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>
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>
32 /* Max number of bytes transferred before requeueing the job.
33 * Using this limit prevents one virtqueue from starving others. */
34 #define VHOST_NET_WEIGHT 0x80000
42 enum vhost_net_poll_state {
43 VHOST_NET_POLL_DISABLED = 0,
44 VHOST_NET_POLL_STARTED = 1,
45 VHOST_NET_POLL_STOPPED = 2,
50 struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
51 struct vhost_poll poll[VHOST_NET_VQ_MAX];
52 /* Tells us whether we are polling a socket for TX.
53 * We only do this when socket buffer fills up.
54 * Protected by tx vq lock. */
55 enum vhost_net_poll_state tx_poll_state;
58 /* Pop first len bytes from iovec. Return number of segments used. */
59 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
60 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)
83 while (len && seg < iovcount) {
84 size = min(from->iov_len, len);
85 to->iov_base = from->iov_base;
94 /* Caller must have TX VQ lock */
95 static void tx_poll_stop(struct vhost_net *net)
97 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
99 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
100 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
103 /* Caller must have TX VQ lock */
104 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
106 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
108 vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
109 net->tx_poll_state = VHOST_NET_POLL_STARTED;
112 /* Expects to be always run from workqueue - which acts as
113 * read-size critical section for our kind of RCU. */
114 static void handle_tx(struct vhost_net *net)
116 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
119 struct msghdr msg = {
125 .msg_flags = MSG_DONTWAIT,
127 size_t len, total_len = 0;
132 sock = rcu_dereference_check(vq->private_data,
133 lockdep_is_held(&vq->mutex));
137 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
138 if (wmem >= sock->sk->sk_sndbuf) {
139 mutex_lock(&vq->mutex);
140 tx_poll_start(net, sock);
141 mutex_unlock(&vq->mutex);
146 mutex_lock(&vq->mutex);
147 vhost_disable_notify(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(vq))) {
170 vhost_disable_notify(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);
210 unuse_mm(net->dev.mm);
213 static int peek_head_len(struct sock *sk)
215 struct sk_buff *head;
219 head = skb_peek(&sk->sk_receive_queue);
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 * returns number of buffer heads allocated, negative on error
235 static int get_rx_bufs(struct vhost_virtqueue *vq,
236 struct vring_used_elem *heads,
239 struct vhost_log *log,
242 unsigned int out, in;
248 while (datalen > 0) {
249 if (unlikely(seg >= VHOST_NET_MAX_SG)) {
253 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
254 ARRAY_SIZE(vq->iov) - seg, &out,
260 if (unlikely(out || in <= 0)) {
261 vq_err(vq, "unexpected descriptor format for RX: "
262 "out %d, in %d\n", out, in);
270 heads[headcount].id = d;
271 heads[headcount].len = iov_length(vq->iov + seg, in);
272 datalen -= heads[headcount].len;
276 heads[headcount - 1].len += datalen;
282 vhost_discard_vq_desc(vq, headcount);
286 /* Expects to be always run from workqueue - which acts as
287 * read-size critical section for our kind of RCU. */
288 static void handle_rx_big(struct vhost_net *net)
290 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
291 unsigned out, in, log, s;
293 struct vhost_log *vq_log;
294 struct msghdr msg = {
297 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
300 .msg_flags = MSG_DONTWAIT,
303 struct virtio_net_hdr hdr = {
305 .gso_type = VIRTIO_NET_HDR_GSO_NONE
308 size_t len, total_len = 0;
311 struct socket *sock = rcu_dereference(vq->private_data);
312 if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
316 mutex_lock(&vq->mutex);
317 vhost_disable_notify(vq);
318 hdr_size = vq->vhost_hlen;
320 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
324 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
328 /* On error, stop handling until the next kick. */
329 if (unlikely(head < 0))
331 /* OK, now we need to know about added descriptors. */
332 if (head == vq->num) {
333 if (unlikely(vhost_enable_notify(vq))) {
334 /* They have slipped one in as we were
335 * doing that: check again. */
336 vhost_disable_notify(vq);
339 /* Nothing new? Wait for eventfd to tell us
343 /* We don't need to be notified again. */
345 vq_err(vq, "Unexpected descriptor format for RX: "
350 /* Skip header. TODO: support TSO/mergeable rx buffers. */
351 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, in);
353 len = iov_length(vq->iov, in);
356 vq_err(vq, "Unexpected header len for RX: "
357 "%zd expected %zd\n",
358 iov_length(vq->hdr, s), hdr_size);
361 err = sock->ops->recvmsg(NULL, sock, &msg,
362 len, MSG_DONTWAIT | MSG_TRUNC);
363 /* TODO: Check specific error and bomb out unless EAGAIN? */
365 vhost_discard_vq_desc(vq, 1);
368 /* TODO: Should check and handle checksum. */
370 pr_debug("Discarded truncated rx packet: "
371 " len %d > %zd\n", err, len);
372 vhost_discard_vq_desc(vq, 1);
376 err = memcpy_toiovec(vq->hdr, (unsigned char *)&hdr, hdr_size);
378 vq_err(vq, "Unable to write vnet_hdr at addr %p: %d\n",
379 vq->iov->iov_base, err);
383 vhost_add_used_and_signal(&net->dev, vq, head, len);
384 if (unlikely(vq_log))
385 vhost_log_write(vq, vq_log, log, len);
387 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
388 vhost_poll_queue(&vq->poll);
393 mutex_unlock(&vq->mutex);
394 unuse_mm(net->dev.mm);
397 /* Expects to be always run from workqueue - which acts as
398 * read-size critical section for our kind of RCU. */
399 static void handle_rx_mergeable(struct vhost_net *net)
401 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
402 unsigned uninitialized_var(in), log;
403 struct vhost_log *vq_log;
404 struct msghdr msg = {
407 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
410 .msg_flags = MSG_DONTWAIT,
413 struct virtio_net_hdr_mrg_rxbuf hdr = {
415 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
418 size_t total_len = 0;
420 size_t vhost_hlen, sock_hlen;
421 size_t vhost_len, sock_len;
422 struct socket *sock = rcu_dereference(vq->private_data);
423 if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
427 mutex_lock(&vq->mutex);
428 vhost_disable_notify(vq);
429 vhost_hlen = vq->vhost_hlen;
430 sock_hlen = vq->sock_hlen;
432 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
435 while ((sock_len = peek_head_len(sock->sk))) {
436 sock_len += sock_hlen;
437 vhost_len = sock_len + vhost_hlen;
438 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
440 /* On error, stop handling until the next kick. */
441 if (unlikely(headcount < 0))
443 /* OK, now we need to know about added descriptors. */
445 if (unlikely(vhost_enable_notify(vq))) {
446 /* They have slipped one in as we were
447 * doing that: check again. */
448 vhost_disable_notify(vq);
451 /* Nothing new? Wait for eventfd to tell us
455 /* We don't need to be notified again. */
456 if (unlikely((vhost_hlen)))
457 /* Skip header. TODO: support TSO. */
458 move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
460 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
461 * needed because sendmsg can modify msg_iov. */
462 copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
464 err = sock->ops->recvmsg(NULL, sock, &msg,
465 sock_len, MSG_DONTWAIT | MSG_TRUNC);
466 /* Userspace might have consumed the packet meanwhile:
467 * it's not supposed to do this usually, but might be hard
468 * to prevent. Discard data we got (if any) and keep going. */
469 if (unlikely(err != sock_len)) {
470 pr_debug("Discarded rx packet: "
471 " len %d, expected %zd\n", err, sock_len);
472 vhost_discard_vq_desc(vq, headcount);
475 if (unlikely(vhost_hlen) &&
476 memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
478 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
482 /* TODO: Should check and handle checksum. */
483 if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF) &&
484 memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
485 offsetof(typeof(hdr), num_buffers),
486 sizeof hdr.num_buffers)) {
487 vq_err(vq, "Failed num_buffers write");
488 vhost_discard_vq_desc(vq, headcount);
491 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
493 if (unlikely(vq_log))
494 vhost_log_write(vq, vq_log, log, vhost_len);
495 total_len += vhost_len;
496 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
497 vhost_poll_queue(&vq->poll);
502 mutex_unlock(&vq->mutex);
503 unuse_mm(net->dev.mm);
506 static void handle_rx(struct vhost_net *net)
508 if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF))
509 handle_rx_mergeable(net);
514 static void handle_tx_kick(struct vhost_work *work)
516 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
518 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
523 static void handle_rx_kick(struct vhost_work *work)
525 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
527 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
532 static void handle_tx_net(struct vhost_work *work)
534 struct vhost_net *net = container_of(work, struct vhost_net,
535 poll[VHOST_NET_VQ_TX].work);
539 static void handle_rx_net(struct vhost_work *work)
541 struct vhost_net *net = container_of(work, struct vhost_net,
542 poll[VHOST_NET_VQ_RX].work);
546 static int vhost_net_open(struct inode *inode, struct file *f)
548 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
549 struct vhost_dev *dev;
556 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
557 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
558 r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
564 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
565 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
566 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
573 static void vhost_net_disable_vq(struct vhost_net *n,
574 struct vhost_virtqueue *vq)
576 if (!vq->private_data)
578 if (vq == n->vqs + VHOST_NET_VQ_TX) {
580 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
582 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
585 static void vhost_net_enable_vq(struct vhost_net *n,
586 struct vhost_virtqueue *vq)
590 sock = rcu_dereference_protected(vq->private_data,
591 lockdep_is_held(&vq->mutex));
594 if (vq == n->vqs + VHOST_NET_VQ_TX) {
595 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
596 tx_poll_start(n, sock);
598 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
601 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
602 struct vhost_virtqueue *vq)
606 mutex_lock(&vq->mutex);
607 sock = rcu_dereference_protected(vq->private_data,
608 lockdep_is_held(&vq->mutex));
609 vhost_net_disable_vq(n, vq);
610 rcu_assign_pointer(vq->private_data, NULL);
611 mutex_unlock(&vq->mutex);
615 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
616 struct socket **rx_sock)
618 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
619 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
622 static void vhost_net_flush_vq(struct vhost_net *n, int index)
624 vhost_poll_flush(n->poll + index);
625 vhost_poll_flush(&n->dev.vqs[index].poll);
628 static void vhost_net_flush(struct vhost_net *n)
630 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
631 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
634 static int vhost_net_release(struct inode *inode, struct file *f)
636 struct vhost_net *n = f->private_data;
637 struct socket *tx_sock;
638 struct socket *rx_sock;
640 vhost_net_stop(n, &tx_sock, &rx_sock);
642 vhost_dev_cleanup(&n->dev);
647 /* We do an extra flush before freeing memory,
648 * since jobs can re-queue themselves. */
654 static struct socket *get_raw_socket(int fd)
657 struct sockaddr_ll sa;
658 char buf[MAX_ADDR_LEN];
660 int uaddr_len = sizeof uaddr, r;
661 struct socket *sock = sockfd_lookup(fd, &r);
663 return ERR_PTR(-ENOTSOCK);
665 /* Parameter checking */
666 if (sock->sk->sk_type != SOCK_RAW) {
667 r = -ESOCKTNOSUPPORT;
671 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
676 if (uaddr.sa.sll_family != AF_PACKET) {
686 static struct socket *get_tap_socket(int fd)
688 struct file *file = fget(fd);
691 return ERR_PTR(-EBADF);
692 sock = tun_get_socket(file);
695 sock = macvtap_get_socket(file);
701 static struct socket *get_socket(int fd)
704 /* special case to disable backend */
707 sock = get_raw_socket(fd);
710 sock = get_tap_socket(fd);
713 return ERR_PTR(-ENOTSOCK);
716 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
718 struct socket *sock, *oldsock;
719 struct vhost_virtqueue *vq;
722 mutex_lock(&n->dev.mutex);
723 r = vhost_dev_check_owner(&n->dev);
727 if (index >= VHOST_NET_VQ_MAX) {
732 mutex_lock(&vq->mutex);
734 /* Verify that ring has been setup correctly. */
735 if (!vhost_vq_access_ok(vq)) {
739 sock = get_socket(fd);
745 /* start polling new socket */
746 oldsock = rcu_dereference_protected(vq->private_data,
747 lockdep_is_held(&vq->mutex));
748 if (sock != oldsock) {
749 vhost_net_disable_vq(n, vq);
750 rcu_assign_pointer(vq->private_data, sock);
751 vhost_net_enable_vq(n, vq);
754 mutex_unlock(&vq->mutex);
757 vhost_net_flush_vq(n, index);
761 mutex_unlock(&n->dev.mutex);
765 mutex_unlock(&vq->mutex);
767 mutex_unlock(&n->dev.mutex);
771 static long vhost_net_reset_owner(struct vhost_net *n)
773 struct socket *tx_sock = NULL;
774 struct socket *rx_sock = NULL;
776 mutex_lock(&n->dev.mutex);
777 err = vhost_dev_check_owner(&n->dev);
780 vhost_net_stop(n, &tx_sock, &rx_sock);
782 err = vhost_dev_reset_owner(&n->dev);
784 mutex_unlock(&n->dev.mutex);
792 static int vhost_net_set_features(struct vhost_net *n, u64 features)
794 size_t vhost_hlen, sock_hlen, hdr_len;
797 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
798 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
799 sizeof(struct virtio_net_hdr);
800 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
801 /* vhost provides vnet_hdr */
802 vhost_hlen = hdr_len;
805 /* socket provides vnet_hdr */
809 mutex_lock(&n->dev.mutex);
810 if ((features & (1 << VHOST_F_LOG_ALL)) &&
811 !vhost_log_access_ok(&n->dev)) {
812 mutex_unlock(&n->dev.mutex);
815 n->dev.acked_features = features;
817 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
818 mutex_lock(&n->vqs[i].mutex);
819 n->vqs[i].vhost_hlen = vhost_hlen;
820 n->vqs[i].sock_hlen = sock_hlen;
821 mutex_unlock(&n->vqs[i].mutex);
824 mutex_unlock(&n->dev.mutex);
828 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
831 struct vhost_net *n = f->private_data;
832 void __user *argp = (void __user *)arg;
833 u64 __user *featurep = argp;
834 struct vhost_vring_file backend;
838 case VHOST_NET_SET_BACKEND:
839 if (copy_from_user(&backend, argp, sizeof backend))
841 return vhost_net_set_backend(n, backend.index, backend.fd);
842 case VHOST_GET_FEATURES:
843 features = VHOST_FEATURES;
844 if (copy_to_user(featurep, &features, sizeof features))
847 case VHOST_SET_FEATURES:
848 if (copy_from_user(&features, featurep, sizeof features))
850 if (features & ~VHOST_FEATURES)
852 return vhost_net_set_features(n, features);
853 case VHOST_RESET_OWNER:
854 return vhost_net_reset_owner(n);
856 mutex_lock(&n->dev.mutex);
857 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
859 mutex_unlock(&n->dev.mutex);
865 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
868 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
872 static const struct file_operations vhost_net_fops = {
873 .owner = THIS_MODULE,
874 .release = vhost_net_release,
875 .unlocked_ioctl = vhost_net_ioctl,
877 .compat_ioctl = vhost_net_compat_ioctl,
879 .open = vhost_net_open,
880 .llseek = noop_llseek,
883 static struct miscdevice vhost_net_misc = {
889 static int vhost_net_init(void)
891 return misc_register(&vhost_net_misc);
893 module_init(vhost_net_init);
895 static void vhost_net_exit(void)
897 misc_deregister(&vhost_net_misc);
899 module_exit(vhost_net_exit);
901 MODULE_VERSION("0.0.1");
902 MODULE_LICENSE("GPL v2");
903 MODULE_AUTHOR("Michael S. Tsirkin");
904 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");