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>
32 static int experimental_zcopytx;
33 module_param(experimental_zcopytx, int, 0444);
34 MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
36 /* Max number of bytes transferred before requeueing the job.
37 * Using this limit prevents one virtqueue from starving others. */
38 #define VHOST_NET_WEIGHT 0x80000
40 /* MAX number of TX used buffers for outstanding zerocopy */
41 #define VHOST_MAX_PEND 128
42 #define VHOST_GOODCOPY_LEN 256
50 enum vhost_net_poll_state {
51 VHOST_NET_POLL_DISABLED = 0,
52 VHOST_NET_POLL_STARTED = 1,
53 VHOST_NET_POLL_STOPPED = 2,
58 struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
59 struct vhost_poll poll[VHOST_NET_VQ_MAX];
60 /* Tells us whether we are polling a socket for TX.
61 * We only do this when socket buffer fills up.
62 * Protected by tx vq lock. */
63 enum vhost_net_poll_state tx_poll_state;
66 static bool vhost_sock_zcopy(struct socket *sock)
68 return unlikely(experimental_zcopytx) &&
69 sock_flag(sock->sk, SOCK_ZEROCOPY);
72 /* Pop first len bytes from iovec. Return number of segments used. */
73 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
74 size_t len, int iov_count)
79 while (len && seg < iov_count) {
80 size = min(from->iov_len, len);
81 to->iov_base = from->iov_base;
83 from->iov_len -= size;
84 from->iov_base += size;
92 /* Copy iovec entries for len bytes from iovec. */
93 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
94 size_t len, int iovcount)
99 while (len && seg < iovcount) {
100 size = min(from->iov_len, len);
101 to->iov_base = from->iov_base;
110 /* Caller must have TX VQ lock */
111 static void tx_poll_stop(struct vhost_net *net)
113 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
115 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
116 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
119 /* Caller must have TX VQ lock */
120 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
122 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
124 vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
125 net->tx_poll_state = VHOST_NET_POLL_STARTED;
128 /* Expects to be always run from workqueue - which acts as
129 * read-size critical section for our kind of RCU. */
130 static void handle_tx(struct vhost_net *net)
132 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
135 struct msghdr msg = {
141 .msg_flags = MSG_DONTWAIT,
143 size_t len, total_len = 0;
147 struct vhost_ubuf_ref *uninitialized_var(ubufs);
150 /* TODO: check that we are running from vhost_worker? */
151 sock = rcu_dereference_check(vq->private_data, 1);
155 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
156 if (wmem >= sock->sk->sk_sndbuf) {
157 mutex_lock(&vq->mutex);
158 tx_poll_start(net, sock);
159 mutex_unlock(&vq->mutex);
163 mutex_lock(&vq->mutex);
164 vhost_disable_notify(&net->dev, vq);
166 if (wmem < sock->sk->sk_sndbuf / 2)
168 hdr_size = vq->vhost_hlen;
169 zcopy = vhost_sock_zcopy(sock);
172 /* Release DMAs done buffers first */
174 vhost_zerocopy_signal_used(vq);
176 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
180 /* On error, stop handling until the next kick. */
181 if (unlikely(head < 0))
183 /* Nothing new? Wait for eventfd to tell us they refilled. */
184 if (head == vq->num) {
185 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
186 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
187 tx_poll_start(net, sock);
188 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
191 /* If more outstanding DMAs, queue the work */
192 if (unlikely(vq->upend_idx - vq->done_idx >
194 tx_poll_start(net, sock);
195 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
198 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
199 vhost_disable_notify(&net->dev, vq);
205 vq_err(vq, "Unexpected descriptor format for TX: "
206 "out %d, int %d\n", out, in);
209 /* Skip header. TODO: support TSO. */
210 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
211 msg.msg_iovlen = out;
212 len = iov_length(vq->iov, out);
215 vq_err(vq, "Unexpected header len for TX: "
216 "%zd expected %zd\n",
217 iov_length(vq->hdr, s), hdr_size);
220 /* use msg_control to pass vhost zerocopy ubuf info to skb */
222 vq->heads[vq->upend_idx].id = head;
223 if (len < VHOST_GOODCOPY_LEN) {
224 /* copy don't need to wait for DMA done */
225 vq->heads[vq->upend_idx].len =
227 msg.msg_control = NULL;
228 msg.msg_controllen = 0;
231 struct ubuf_info *ubuf = &vq->ubuf_info[head];
233 vq->heads[vq->upend_idx].len = len;
234 ubuf->callback = vhost_zerocopy_callback;
235 ubuf->arg = vq->ubufs;
236 ubuf->desc = vq->upend_idx;
237 msg.msg_control = ubuf;
238 msg.msg_controllen = sizeof(ubuf);
240 kref_get(&ubufs->kref);
242 vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
244 /* TODO: Check specific error and bomb out unless ENOBUFS? */
245 err = sock->ops->sendmsg(NULL, sock, &msg, len);
246 if (unlikely(err < 0)) {
249 vhost_ubuf_put(ubufs);
250 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
253 vhost_discard_vq_desc(vq, 1);
254 tx_poll_start(net, sock);
258 pr_debug("Truncated TX packet: "
259 " len %d != %zd\n", err, len);
261 vhost_add_used_and_signal(&net->dev, vq, head, 0);
263 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
264 vhost_poll_queue(&vq->poll);
269 mutex_unlock(&vq->mutex);
272 static int peek_head_len(struct sock *sk)
274 struct sk_buff *head;
278 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
279 head = skb_peek(&sk->sk_receive_queue);
282 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
286 /* This is a multi-buffer version of vhost_get_desc, that works if
287 * vq has read descriptors only.
288 * @vq - the relevant virtqueue
289 * @datalen - data length we'll be reading
290 * @iovcount - returned count of io vectors we fill
292 * @log_num - log offset
293 * @quota - headcount quota, 1 for big buffer
294 * returns number of buffer heads allocated, negative on error
296 static int get_rx_bufs(struct vhost_virtqueue *vq,
297 struct vring_used_elem *heads,
300 struct vhost_log *log,
304 unsigned int out, in;
310 while (datalen > 0 && headcount < quota) {
311 if (unlikely(seg >= UIO_MAXIOV)) {
315 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
316 ARRAY_SIZE(vq->iov) - seg, &out,
322 if (unlikely(out || in <= 0)) {
323 vq_err(vq, "unexpected descriptor format for RX: "
324 "out %d, in %d\n", out, in);
332 heads[headcount].id = d;
333 heads[headcount].len = iov_length(vq->iov + seg, in);
334 datalen -= heads[headcount].len;
338 heads[headcount - 1].len += datalen;
344 vhost_discard_vq_desc(vq, headcount);
348 /* Expects to be always run from workqueue - which acts as
349 * read-size critical section for our kind of RCU. */
350 static void handle_rx(struct vhost_net *net)
352 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
353 unsigned uninitialized_var(in), log;
354 struct vhost_log *vq_log;
355 struct msghdr msg = {
358 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
361 .msg_flags = MSG_DONTWAIT,
363 struct virtio_net_hdr_mrg_rxbuf hdr = {
365 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
367 size_t total_len = 0;
368 int err, headcount, mergeable;
369 size_t vhost_hlen, sock_hlen;
370 size_t vhost_len, sock_len;
371 /* TODO: check that we are running from vhost_worker? */
372 struct socket *sock = rcu_dereference_check(vq->private_data, 1);
377 mutex_lock(&vq->mutex);
378 vhost_disable_notify(&net->dev, vq);
379 vhost_hlen = vq->vhost_hlen;
380 sock_hlen = vq->sock_hlen;
382 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
384 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
386 while ((sock_len = peek_head_len(sock->sk))) {
387 sock_len += sock_hlen;
388 vhost_len = sock_len + vhost_hlen;
389 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
391 likely(mergeable) ? UIO_MAXIOV : 1);
392 /* On error, stop handling until the next kick. */
393 if (unlikely(headcount < 0))
395 /* OK, now we need to know about added descriptors. */
397 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
398 /* They have slipped one in as we were
399 * doing that: check again. */
400 vhost_disable_notify(&net->dev, vq);
403 /* Nothing new? Wait for eventfd to tell us
407 /* We don't need to be notified again. */
408 if (unlikely((vhost_hlen)))
409 /* Skip header. TODO: support TSO. */
410 move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
412 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
413 * needed because recvmsg can modify msg_iov. */
414 copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
416 err = sock->ops->recvmsg(NULL, sock, &msg,
417 sock_len, MSG_DONTWAIT | MSG_TRUNC);
418 /* Userspace might have consumed the packet meanwhile:
419 * it's not supposed to do this usually, but might be hard
420 * to prevent. Discard data we got (if any) and keep going. */
421 if (unlikely(err != sock_len)) {
422 pr_debug("Discarded rx packet: "
423 " len %d, expected %zd\n", err, sock_len);
424 vhost_discard_vq_desc(vq, headcount);
427 if (unlikely(vhost_hlen) &&
428 memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
430 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
434 /* TODO: Should check and handle checksum. */
435 if (likely(mergeable) &&
436 memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
437 offsetof(typeof(hdr), num_buffers),
438 sizeof hdr.num_buffers)) {
439 vq_err(vq, "Failed num_buffers write");
440 vhost_discard_vq_desc(vq, headcount);
443 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
445 if (unlikely(vq_log))
446 vhost_log_write(vq, vq_log, log, vhost_len);
447 total_len += vhost_len;
448 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
449 vhost_poll_queue(&vq->poll);
454 mutex_unlock(&vq->mutex);
457 static void handle_tx_kick(struct vhost_work *work)
459 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
461 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
466 static void handle_rx_kick(struct vhost_work *work)
468 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
470 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
475 static void handle_tx_net(struct vhost_work *work)
477 struct vhost_net *net = container_of(work, struct vhost_net,
478 poll[VHOST_NET_VQ_TX].work);
482 static void handle_rx_net(struct vhost_work *work)
484 struct vhost_net *net = container_of(work, struct vhost_net,
485 poll[VHOST_NET_VQ_RX].work);
489 static int vhost_net_open(struct inode *inode, struct file *f)
491 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
492 struct vhost_dev *dev;
499 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
500 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
501 r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
507 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
508 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
509 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
516 static void vhost_net_disable_vq(struct vhost_net *n,
517 struct vhost_virtqueue *vq)
519 if (!vq->private_data)
521 if (vq == n->vqs + VHOST_NET_VQ_TX) {
523 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
525 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
528 static void vhost_net_enable_vq(struct vhost_net *n,
529 struct vhost_virtqueue *vq)
533 sock = rcu_dereference_protected(vq->private_data,
534 lockdep_is_held(&vq->mutex));
537 if (vq == n->vqs + VHOST_NET_VQ_TX) {
538 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
539 tx_poll_start(n, sock);
541 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
544 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
545 struct vhost_virtqueue *vq)
549 mutex_lock(&vq->mutex);
550 sock = rcu_dereference_protected(vq->private_data,
551 lockdep_is_held(&vq->mutex));
552 vhost_net_disable_vq(n, vq);
553 rcu_assign_pointer(vq->private_data, NULL);
554 mutex_unlock(&vq->mutex);
558 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
559 struct socket **rx_sock)
561 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
562 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
565 static void vhost_net_flush_vq(struct vhost_net *n, int index)
567 vhost_poll_flush(n->poll + index);
568 vhost_poll_flush(&n->dev.vqs[index].poll);
571 static void vhost_net_flush(struct vhost_net *n)
573 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
574 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
577 static int vhost_net_release(struct inode *inode, struct file *f)
579 struct vhost_net *n = f->private_data;
580 struct socket *tx_sock;
581 struct socket *rx_sock;
583 vhost_net_stop(n, &tx_sock, &rx_sock);
585 vhost_dev_cleanup(&n->dev);
590 /* We do an extra flush before freeing memory,
591 * since jobs can re-queue themselves. */
597 static struct socket *get_raw_socket(int fd)
600 struct sockaddr_ll sa;
601 char buf[MAX_ADDR_LEN];
603 int uaddr_len = sizeof uaddr, r;
604 struct socket *sock = sockfd_lookup(fd, &r);
607 return ERR_PTR(-ENOTSOCK);
609 /* Parameter checking */
610 if (sock->sk->sk_type != SOCK_RAW) {
611 r = -ESOCKTNOSUPPORT;
615 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
620 if (uaddr.sa.sll_family != AF_PACKET) {
630 static struct socket *get_tap_socket(int fd)
632 struct file *file = fget(fd);
636 return ERR_PTR(-EBADF);
637 sock = tun_get_socket(file);
640 sock = macvtap_get_socket(file);
646 static struct socket *get_socket(int fd)
650 /* special case to disable backend */
653 sock = get_raw_socket(fd);
656 sock = get_tap_socket(fd);
659 return ERR_PTR(-ENOTSOCK);
662 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
664 struct socket *sock, *oldsock;
665 struct vhost_virtqueue *vq;
666 struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
669 mutex_lock(&n->dev.mutex);
670 r = vhost_dev_check_owner(&n->dev);
674 if (index >= VHOST_NET_VQ_MAX) {
679 mutex_lock(&vq->mutex);
681 /* Verify that ring has been setup correctly. */
682 if (!vhost_vq_access_ok(vq)) {
686 sock = get_socket(fd);
692 /* start polling new socket */
693 oldsock = rcu_dereference_protected(vq->private_data,
694 lockdep_is_held(&vq->mutex));
695 if (sock != oldsock) {
696 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
701 oldubufs = vq->ubufs;
703 vhost_net_disable_vq(n, vq);
704 rcu_assign_pointer(vq->private_data, sock);
705 vhost_net_enable_vq(n, vq);
708 mutex_unlock(&vq->mutex);
711 vhost_ubuf_put_and_wait(oldubufs);
714 vhost_net_flush_vq(n, index);
718 mutex_unlock(&n->dev.mutex);
724 mutex_unlock(&vq->mutex);
726 mutex_unlock(&n->dev.mutex);
730 static long vhost_net_reset_owner(struct vhost_net *n)
732 struct socket *tx_sock = NULL;
733 struct socket *rx_sock = NULL;
736 mutex_lock(&n->dev.mutex);
737 err = vhost_dev_check_owner(&n->dev);
740 vhost_net_stop(n, &tx_sock, &rx_sock);
742 err = vhost_dev_reset_owner(&n->dev);
744 mutex_unlock(&n->dev.mutex);
752 static int vhost_net_set_features(struct vhost_net *n, u64 features)
754 size_t vhost_hlen, sock_hlen, hdr_len;
757 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
758 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
759 sizeof(struct virtio_net_hdr);
760 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
761 /* vhost provides vnet_hdr */
762 vhost_hlen = hdr_len;
765 /* socket provides vnet_hdr */
769 mutex_lock(&n->dev.mutex);
770 if ((features & (1 << VHOST_F_LOG_ALL)) &&
771 !vhost_log_access_ok(&n->dev)) {
772 mutex_unlock(&n->dev.mutex);
775 n->dev.acked_features = features;
777 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
778 mutex_lock(&n->vqs[i].mutex);
779 n->vqs[i].vhost_hlen = vhost_hlen;
780 n->vqs[i].sock_hlen = sock_hlen;
781 mutex_unlock(&n->vqs[i].mutex);
784 mutex_unlock(&n->dev.mutex);
788 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
791 struct vhost_net *n = f->private_data;
792 void __user *argp = (void __user *)arg;
793 u64 __user *featurep = argp;
794 struct vhost_vring_file backend;
799 case VHOST_NET_SET_BACKEND:
800 if (copy_from_user(&backend, argp, sizeof backend))
802 return vhost_net_set_backend(n, backend.index, backend.fd);
803 case VHOST_GET_FEATURES:
804 features = VHOST_FEATURES;
805 if (copy_to_user(featurep, &features, sizeof features))
808 case VHOST_SET_FEATURES:
809 if (copy_from_user(&features, featurep, sizeof features))
811 if (features & ~VHOST_FEATURES)
813 return vhost_net_set_features(n, features);
814 case VHOST_RESET_OWNER:
815 return vhost_net_reset_owner(n);
817 mutex_lock(&n->dev.mutex);
818 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
820 mutex_unlock(&n->dev.mutex);
826 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
829 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
833 static const struct file_operations vhost_net_fops = {
834 .owner = THIS_MODULE,
835 .release = vhost_net_release,
836 .unlocked_ioctl = vhost_net_ioctl,
838 .compat_ioctl = vhost_net_compat_ioctl,
840 .open = vhost_net_open,
841 .llseek = noop_llseek,
844 static struct miscdevice vhost_net_misc = {
850 static int vhost_net_init(void)
852 if (experimental_zcopytx)
853 vhost_enable_zcopy(VHOST_NET_VQ_TX);
854 return misc_register(&vhost_net_misc);
856 module_init(vhost_net_init);
858 static void vhost_net_exit(void)
860 misc_deregister(&vhost_net_misc);
862 module_exit(vhost_net_exit);
864 MODULE_VERSION("0.0.1");
865 MODULE_LICENSE("GPL v2");
866 MODULE_AUTHOR("Michael S. Tsirkin");
867 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");