err;
114 static bool vhost_sock_zcopy(struct socket *sock)
116 return unlikely(experimental_zcopytx) &&
117 sock_flag(sock->sk, SOCK_ZEROCOPY);
120 /* Pop first len bytes from iovec. Return number of segments used. */
121 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
122 size_t len, int iov_count)
127 while (len && seg < iov_count) {
128 size = min(from->iov_len, len);
129 to->iov_base = from->iov_base;
131 from->iov_len -= size;
132 from->iov_base += size;
140 /* Copy iovec entries for len bytes from iovec. */
141 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
142 size_t len, int iovcount)
147 while (len && seg < iovcount) {
148 size = min(from->iov_len, len);
149 to->iov_base = from->iov_base;
158 /* Caller must have TX VQ lock */
159 static void tx_poll_stop(struct vhost_net *net)
161 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
163 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
164 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
167 /* Caller must have TX VQ lock */
168 static int tx_poll_start(struct vhost_net *net, struct socket *sock)
172 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
174 ret = vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
176 net->tx_poll_state = VHOST_NET_POLL_STARTED;
180 /* In case of DMA done not in order in lower device driver for some reason.
181 * upend_idx is used to track end of used idx, done_idx is used to track head
182 * of used idx. Once lower device DMA done contiguously, we will signal KVM
185 static int vhost_zerocopy_signal_used(struct vhost_net *net,
186 struct vhost_virtqueue *vq)
191 for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
192 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
193 vhost_net_tx_err(net);
194 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
195 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
196 vhost_add_used_and_signal(vq->dev, vq,
207 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
209 struct vhost_ubuf_ref *ubufs = ubuf->ctx;
210 struct vhost_virtqueue *vq = ubufs->vq;
211 int cnt = atomic_read(&ubufs->kref.refcount);
214 * Trigger polling thread if guest stopped submitting new buffers:
215 * in this case, the refcount after decrement will eventually reach 1
217 * We also trigger polling periodically after each 16 packets
218 * (the value 16 here is more or less arbitrary, it's tuned to trigger
219 * less than 10% of times).
221 if (cnt <= 2 || !(cnt % 16))
222 vhost_poll_queue(&vq->poll);
223 /* set len to mark this desc buffers done DMA */
224 vq->heads[ubuf->desc].len = success ?
225 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
226 vhost_ubuf_put(ubufs);
229 /* Expects to be always run from workqueue - which acts as
230 * read-size critical section for our kind of RCU. */
231 static void handle_tx(struct vhost_net *net)
233 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
236 struct msghdr msg = {
242 .msg_flags = MSG_DONTWAIT,
244 size_t len, total_len = 0;
248 struct vhost_ubuf_ref *uninitialized_var(ubufs);
249 bool zcopy, zcopy_used;
251 /* TODO: check that we are running from vhost_worker? */
252 sock = rcu_dereference_check(vq->private_data, 1);
256 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
257 if (wmem >= sock->sk->sk_sndbuf) {
258 mutex_lock(&vq->mutex);
259 tx_poll_start(net, sock);
260 mutex_unlock(&vq->mutex);
264 mutex_lock(&vq->mutex);
265 vhost_disable_notify(&net->dev, vq);
267 if (wmem < sock->sk->sk_sndbuf / 2)
269 hdr_size = vq->vhost_hlen;
273 /* Release DMAs done buffers first */
275 vhost_zerocopy_signal_used(net, vq);
277 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
281 /* On error, stop handling until the next kick. */
282 if (unlikely(head < 0))
284 /* Nothing new? Wait for eventfd to tell us they refilled. */
285 if (head == vq->num) {
288 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
289 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
290 tx_poll_start(net, sock);
291 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
294 /* If more outstanding DMAs, queue the work.
295 * Handle upend_idx wrap around
297 num_pends = likely(vq->upend_idx >= vq->done_idx) ?
298 (vq->upend_idx - vq->done_idx) :
299 (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
300 if (unlikely(num_pends > VHOST_MAX_PEND)) {
301 tx_poll_start(net, sock);
302 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
305 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
306 vhost_disable_notify(&net->dev, vq);
312 vq_err(vq, "Unexpected descriptor format for TX: "
313 "out %d, int %d\n", out, in);
316 /* Skip header. TODO: support TSO. */
317 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
318 msg.msg_iovlen = out;
319 len = iov_length(vq->iov, out);
322 vq_err(vq, "Unexpected header len for TX: "
323 "%zd expected %zd\n",
324 iov_length(vq->hdr, s), hdr_size);
327 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
328 vq->upend_idx != vq->done_idx);
330 /* use msg_control to pass vhost zerocopy ubuf info to skb */
332 vq->heads[vq->upend_idx].id = head;
333 if (!vhost_net_tx_select_zcopy(net) ||
334 len < VHOST_GOODCOPY_LEN) {
335 /* copy don't need to wait for DMA done */
336 vq->heads[vq->upend_idx].len =
338 msg.msg_control = NULL;
339 msg.msg_controllen = 0;
342 struct ubuf_info *ubuf = &vq->ubuf_info[head];
344 vq->heads[vq->upend_idx].len =
345 VHOST_DMA_IN_PROGRESS;
346 ubuf->callback = vhost_zerocopy_callback;
347 ubuf->ctx = vq->ubufs;
348 ubuf->desc = vq->upend_idx;
349 msg.msg_control = ubuf;
350 msg.msg_controllen = sizeof(ubuf);
352 kref_get(&ubufs->kref);
354 vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
356 /* TODO: Check specific error and bomb out unless ENOBUFS? */
357 err = sock->ops->sendmsg(NULL, sock, &msg, len);
358 if (unlikely(err < 0)) {
361 vhost_ubuf_put(ubufs);
362 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
365 vhost_discard_vq_desc(vq, 1);
366 if (err == -EAGAIN || err == -ENOBUFS)
367 tx_poll_start(net, sock);
371 pr_debug("Truncated TX packet: "
372 " len %d != %zd\n", err, len);
374 vhost_add_used_and_signal(&net->dev, vq, head, 0);
376 vhost_zerocopy_signal_used(net, vq);
378 vhost_net_tx_packet(net);
379 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
380 vhost_poll_queue(&vq->poll);
385 mutex_unlock(&vq->mutex);
388 static int peek_head_len(struct sock *sk)
390 struct sk_buff *head;
394 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
395 head = skb_peek(&sk->sk_receive_queue);
398 if (vlan_tx_tag_present(head))
402 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
406 /* This is a multi-buffer version of vhost_get_desc, that works if
407 * vq has read descriptors only.
408 * @vq - the relevant virtqueue
409 * @datalen - data length we'll be reading
410 * @iovcount - returned count of io vectors we fill
412 * @log_num - log offset
413 * @quota - headcount quota, 1 for big buffer
414 * returns number of buffer heads allocated, negative on error
416 static int get_rx_bufs(struct vhost_virtqueue *vq,
417 struct vring_used_elem *heads,
420 struct vhost_log *log,
424 unsigned int out, in;
430 while (datalen > 0 && headcount < quota) {
431 if (unlikely(seg >= UIO_MAXIOV)) {
435 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
436 ARRAY_SIZE(vq->iov) - seg, &out,
442 if (unlikely(out || in <= 0)) {
443 vq_err(vq, "unexpected descriptor format for RX: "
444 "out %d, in %d\n", out, in);
452 heads[headcount].id = d;
453 heads[headcount].len = iov_length(vq->iov + seg, in);
454 datalen -= heads[headcount].len;
458 heads[headcount - 1].len += datalen;
464 vhost_discard_vq_desc(vq, headcount);
468 /* Expects to be always run from workqueue - which acts as
469 * read-size critical section for our kind of RCU. */
470 static void handle_rx(struct vhost_net *net)
472 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
473 unsigned uninitialized_var(in), log;
474 struct vhost_log *vq_log;
475 struct msghdr msg = {
478 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
481 .msg_flags = MSG_DONTWAIT,
483 struct virtio_net_hdr_mrg_rxbuf hdr = {
485 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
487 size_t total_len = 0;
490 size_t vhost_hlen, sock_hlen;
491 size_t vhost_len, sock_len;
492 /* TODO: check that we are running from vhost_worker? */
493 struct socket *sock = rcu_dereference_check(vq->private_data, 1);
498 mutex_lock(&vq->mutex);
499 vhost_disable_notify(&net->dev, vq);
500 vhost_hlen = vq->vhost_hlen;
501 sock_hlen = vq->sock_hlen;
503 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
505 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
507 while ((sock_len = peek_head_len(sock->sk))) {
508 sock_len += sock_hlen;
509 vhost_len = sock_len + vhost_hlen;
510 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
512 likely(mergeable) ? UIO_MAXIOV : 1);
513 /* On error, stop handling until the next kick. */
514 if (unlikely(headcount < 0))
516 /* OK, now we need to know about added descriptors. */
518 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
519 /* They have slipped one in as we were
520 * doing that: check again. */
521 vhost_disable_notify(&net->dev, vq);
524 /* Nothing new? Wait for eventfd to tell us
528 /* We don't need to be notified again. */
529 if (unlikely((vhost_hlen)))
530 /* Skip header. TODO: support TSO. */
531 move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
533 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
534 * needed because recvmsg can modify msg_iov. */
535 copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
537 err = sock->ops->recvmsg(NULL, sock, &msg,
538 sock_len, MSG_DONTWAIT | MSG_TRUNC);
539 /* Userspace might have consumed the packet meanwhile:
540 * it's not supposed to do this usually, but might be hard
541 * to prevent. Discard data we got (if any) and keep going. */
542 if (unlikely(err != sock_len)) {
543 pr_debug("Discarded rx packet: "
544 " len %d, expected %zd\n", err, sock_len);
545 vhost_discard_vq_desc(vq, headcount);
548 if (unlikely(vhost_hlen) &&
549 memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
551 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
555 /* TODO: Should check and handle checksum. */
556 if (likely(mergeable) &&
557 memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
558 offsetof(typeof(hdr), num_buffers),
559 sizeof hdr.num_buffers)) {
560 vq_err(vq, "Failed num_buffers write");
561 vhost_discard_vq_desc(vq, headcount);
564 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
566 if (unlikely(vq_log))
567 vhost_log_write(vq, vq_log, log, vhost_len);
568 total_len += vhost_len;
569 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
570 vhost_poll_queue(&vq->poll);
575 mutex_unlock(&vq->mutex);
578 static void handle_tx_kick(struct vhost_work *work)
580 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
582 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
587 static void handle_rx_kick(struct vhost_work *work)
589 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
591 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
596 static void handle_tx_net(struct vhost_work *work)
598 struct vhost_net *net = container_of(work, struct vhost_net,
599 poll[VHOST_NET_VQ_TX].work);
603 static void handle_rx_net(struct vhost_work *work)
605 struct vhost_net *net = container_of(work, struct vhost_net,
606 poll[VHOST_NET_VQ_RX].work);
610 static int vhost_net_open(struct inode *inode, struct file *f)
612 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
613 struct vhost_dev *dev;
620 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
621 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
622 r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
628 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
629 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
630 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
637 static void vhost_net_disable_vq(struct vhost_net *n,
638 struct vhost_virtqueue *vq)
640 if (!vq->private_data)
642 if (vq == n->vqs + VHOST_NET_VQ_TX) {
644 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
646 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
649 static int vhost_net_enable_vq(struct vhost_net *n,
650 struct vhost_virtqueue *vq)
655 sock = rcu_dereference_protected(vq->private_data,
656 lockdep_is_held(&vq->mutex));
659 if (vq == n->vqs + VHOST_NET_VQ_TX) {
660 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
661 ret = tx_poll_start(n, sock);
663 ret = vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
668 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
669 struct vhost_virtqueue *vq)
673 mutex_lock(&vq->mutex);
674 sock = rcu_dereference_protected(vq->private_data,
675 lockdep_is_held(&vq->mutex));
676 vhost_net_disable_vq(n, vq);
677 rcu_assign_pointer(vq->private_data, NULL);
678 mutex_unlock(&vq->mutex);
682 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
683 struct socket **rx_sock)
685 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
686 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
689 static void vhost_net_flush_vq(struct vhost_net *n, int index)
691 vhost_poll_flush(n->poll + index);
692 vhost_poll_flush(&n->dev.vqs[index].poll);
695 static void vhost_net_flush(struct vhost_net *n)
697 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
698 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
699 if (n->dev.vqs[VHOST_NET_VQ_TX].ubufs) {
700 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
702 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
703 /* Wait for all lower device DMAs done. */
704 vhost_ubuf_put_and_wait(n->dev.vqs[VHOST_NET_VQ_TX].ubufs);
705 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
707 kref_init(&n->dev.vqs[VHOST_NET_VQ_TX].ubufs->kref);
708 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
712 static int vhost_net_release(struct inode *inode, struct file *f)
714 struct vhost_net *n = f->private_data;
715 struct socket *tx_sock;
716 struct socket *rx_sock;
718 vhost_net_stop(n, &tx_sock, &rx_sock);
720 vhost_dev_stop(&n->dev);
721 vhost_dev_cleanup(&n->dev, false);
726 /* We do an extra flush before freeing memory,
727 * since jobs can re-queue themselves. */
733 static struct socket *get_raw_socket(int fd)
736 struct sockaddr_ll sa;
737 char buf[MAX_ADDR_LEN];
739 int uaddr_len = sizeof uaddr, r;
740 struct socket *sock = sockfd_lookup(fd, &r);
743 return ERR_PTR(-ENOTSOCK);
745 /* Parameter checking */
746 if (sock->sk->sk_type != SOCK_RAW) {
747 r = -ESOCKTNOSUPPORT;
751 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
756 if (uaddr.sa.sll_family != AF_PACKET) {
766 static struct socket *get_tap_socket(int fd)
768 struct file *file = fget(fd);
772 return ERR_PTR(-EBADF);
773 sock = tun_get_socket(file);
776 sock = macvtap_get_socket(file);
782 static struct socket *get_socket(int fd)
786 /* special case to disable backend */
789 sock = get_raw_socket(fd);
792 sock = get_tap_socket(fd);
795 return ERR_PTR(-ENOTSOCK);
798 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
800 struct socket *sock, *oldsock;
801 struct vhost_virtqueue *vq;
802 struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
805 mutex_lock(&n->dev.mutex);
806 r = vhost_dev_check_owner(&n->dev);
810 if (index >= VHOST_NET_VQ_MAX) {
815 mutex_lock(&vq->mutex);
817 /* Verify that ring has been setup correctly. */
818 if (!vhost_vq_access_ok(vq)) {
822 sock = get_socket(fd);
828 /* start polling new socket */
829 oldsock = rcu_dereference_protected(vq->private_data,
830 lockdep_is_held(&vq->mutex));
831 if (sock != oldsock) {
832 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
838 vhost_net_disable_vq(n, vq);
839 rcu_assign_pointer(vq->private_data, sock);
840 r = vhost_init_used(vq);
843 r = vhost_net_enable_vq(n, vq);
847 oldubufs = vq->ubufs;
855 mutex_unlock(&vq->mutex);
858 vhost_ubuf_put_and_wait(oldubufs);
859 mutex_lock(&vq->mutex);
860 vhost_zerocopy_signal_used(n, vq);
861 mutex_unlock(&vq->mutex);
865 vhost_net_flush_vq(n, index);
869 mutex_unlock(&n->dev.mutex);
873 rcu_assign_pointer(vq->private_data, oldsock);
874 vhost_net_enable_vq(n, vq);
876 vhost_ubuf_put_and_wait(ubufs);
880 mutex_unlock(&vq->mutex);
882 mutex_unlock(&n->dev.mutex);
886 static long vhost_net_reset_owner(struct vhost_net *n)
888 struct socket *tx_sock = NULL;
889 struct socket *rx_sock = NULL;
892 mutex_lock(&n->dev.mutex);
893 err = vhost_dev_check_owner(&n->dev);
896 vhost_net_stop(n, &tx_sock, &rx_sock);
898 err = vhost_dev_reset_owner(&n->dev);
900 mutex_unlock(&n->dev.mutex);
908 static int vhost_net_set_features(struct vhost_net *n, u64 features)
910 size_t vhost_hlen, sock_hlen, hdr_len;
913 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
914 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
915 sizeof(struct virtio_net_hdr);
916 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
917 /* vhost provides vnet_hdr */
918 vhost_hlen = hdr_len;
921 /* socket provides vnet_hdr */
925 mutex_lock(&n->dev.mutex);
926 if ((features & (1 << VHOST_F_LOG_ALL)) &&
927 !vhost_log_access_ok(&n->dev)) {
928 mutex_unlock(&n->dev.mutex);
931 n->dev.acked_features = features;
933 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
934 mutex_lock(&n->vqs[i].mutex);
935 n->vqs[i].vhost_hlen = vhost_hlen;
936 n->vqs[i].sock_hlen = sock_hlen;
937 mutex_unlock(&n->vqs[i].mutex);
940 mutex_unlock(&n->dev.mutex);
944 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
947 struct vhost_net *n = f->private_data;
948 void __user *argp = (void __user *)arg;
949 u64 __user *featurep = argp;
950 struct vhost_vring_file backend;
955 case VHOST_NET_SET_BACKEND:
956 if (copy_from_user(&backend, argp, sizeof backend))
958 return vhost_net_set_backend(n, backend.index, backend.fd);
959 case VHOST_GET_FEATURES:
960 features = VHOST_NET_FEATURES;
961 if (copy_to_user(featurep, &features, sizeof features))
964 case VHOST_SET_FEATURES:
965 if (copy_from_user(&features, featurep, sizeof features))
967 if (features & ~VHOST_NET_FEATURES)
969 return vhost_net_set_features(n, features);
970 case VHOST_RESET_OWNER:
971 return vhost_net_reset_owner(n);
973 mutex_lock(&n->dev.mutex);
974 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
975 if (r == -ENOIOCTLCMD)
976 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
979 mutex_unlock(&n->dev.mutex);
985 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
988 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
992 static const struct file_operations vhost_net_fops = {
993 .owner = THIS_MODULE,
994 .release = vhost_net_release,
995 .unlocked_ioctl = vhost_net_ioctl,
997 .compat_ioctl = vhost_net_compat_ioctl,
999 .open = vhost_net_open,
1000 .llseek = noop_llseek,
1003 static struct miscdevice vhost_net_misc = {
1004 .minor = VHOST_NET_MINOR,
1005 .name = "vhost-net",
1006 .fops = &vhost_net_fops,
1009 static int vhost_net_init(void)
1011 if (experimental_zcopytx)
1012 vhost_enable_zcopy(VHOST_NET_VQ_TX);
1013 return misc_register(&vhost_net_misc);
1015 module_init(vhost_net_init);
1017 static void vhost_net_exit(void)
1019 misc_deregister(&vhost_net_misc);
1021 module_exit(vhost_net_exit);
1023 MODULE_VERSION("0.0.1");
1024 MODULE_LICENSE("GPL v2");
1025 MODULE_AUTHOR("Michael S. Tsirkin");
1026 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1027 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1028 MODULE_ALIAS("devname:vhost-net");