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/mutex.h>
16 #include <linux/workqueue.h>
17 #include <linux/rcupdate.h>
18 #include <linux/file.h>
19 #include <linux/slab.h>
21 #include <linux/net.h>
22 #include <linux/if_packet.h>
23 #include <linux/if_arp.h>
24 #include <linux/if_tun.h>
25 #include <linux/if_macvlan.h>
31 /* Max number of bytes transferred before requeueing the job.
32 * Using this limit prevents one virtqueue from starving others. */
33 #define VHOST_NET_WEIGHT 0x80000
41 enum vhost_net_poll_state {
42 VHOST_NET_POLL_DISABLED = 0,
43 VHOST_NET_POLL_STARTED = 1,
44 VHOST_NET_POLL_STOPPED = 2,
49 struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
50 struct vhost_poll poll[VHOST_NET_VQ_MAX];
51 /* Tells us whether we are polling a socket for TX.
52 * We only do this when socket buffer fills up.
53 * Protected by tx vq lock. */
54 enum vhost_net_poll_state tx_poll_state;
57 /* Pop first len bytes from iovec. Return number of segments used. */
58 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
59 size_t len, int iov_count)
64 while (len && seg < iov_count) {
65 size = min(from->iov_len, len);
66 to->iov_base = from->iov_base;
68 from->iov_len -= size;
69 from->iov_base += size;
77 /* Copy iovec entries for len bytes from iovec. */
78 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
79 size_t len, int iovcount)
84 while (len && seg < iovcount) {
85 size = min(from->iov_len, len);
86 to->iov_base = from->iov_base;
95 /* Caller must have TX VQ lock */
96 static void tx_poll_stop(struct vhost_net *net)
98 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
100 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
101 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
104 /* Caller must have TX VQ lock */
105 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
107 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
109 vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
110 net->tx_poll_state = VHOST_NET_POLL_STARTED;
113 /* Expects to be always run from workqueue - which acts as
114 * read-size critical section for our kind of RCU. */
115 static void handle_tx(struct vhost_net *net)
117 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
120 struct msghdr msg = {
126 .msg_flags = MSG_DONTWAIT,
128 size_t len, total_len = 0;
133 /* TODO: check that we are running from vhost_worker? */
134 sock = rcu_dereference_check(vq->private_data, 1);
138 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
139 if (wmem >= sock->sk->sk_sndbuf) {
140 mutex_lock(&vq->mutex);
141 tx_poll_start(net, sock);
142 mutex_unlock(&vq->mutex);
146 mutex_lock(&vq->mutex);
147 vhost_disable_notify(&net->dev, vq);
149 if (wmem < sock->sk->sk_sndbuf / 2)
151 hdr_size = vq->vhost_hlen;
154 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
158 /* On error, stop handling until the next kick. */
159 if (unlikely(head < 0))
161 /* Nothing new? Wait for eventfd to tell us they refilled. */
162 if (head == vq->num) {
163 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
164 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
165 tx_poll_start(net, sock);
166 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
169 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
170 vhost_disable_notify(&net->dev, vq);
176 vq_err(vq, "Unexpected descriptor format for TX: "
177 "out %d, int %d\n", out, in);
180 /* Skip header. TODO: support TSO. */
181 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
182 msg.msg_iovlen = out;
183 len = iov_length(vq->iov, out);
186 vq_err(vq, "Unexpected header len for TX: "
187 "%zd expected %zd\n",
188 iov_length(vq->hdr, s), hdr_size);
191 /* TODO: Check specific error and bomb out unless ENOBUFS? */
192 err = sock->ops->sendmsg(NULL, sock, &msg, len);
193 if (unlikely(err < 0)) {
194 vhost_discard_vq_desc(vq, 1);
195 tx_poll_start(net, sock);
199 pr_debug("Truncated TX packet: "
200 " len %d != %zd\n", err, len);
201 vhost_add_used_and_signal(&net->dev, vq, head, 0);
203 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
204 vhost_poll_queue(&vq->poll);
209 mutex_unlock(&vq->mutex);
212 static int peek_head_len(struct sock *sk)
214 struct sk_buff *head;
218 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
219 head = skb_peek(&sk->sk_receive_queue);
222 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
226 /* This is a multi-buffer version of vhost_get_desc, that works if
227 * vq has read descriptors only.
228 * @vq - the relevant virtqueue
229 * @datalen - data length we'll be reading
230 * @iovcount - returned count of io vectors we fill
232 * @log_num - log offset
233 * @quota - headcount quota, 1 for big buffer
234 * returns number of buffer heads allocated, negative on error
236 static int get_rx_bufs(struct vhost_virtqueue *vq,
237 struct vring_used_elem *heads,
240 struct vhost_log *log,
244 unsigned int out, in;
250 while (datalen > 0 && headcount < quota) {
251 if (unlikely(seg >= UIO_MAXIOV)) {
255 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
256 ARRAY_SIZE(vq->iov) - seg, &out,
262 if (unlikely(out || in <= 0)) {
263 vq_err(vq, "unexpected descriptor format for RX: "
264 "out %d, in %d\n", out, in);
272 heads[headcount].id = d;
273 heads[headcount].len = iov_length(vq->iov + seg, in);
274 datalen -= heads[headcount].len;
278 heads[headcount - 1].len += datalen;
284 vhost_discard_vq_desc(vq, headcount);
288 /* Expects to be always run from workqueue - which acts as
289 * read-size critical section for our kind of RCU. */
290 static void handle_rx(struct vhost_net *net)
292 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
293 unsigned uninitialized_var(in), log;
294 struct vhost_log *vq_log;
295 struct msghdr msg = {
298 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
301 .msg_flags = MSG_DONTWAIT,
303 struct virtio_net_hdr_mrg_rxbuf hdr = {
305 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
307 size_t total_len = 0;
308 int err, headcount, mergeable;
309 size_t vhost_hlen, sock_hlen;
310 size_t vhost_len, sock_len;
311 /* TODO: check that we are running from vhost_worker? */
312 struct socket *sock = rcu_dereference_check(vq->private_data, 1);
317 mutex_lock(&vq->mutex);
318 vhost_disable_notify(&net->dev, vq);
319 vhost_hlen = vq->vhost_hlen;
320 sock_hlen = vq->sock_hlen;
322 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
324 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
326 while ((sock_len = peek_head_len(sock->sk))) {
327 sock_len += sock_hlen;
328 vhost_len = sock_len + vhost_hlen;
329 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
331 likely(mergeable) ? UIO_MAXIOV : 1);
332 /* On error, stop handling until the next kick. */
333 if (unlikely(headcount < 0))
335 /* OK, now we need to know about added descriptors. */
337 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
338 /* They have slipped one in as we were
339 * doing that: check again. */
340 vhost_disable_notify(&net->dev, vq);
343 /* Nothing new? Wait for eventfd to tell us
347 /* We don't need to be notified again. */
348 if (unlikely((vhost_hlen)))
349 /* Skip header. TODO: support TSO. */
350 move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
352 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
353 * needed because recvmsg can modify msg_iov. */
354 copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
356 err = sock->ops->recvmsg(NULL, sock, &msg,
357 sock_len, MSG_DONTWAIT | MSG_TRUNC);
358 /* Userspace might have consumed the packet meanwhile:
359 * it's not supposed to do this usually, but might be hard
360 * to prevent. Discard data we got (if any) and keep going. */
361 if (unlikely(err != sock_len)) {
362 pr_debug("Discarded rx packet: "
363 " len %d, expected %zd\n", err, sock_len);
364 vhost_discard_vq_desc(vq, headcount);
367 if (unlikely(vhost_hlen) &&
368 memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
370 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
374 /* TODO: Should check and handle checksum. */
375 if (likely(mergeable) &&
376 memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
377 offsetof(typeof(hdr), num_buffers),
378 sizeof hdr.num_buffers)) {
379 vq_err(vq, "Failed num_buffers write");
380 vhost_discard_vq_desc(vq, headcount);
383 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
385 if (unlikely(vq_log))
386 vhost_log_write(vq, vq_log, log, vhost_len);
387 total_len += vhost_len;
388 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
389 vhost_poll_queue(&vq->poll);
394 mutex_unlock(&vq->mutex);
397 static void handle_tx_kick(struct vhost_work *work)
399 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
401 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
406 static void handle_rx_kick(struct vhost_work *work)
408 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
410 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
415 static void handle_tx_net(struct vhost_work *work)
417 struct vhost_net *net = container_of(work, struct vhost_net,
418 poll[VHOST_NET_VQ_TX].work);
422 static void handle_rx_net(struct vhost_work *work)
424 struct vhost_net *net = container_of(work, struct vhost_net,
425 poll[VHOST_NET_VQ_RX].work);
429 static int vhost_net_open(struct inode *inode, struct file *f)
431 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
432 struct vhost_dev *dev;
439 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
440 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
441 r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
447 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
448 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
449 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
456 static void vhost_net_disable_vq(struct vhost_net *n,
457 struct vhost_virtqueue *vq)
459 if (!vq->private_data)
461 if (vq == n->vqs + VHOST_NET_VQ_TX) {
463 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
465 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
468 static void vhost_net_enable_vq(struct vhost_net *n,
469 struct vhost_virtqueue *vq)
473 sock = rcu_dereference_protected(vq->private_data,
474 lockdep_is_held(&vq->mutex));
477 if (vq == n->vqs + VHOST_NET_VQ_TX) {
478 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
479 tx_poll_start(n, sock);
481 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
484 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
485 struct vhost_virtqueue *vq)
489 mutex_lock(&vq->mutex);
490 sock = rcu_dereference_protected(vq->private_data,
491 lockdep_is_held(&vq->mutex));
492 vhost_net_disable_vq(n, vq);
493 rcu_assign_pointer(vq->private_data, NULL);
494 mutex_unlock(&vq->mutex);
498 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
499 struct socket **rx_sock)
501 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
502 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
505 static void vhost_net_flush_vq(struct vhost_net *n, int index)
507 vhost_poll_flush(n->poll + index);
508 vhost_poll_flush(&n->dev.vqs[index].poll);
511 static void vhost_net_flush(struct vhost_net *n)
513 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
514 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
517 static int vhost_net_release(struct inode *inode, struct file *f)
519 struct vhost_net *n = f->private_data;
520 struct socket *tx_sock;
521 struct socket *rx_sock;
523 vhost_net_stop(n, &tx_sock, &rx_sock);
525 vhost_dev_cleanup(&n->dev);
530 /* We do an extra flush before freeing memory,
531 * since jobs can re-queue themselves. */
537 static struct socket *get_raw_socket(int fd)
540 struct sockaddr_ll sa;
541 char buf[MAX_ADDR_LEN];
543 int uaddr_len = sizeof uaddr, r;
544 struct socket *sock = sockfd_lookup(fd, &r);
547 return ERR_PTR(-ENOTSOCK);
549 /* Parameter checking */
550 if (sock->sk->sk_type != SOCK_RAW) {
551 r = -ESOCKTNOSUPPORT;
555 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
560 if (uaddr.sa.sll_family != AF_PACKET) {
570 static struct socket *get_tap_socket(int fd)
572 struct file *file = fget(fd);
576 return ERR_PTR(-EBADF);
577 sock = tun_get_socket(file);
580 sock = macvtap_get_socket(file);
586 static struct socket *get_socket(int fd)
590 /* special case to disable backend */
593 sock = get_raw_socket(fd);
596 sock = get_tap_socket(fd);
599 return ERR_PTR(-ENOTSOCK);
602 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
604 struct socket *sock, *oldsock;
605 struct vhost_virtqueue *vq;
608 mutex_lock(&n->dev.mutex);
609 r = vhost_dev_check_owner(&n->dev);
613 if (index >= VHOST_NET_VQ_MAX) {
618 mutex_lock(&vq->mutex);
620 /* Verify that ring has been setup correctly. */
621 if (!vhost_vq_access_ok(vq)) {
625 sock = get_socket(fd);
631 /* start polling new socket */
632 oldsock = rcu_dereference_protected(vq->private_data,
633 lockdep_is_held(&vq->mutex));
634 if (sock != oldsock) {
635 vhost_net_disable_vq(n, vq);
636 rcu_assign_pointer(vq->private_data, sock);
637 vhost_net_enable_vq(n, vq);
640 mutex_unlock(&vq->mutex);
643 vhost_net_flush_vq(n, index);
647 mutex_unlock(&n->dev.mutex);
651 mutex_unlock(&vq->mutex);
653 mutex_unlock(&n->dev.mutex);
657 static long vhost_net_reset_owner(struct vhost_net *n)
659 struct socket *tx_sock = NULL;
660 struct socket *rx_sock = NULL;
663 mutex_lock(&n->dev.mutex);
664 err = vhost_dev_check_owner(&n->dev);
667 vhost_net_stop(n, &tx_sock, &rx_sock);
669 err = vhost_dev_reset_owner(&n->dev);
671 mutex_unlock(&n->dev.mutex);
679 static int vhost_net_set_features(struct vhost_net *n, u64 features)
681 size_t vhost_hlen, sock_hlen, hdr_len;
684 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
685 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
686 sizeof(struct virtio_net_hdr);
687 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
688 /* vhost provides vnet_hdr */
689 vhost_hlen = hdr_len;
692 /* socket provides vnet_hdr */
696 mutex_lock(&n->dev.mutex);
697 if ((features & (1 << VHOST_F_LOG_ALL)) &&
698 !vhost_log_access_ok(&n->dev)) {
699 mutex_unlock(&n->dev.mutex);
702 n->dev.acked_features = features;
704 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
705 mutex_lock(&n->vqs[i].mutex);
706 n->vqs[i].vhost_hlen = vhost_hlen;
707 n->vqs[i].sock_hlen = sock_hlen;
708 mutex_unlock(&n->vqs[i].mutex);
711 mutex_unlock(&n->dev.mutex);
715 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
718 struct vhost_net *n = f->private_data;
719 void __user *argp = (void __user *)arg;
720 u64 __user *featurep = argp;
721 struct vhost_vring_file backend;
726 case VHOST_NET_SET_BACKEND:
727 if (copy_from_user(&backend, argp, sizeof backend))
729 return vhost_net_set_backend(n, backend.index, backend.fd);
730 case VHOST_GET_FEATURES:
731 features = VHOST_FEATURES;
732 if (copy_to_user(featurep, &features, sizeof features))
735 case VHOST_SET_FEATURES:
736 if (copy_from_user(&features, featurep, sizeof features))
738 if (features & ~VHOST_FEATURES)
740 return vhost_net_set_features(n, features);
741 case VHOST_RESET_OWNER:
742 return vhost_net_reset_owner(n);
744 mutex_lock(&n->dev.mutex);
745 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
747 mutex_unlock(&n->dev.mutex);
753 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
756 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
760 static const struct file_operations vhost_net_fops = {
761 .owner = THIS_MODULE,
762 .release = vhost_net_release,
763 .unlocked_ioctl = vhost_net_ioctl,
765 .compat_ioctl = vhost_net_compat_ioctl,
767 .open = vhost_net_open,
768 .llseek = noop_llseek,
771 static struct miscdevice vhost_net_misc = {
777 static int vhost_net_init(void)
779 return misc_register(&vhost_net_misc);
781 module_init(vhost_net_init);
783 static void vhost_net_exit(void)
785 misc_deregister(&vhost_net_misc);
787 module_exit(vhost_net_exit);
789 MODULE_VERSION("0.0.1");
790 MODULE_LICENSE("GPL v2");
791 MODULE_AUTHOR("Michael S. Tsirkin");
792 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");