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/mmu_context.h>
14 #include <linux/miscdevice.h>
15 #include <linux/module.h>
16 #include <linux/mutex.h>
17 #include <linux/workqueue.h>
18 #include <linux/rcupdate.h>
19 #include <linux/file.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)
63 while (len && seg < iov_count) {
64 size = min(from->iov_len, len);
65 to->iov_base = from->iov_base;
67 from->iov_len -= size;
68 from->iov_base += size;
77 /* Caller must have TX VQ lock */
78 static void tx_poll_stop(struct vhost_net *net)
80 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
82 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
83 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
86 /* Caller must have TX VQ lock */
87 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
89 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
91 vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
92 net->tx_poll_state = VHOST_NET_POLL_STARTED;
95 /* Expects to be always run from workqueue - which acts as
96 * read-size critical section for our kind of RCU. */
97 static void handle_tx(struct vhost_net *net)
99 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
100 unsigned head, out, in, s;
101 struct msghdr msg = {
107 .msg_flags = MSG_DONTWAIT,
109 size_t len, total_len = 0;
112 struct socket *sock = rcu_dereference(vq->private_data);
116 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
117 if (wmem >= sock->sk->sk_sndbuf) {
118 mutex_lock(&vq->mutex);
119 tx_poll_start(net, sock);
120 mutex_unlock(&vq->mutex);
125 mutex_lock(&vq->mutex);
126 vhost_disable_notify(vq);
128 if (wmem < sock->sk->sk_sndbuf / 2)
130 hdr_size = vq->hdr_size;
133 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
137 /* Nothing new? Wait for eventfd to tell us they refilled. */
138 if (head == vq->num) {
139 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
140 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
141 tx_poll_start(net, sock);
142 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
145 if (unlikely(vhost_enable_notify(vq))) {
146 vhost_disable_notify(vq);
152 vq_err(vq, "Unexpected descriptor format for TX: "
153 "out %d, int %d\n", out, in);
156 /* Skip header. TODO: support TSO. */
157 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
158 msg.msg_iovlen = out;
159 len = iov_length(vq->iov, out);
162 vq_err(vq, "Unexpected header len for TX: "
163 "%zd expected %zd\n",
164 iov_length(vq->hdr, s), hdr_size);
167 /* TODO: Check specific error and bomb out unless ENOBUFS? */
168 err = sock->ops->sendmsg(NULL, sock, &msg, len);
169 if (unlikely(err < 0)) {
170 vhost_discard_vq_desc(vq);
171 tx_poll_start(net, sock);
175 pr_err("Truncated TX packet: "
176 " len %d != %zd\n", err, len);
177 vhost_add_used_and_signal(&net->dev, vq, head, 0);
179 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
180 vhost_poll_queue(&vq->poll);
185 mutex_unlock(&vq->mutex);
186 unuse_mm(net->dev.mm);
189 /* Expects to be always run from workqueue - which acts as
190 * read-size critical section for our kind of RCU. */
191 static void handle_rx(struct vhost_net *net)
193 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
194 unsigned head, out, in, log, s;
195 struct vhost_log *vq_log;
196 struct msghdr msg = {
199 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
202 .msg_flags = MSG_DONTWAIT,
205 struct virtio_net_hdr hdr = {
207 .gso_type = VIRTIO_NET_HDR_GSO_NONE
210 size_t len, total_len = 0;
213 struct socket *sock = rcu_dereference(vq->private_data);
214 if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
218 mutex_lock(&vq->mutex);
219 vhost_disable_notify(vq);
220 hdr_size = vq->hdr_size;
222 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
226 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
230 /* OK, now we need to know about added descriptors. */
231 if (head == vq->num) {
232 if (unlikely(vhost_enable_notify(vq))) {
233 /* They have slipped one in as we were
234 * doing that: check again. */
235 vhost_disable_notify(vq);
238 /* Nothing new? Wait for eventfd to tell us
242 /* We don't need to be notified again. */
244 vq_err(vq, "Unexpected descriptor format for RX: "
249 /* Skip header. TODO: support TSO/mergeable rx buffers. */
250 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, in);
252 len = iov_length(vq->iov, in);
255 vq_err(vq, "Unexpected header len for RX: "
256 "%zd expected %zd\n",
257 iov_length(vq->hdr, s), hdr_size);
260 err = sock->ops->recvmsg(NULL, sock, &msg,
261 len, MSG_DONTWAIT | MSG_TRUNC);
262 /* TODO: Check specific error and bomb out unless EAGAIN? */
264 vhost_discard_vq_desc(vq);
267 /* TODO: Should check and handle checksum. */
269 pr_err("Discarded truncated rx packet: "
270 " len %d > %zd\n", err, len);
271 vhost_discard_vq_desc(vq);
275 err = memcpy_toiovec(vq->hdr, (unsigned char *)&hdr, hdr_size);
277 vq_err(vq, "Unable to write vnet_hdr at addr %p: %d\n",
278 vq->iov->iov_base, err);
282 vhost_add_used_and_signal(&net->dev, vq, head, len);
283 if (unlikely(vq_log))
284 vhost_log_write(vq, vq_log, log, len);
286 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
287 vhost_poll_queue(&vq->poll);
292 mutex_unlock(&vq->mutex);
293 unuse_mm(net->dev.mm);
296 static void handle_tx_kick(struct work_struct *work)
298 struct vhost_virtqueue *vq;
299 struct vhost_net *net;
300 vq = container_of(work, struct vhost_virtqueue, poll.work);
301 net = container_of(vq->dev, struct vhost_net, dev);
305 static void handle_rx_kick(struct work_struct *work)
307 struct vhost_virtqueue *vq;
308 struct vhost_net *net;
309 vq = container_of(work, struct vhost_virtqueue, poll.work);
310 net = container_of(vq->dev, struct vhost_net, dev);
314 static void handle_tx_net(struct work_struct *work)
316 struct vhost_net *net;
317 net = container_of(work, struct vhost_net, poll[VHOST_NET_VQ_TX].work);
321 static void handle_rx_net(struct work_struct *work)
323 struct vhost_net *net;
324 net = container_of(work, struct vhost_net, poll[VHOST_NET_VQ_RX].work);
328 static int vhost_net_open(struct inode *inode, struct file *f)
330 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
334 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
335 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
336 r = vhost_dev_init(&n->dev, n->vqs, VHOST_NET_VQ_MAX);
342 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT);
343 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN);
344 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
351 static void vhost_net_disable_vq(struct vhost_net *n,
352 struct vhost_virtqueue *vq)
354 if (!vq->private_data)
356 if (vq == n->vqs + VHOST_NET_VQ_TX) {
358 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
360 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
363 static void vhost_net_enable_vq(struct vhost_net *n,
364 struct vhost_virtqueue *vq)
366 struct socket *sock = vq->private_data;
369 if (vq == n->vqs + VHOST_NET_VQ_TX) {
370 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
371 tx_poll_start(n, sock);
373 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
376 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
377 struct vhost_virtqueue *vq)
381 mutex_lock(&vq->mutex);
382 sock = vq->private_data;
383 vhost_net_disable_vq(n, vq);
384 rcu_assign_pointer(vq->private_data, NULL);
385 mutex_unlock(&vq->mutex);
389 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
390 struct socket **rx_sock)
392 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
393 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
396 static void vhost_net_flush_vq(struct vhost_net *n, int index)
398 vhost_poll_flush(n->poll + index);
399 vhost_poll_flush(&n->dev.vqs[index].poll);
402 static void vhost_net_flush(struct vhost_net *n)
404 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
405 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
408 static int vhost_net_release(struct inode *inode, struct file *f)
410 struct vhost_net *n = f->private_data;
411 struct socket *tx_sock;
412 struct socket *rx_sock;
414 vhost_net_stop(n, &tx_sock, &rx_sock);
416 vhost_dev_cleanup(&n->dev);
421 /* We do an extra flush before freeing memory,
422 * since jobs can re-queue themselves. */
428 static struct socket *get_raw_socket(int fd)
431 struct sockaddr_ll sa;
432 char buf[MAX_ADDR_LEN];
434 int uaddr_len = sizeof uaddr, r;
435 struct socket *sock = sockfd_lookup(fd, &r);
437 return ERR_PTR(-ENOTSOCK);
439 /* Parameter checking */
440 if (sock->sk->sk_type != SOCK_RAW) {
441 r = -ESOCKTNOSUPPORT;
445 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
450 if (uaddr.sa.sll_family != AF_PACKET) {
460 static struct socket *get_tap_socket(int fd)
462 struct file *file = fget(fd);
465 return ERR_PTR(-EBADF);
466 sock = tun_get_socket(file);
469 sock = macvtap_get_socket(file);
475 static struct socket *get_socket(int fd)
478 /* special case to disable backend */
481 sock = get_raw_socket(fd);
484 sock = get_tap_socket(fd);
487 return ERR_PTR(-ENOTSOCK);
490 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
492 struct socket *sock, *oldsock;
493 struct vhost_virtqueue *vq;
496 mutex_lock(&n->dev.mutex);
497 r = vhost_dev_check_owner(&n->dev);
501 if (index >= VHOST_NET_VQ_MAX) {
506 mutex_lock(&vq->mutex);
508 /* Verify that ring has been setup correctly. */
509 if (!vhost_vq_access_ok(vq)) {
513 sock = get_socket(fd);
519 /* start polling new socket */
520 oldsock = vq->private_data;
524 vhost_net_disable_vq(n, vq);
525 rcu_assign_pointer(vq->private_data, sock);
526 vhost_net_enable_vq(n, vq);
529 vhost_net_flush_vq(n, index);
534 mutex_unlock(&vq->mutex);
536 mutex_unlock(&n->dev.mutex);
540 static long vhost_net_reset_owner(struct vhost_net *n)
542 struct socket *tx_sock = NULL;
543 struct socket *rx_sock = NULL;
545 mutex_lock(&n->dev.mutex);
546 err = vhost_dev_check_owner(&n->dev);
549 vhost_net_stop(n, &tx_sock, &rx_sock);
551 err = vhost_dev_reset_owner(&n->dev);
553 mutex_unlock(&n->dev.mutex);
561 static int vhost_net_set_features(struct vhost_net *n, u64 features)
563 size_t hdr_size = features & (1 << VHOST_NET_F_VIRTIO_NET_HDR) ?
564 sizeof(struct virtio_net_hdr) : 0;
566 mutex_lock(&n->dev.mutex);
567 if ((features & (1 << VHOST_F_LOG_ALL)) &&
568 !vhost_log_access_ok(&n->dev)) {
569 mutex_unlock(&n->dev.mutex);
572 n->dev.acked_features = features;
574 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
575 mutex_lock(&n->vqs[i].mutex);
576 n->vqs[i].hdr_size = hdr_size;
577 mutex_unlock(&n->vqs[i].mutex);
580 mutex_unlock(&n->dev.mutex);
584 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
587 struct vhost_net *n = f->private_data;
588 void __user *argp = (void __user *)arg;
589 u64 __user *featurep = argp;
590 struct vhost_vring_file backend;
594 case VHOST_NET_SET_BACKEND:
595 r = copy_from_user(&backend, argp, sizeof backend);
598 return vhost_net_set_backend(n, backend.index, backend.fd);
599 case VHOST_GET_FEATURES:
600 features = VHOST_FEATURES;
601 return copy_to_user(featurep, &features, sizeof features);
602 case VHOST_SET_FEATURES:
603 r = copy_from_user(&features, featurep, sizeof features);
606 if (features & ~VHOST_FEATURES)
608 return vhost_net_set_features(n, features);
609 case VHOST_RESET_OWNER:
610 return vhost_net_reset_owner(n);
612 mutex_lock(&n->dev.mutex);
613 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
615 mutex_unlock(&n->dev.mutex);
621 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
624 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
628 const static struct file_operations vhost_net_fops = {
629 .owner = THIS_MODULE,
630 .release = vhost_net_release,
631 .unlocked_ioctl = vhost_net_ioctl,
633 .compat_ioctl = vhost_net_compat_ioctl,
635 .open = vhost_net_open,
638 static struct miscdevice vhost_net_misc = {
644 int vhost_net_init(void)
646 int r = vhost_init();
649 r = misc_register(&vhost_net_misc);
659 module_init(vhost_net_init);
661 void vhost_net_exit(void)
663 misc_deregister(&vhost_net_misc);
666 module_exit(vhost_net_exit);
668 MODULE_VERSION("0.0.1");
669 MODULE_LICENSE("GPL v2");
670 MODULE_AUTHOR("Michael S. Tsirkin");
671 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");