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;
34 module_param(experimental_zcopytx, int, 0444);
35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
37 /* Max number of bytes transferred before requeueing the job.
38 * Using this limit prevents one virtqueue from starving others. */
39 #define VHOST_NET_WEIGHT 0x80000
41 /* MAX number of TX used buffers for outstanding zerocopy */
42 #define VHOST_MAX_PEND 128
43 #define VHOST_GOODCOPY_LEN 256
51 enum vhost_net_poll_state {
52 VHOST_NET_POLL_DISABLED = 0,
53 VHOST_NET_POLL_STARTED = 1,
54 VHOST_NET_POLL_STOPPED = 2,
59 struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
60 struct vhost_poll poll[VHOST_NET_VQ_MAX];
61 /* Tells us whether we are polling a socket for TX.
62 * We only do this when socket buffer fills up.
63 * Protected by tx vq lock. */
64 enum vhost_net_poll_state tx_poll_state;
67 static bool vhost_sock_zcopy(struct socket *sock)
69 return unlikely(experimental_zcopytx) &&
70 sock_flag(sock->sk, SOCK_ZEROCOPY);
73 /* Pop first len bytes from iovec. Return number of segments used. */
74 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
75 size_t len, int iov_count)
80 while (len && seg < iov_count) {
81 size = min(from->iov_len, len);
82 to->iov_base = from->iov_base;
84 from->iov_len -= size;
85 from->iov_base += size;
93 /* Copy iovec entries for len bytes from iovec. */
94 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
95 size_t len, int iovcount)
100 while (len && seg < iovcount) {
101 size = min(from->iov_len, len);
102 to->iov_base = from->iov_base;
111 /* Caller must have TX VQ lock */
112 static void tx_poll_stop(struct vhost_net *net)
114 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
116 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
117 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
120 /* Caller must have TX VQ lock */
121 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
123 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
125 vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
126 net->tx_poll_state = VHOST_NET_POLL_STARTED;
129 /* Expects to be always run from workqueue - which acts as
130 * read-size critical section for our kind of RCU. */
131 static void handle_tx(struct vhost_net *net)
133 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
136 struct msghdr msg = {
142 .msg_flags = MSG_DONTWAIT,
144 size_t len, total_len = 0;
148 struct vhost_ubuf_ref *uninitialized_var(ubufs);
151 /* TODO: check that we are running from vhost_worker? */
152 sock = rcu_dereference_check(vq->private_data, 1);
156 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
157 if (wmem >= sock->sk->sk_sndbuf) {
158 mutex_lock(&vq->mutex);
159 tx_poll_start(net, sock);
160 mutex_unlock(&vq->mutex);
164 mutex_lock(&vq->mutex);
165 vhost_disable_notify(&net->dev, vq);
167 if (wmem < sock->sk->sk_sndbuf / 2)
169 hdr_size = vq->vhost_hlen;
173 /* Release DMAs done buffers first */
175 vhost_zerocopy_signal_used(vq);
177 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
181 /* On error, stop handling until the next kick. */
182 if (unlikely(head < 0))
184 /* Nothing new? Wait for eventfd to tell us they refilled. */
185 if (head == vq->num) {
188 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
189 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
190 tx_poll_start(net, sock);
191 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
194 /* If more outstanding DMAs, queue the work.
195 * Handle upend_idx wrap around
197 num_pends = likely(vq->upend_idx >= vq->done_idx) ?
198 (vq->upend_idx - vq->done_idx) :
199 (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
200 if (unlikely(num_pends > VHOST_MAX_PEND)) {
201 tx_poll_start(net, sock);
202 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
205 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
206 vhost_disable_notify(&net->dev, vq);
212 vq_err(vq, "Unexpected descriptor format for TX: "
213 "out %d, int %d\n", out, in);
216 /* Skip header. TODO: support TSO. */
217 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
218 msg.msg_iovlen = out;
219 len = iov_length(vq->iov, out);
222 vq_err(vq, "Unexpected header len for TX: "
223 "%zd expected %zd\n",
224 iov_length(vq->hdr, s), hdr_size);
227 /* use msg_control to pass vhost zerocopy ubuf info to skb */
229 vq->heads[vq->upend_idx].id = head;
230 if (len < VHOST_GOODCOPY_LEN) {
231 /* copy don't need to wait for DMA done */
232 vq->heads[vq->upend_idx].len =
234 msg.msg_control = NULL;
235 msg.msg_controllen = 0;
238 struct ubuf_info *ubuf = &vq->ubuf_info[head];
240 vq->heads[vq->upend_idx].len = len;
241 ubuf->callback = vhost_zerocopy_callback;
242 ubuf->ctx = vq->ubufs;
243 ubuf->desc = vq->upend_idx;
244 msg.msg_control = ubuf;
245 msg.msg_controllen = sizeof(ubuf);
247 kref_get(&ubufs->kref);
249 vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
251 /* TODO: Check specific error and bomb out unless ENOBUFS? */
252 err = sock->ops->sendmsg(NULL, sock, &msg, len);
253 if (unlikely(err < 0)) {
256 vhost_ubuf_put(ubufs);
257 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
260 vhost_discard_vq_desc(vq, 1);
261 if (err == -EAGAIN || err == -ENOBUFS)
262 tx_poll_start(net, sock);
266 pr_debug("Truncated TX packet: "
267 " len %d != %zd\n", err, len);
269 vhost_add_used_and_signal(&net->dev, vq, head, 0);
271 vhost_zerocopy_signal_used(vq);
273 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
274 vhost_poll_queue(&vq->poll);
279 mutex_unlock(&vq->mutex);
282 static int peek_head_len(struct sock *sk)
284 struct sk_buff *head;
288 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
289 head = skb_peek(&sk->sk_receive_queue);
292 if (vlan_tx_tag_present(head))
296 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
300 /* This is a multi-buffer version of vhost_get_desc, that works if
301 * vq has read descriptors only.
302 * @vq - the relevant virtqueue
303 * @datalen - data length we'll be reading
304 * @iovcount - returned count of io vectors we fill
306 * @log_num - log offset
307 * @quota - headcount quota, 1 for big buffer
308 * returns number of buffer heads allocated, negative on error
310 static int get_rx_bufs(struct vhost_virtqueue *vq,
311 struct vring_used_elem *heads,
314 struct vhost_log *log,
318 unsigned int out, in;
324 while (datalen > 0 && headcount < quota) {
325 if (unlikely(seg >= UIO_MAXIOV)) {
329 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
330 ARRAY_SIZE(vq->iov) - seg, &out,
336 if (unlikely(out || in <= 0)) {
337 vq_err(vq, "unexpected descriptor format for RX: "
338 "out %d, in %d\n", out, in);
346 heads[headcount].id = d;
347 heads[headcount].len = iov_length(vq->iov + seg, in);
348 datalen -= heads[headcount].len;
352 heads[headcount - 1].len += datalen;
358 vhost_discard_vq_desc(vq, headcount);
362 /* Expects to be always run from workqueue - which acts as
363 * read-size critical section for our kind of RCU. */
364 static void handle_rx(struct vhost_net *net)
366 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
367 unsigned uninitialized_var(in), log;
368 struct vhost_log *vq_log;
369 struct msghdr msg = {
372 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
375 .msg_flags = MSG_DONTWAIT,
377 struct virtio_net_hdr_mrg_rxbuf hdr = {
379 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
381 size_t total_len = 0;
384 size_t vhost_hlen, sock_hlen;
385 size_t vhost_len, sock_len;
386 /* TODO: check that we are running from vhost_worker? */
387 struct socket *sock = rcu_dereference_check(vq->private_data, 1);
392 mutex_lock(&vq->mutex);
393 vhost_disable_notify(&net->dev, vq);
394 vhost_hlen = vq->vhost_hlen;
395 sock_hlen = vq->sock_hlen;
397 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
399 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
401 while ((sock_len = peek_head_len(sock->sk))) {
402 sock_len += sock_hlen;
403 vhost_len = sock_len + vhost_hlen;
404 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
406 likely(mergeable) ? UIO_MAXIOV : 1);
407 /* On error, stop handling until the next kick. */
408 if (unlikely(headcount < 0))
410 /* OK, now we need to know about added descriptors. */
412 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
413 /* They have slipped one in as we were
414 * doing that: check again. */
415 vhost_disable_notify(&net->dev, vq);
418 /* Nothing new? Wait for eventfd to tell us
422 /* We don't need to be notified again. */
423 if (unlikely((vhost_hlen)))
424 /* Skip header. TODO: support TSO. */
425 move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
427 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
428 * needed because recvmsg can modify msg_iov. */
429 copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
431 err = sock->ops->recvmsg(NULL, sock, &msg,
432 sock_len, MSG_DONTWAIT | MSG_TRUNC);
433 /* Userspace might have consumed the packet meanwhile:
434 * it's not supposed to do this usually, but might be hard
435 * to prevent. Discard data we got (if any) and keep going. */
436 if (unlikely(err != sock_len)) {
437 pr_debug("Discarded rx packet: "
438 " len %d, expected %zd\n", err, sock_len);
439 vhost_discard_vq_desc(vq, headcount);
442 if (unlikely(vhost_hlen) &&
443 memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
445 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
449 /* TODO: Should check and handle checksum. */
450 if (likely(mergeable) &&
451 memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
452 offsetof(typeof(hdr), num_buffers),
453 sizeof hdr.num_buffers)) {
454 vq_err(vq, "Failed num_buffers write");
455 vhost_discard_vq_desc(vq, headcount);
458 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
460 if (unlikely(vq_log))
461 vhost_log_write(vq, vq_log, log, vhost_len);
462 total_len += vhost_len;
463 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
464 vhost_poll_queue(&vq->poll);
469 mutex_unlock(&vq->mutex);
472 static void handle_tx_kick(struct vhost_work *work)
474 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
476 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
481 static void handle_rx_kick(struct vhost_work *work)
483 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
485 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
490 static void handle_tx_net(struct vhost_work *work)
492 struct vhost_net *net = container_of(work, struct vhost_net,
493 poll[VHOST_NET_VQ_TX].work);
497 static void handle_rx_net(struct vhost_work *work)
499 struct vhost_net *net = container_of(work, struct vhost_net,
500 poll[VHOST_NET_VQ_RX].work);
504 static int vhost_net_open(struct inode *inode, struct file *f)
506 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
507 struct vhost_dev *dev;
514 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
515 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
516 r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
522 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
523 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
524 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
531 static void vhost_net_disable_vq(struct vhost_net *n,
532 struct vhost_virtqueue *vq)
534 if (!vq->private_data)
536 if (vq == n->vqs + VHOST_NET_VQ_TX) {
538 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
540 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
543 static void vhost_net_enable_vq(struct vhost_net *n,
544 struct vhost_virtqueue *vq)
548 sock = rcu_dereference_protected(vq->private_data,
549 lockdep_is_held(&vq->mutex));
552 if (vq == n->vqs + VHOST_NET_VQ_TX) {
553 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
554 tx_poll_start(n, sock);
556 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
559 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
560 struct vhost_virtqueue *vq)
564 mutex_lock(&vq->mutex);
565 sock = rcu_dereference_protected(vq->private_data,
566 lockdep_is_held(&vq->mutex));
567 vhost_net_disable_vq(n, vq);
568 rcu_assign_pointer(vq->private_data, NULL);
569 mutex_unlock(&vq->mutex);
573 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
574 struct socket **rx_sock)
576 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
577 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
580 static void vhost_net_flush_vq(struct vhost_net *n, int index)
582 vhost_poll_flush(n->poll + index);
583 vhost_poll_flush(&n->dev.vqs[index].poll);
586 static void vhost_net_flush(struct vhost_net *n)
588 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
589 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
592 static int vhost_net_release(struct inode *inode, struct file *f)
594 struct vhost_net *n = f->private_data;
595 struct socket *tx_sock;
596 struct socket *rx_sock;
598 vhost_net_stop(n, &tx_sock, &rx_sock);
600 vhost_dev_cleanup(&n->dev, false);
605 /* We do an extra flush before freeing memory,
606 * since jobs can re-queue themselves. */
612 static struct socket *get_raw_socket(int fd)
615 struct sockaddr_ll sa;
616 char buf[MAX_ADDR_LEN];
618 int uaddr_len = sizeof uaddr, r;
619 struct socket *sock = sockfd_lookup(fd, &r);
622 return ERR_PTR(-ENOTSOCK);
624 /* Parameter checking */
625 if (sock->sk->sk_type != SOCK_RAW) {
626 r = -ESOCKTNOSUPPORT;
630 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
635 if (uaddr.sa.sll_family != AF_PACKET) {
645 static struct socket *get_tap_socket(int fd)
647 struct file *file = fget(fd);
651 return ERR_PTR(-EBADF);
652 sock = tun_get_socket(file);
655 sock = macvtap_get_socket(file);
661 static struct socket *get_socket(int fd)
665 /* special case to disable backend */
668 sock = get_raw_socket(fd);
671 sock = get_tap_socket(fd);
674 return ERR_PTR(-ENOTSOCK);
677 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
679 struct socket *sock, *oldsock;
680 struct vhost_virtqueue *vq;
681 struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
684 mutex_lock(&n->dev.mutex);
685 r = vhost_dev_check_owner(&n->dev);
689 if (index >= VHOST_NET_VQ_MAX) {
694 mutex_lock(&vq->mutex);
696 /* Verify that ring has been setup correctly. */
697 if (!vhost_vq_access_ok(vq)) {
701 sock = get_socket(fd);
707 /* start polling new socket */
708 oldsock = rcu_dereference_protected(vq->private_data,
709 lockdep_is_held(&vq->mutex));
710 if (sock != oldsock) {
711 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
716 oldubufs = vq->ubufs;
718 vhost_net_disable_vq(n, vq);
719 rcu_assign_pointer(vq->private_data, sock);
720 vhost_net_enable_vq(n, vq);
722 r = vhost_init_used(vq);
727 mutex_unlock(&vq->mutex);
730 vhost_ubuf_put_and_wait(oldubufs);
731 mutex_lock(&vq->mutex);
732 vhost_zerocopy_signal_used(vq);
733 mutex_unlock(&vq->mutex);
737 vhost_net_flush_vq(n, index);
741 mutex_unlock(&n->dev.mutex);
747 mutex_unlock(&vq->mutex);
749 mutex_unlock(&n->dev.mutex);
753 static long vhost_net_reset_owner(struct vhost_net *n)
755 struct socket *tx_sock = NULL;
756 struct socket *rx_sock = NULL;
759 mutex_lock(&n->dev.mutex);
760 err = vhost_dev_check_owner(&n->dev);
763 vhost_net_stop(n, &tx_sock, &rx_sock);
765 err = vhost_dev_reset_owner(&n->dev);
767 mutex_unlock(&n->dev.mutex);
775 static int vhost_net_set_features(struct vhost_net *n, u64 features)
777 size_t vhost_hlen, sock_hlen, hdr_len;
780 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
781 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
782 sizeof(struct virtio_net_hdr);
783 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
784 /* vhost provides vnet_hdr */
785 vhost_hlen = hdr_len;
788 /* socket provides vnet_hdr */
792 mutex_lock(&n->dev.mutex);
793 if ((features & (1 << VHOST_F_LOG_ALL)) &&
794 !vhost_log_access_ok(&n->dev)) {
795 mutex_unlock(&n->dev.mutex);
798 n->dev.acked_features = features;
800 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
801 mutex_lock(&n->vqs[i].mutex);
802 n->vqs[i].vhost_hlen = vhost_hlen;
803 n->vqs[i].sock_hlen = sock_hlen;
804 mutex_unlock(&n->vqs[i].mutex);
807 mutex_unlock(&n->dev.mutex);
811 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
814 struct vhost_net *n = f->private_data;
815 void __user *argp = (void __user *)arg;
816 u64 __user *featurep = argp;
817 struct vhost_vring_file backend;
822 case VHOST_NET_SET_BACKEND:
823 if (copy_from_user(&backend, argp, sizeof backend))
825 return vhost_net_set_backend(n, backend.index, backend.fd);
826 case VHOST_GET_FEATURES:
827 features = VHOST_NET_FEATURES;
828 if (copy_to_user(featurep, &features, sizeof features))
831 case VHOST_SET_FEATURES:
832 if (copy_from_user(&features, featurep, sizeof features))
834 if (features & ~VHOST_NET_FEATURES)
836 return vhost_net_set_features(n, features);
837 case VHOST_RESET_OWNER:
838 return vhost_net_reset_owner(n);
840 mutex_lock(&n->dev.mutex);
841 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
843 mutex_unlock(&n->dev.mutex);
849 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
852 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
856 static const struct file_operations vhost_net_fops = {
857 .owner = THIS_MODULE,
858 .release = vhost_net_release,
859 .unlocked_ioctl = vhost_net_ioctl,
861 .compat_ioctl = vhost_net_compat_ioctl,
863 .open = vhost_net_open,
864 .llseek = noop_llseek,
867 static struct miscdevice vhost_net_misc = {
868 .minor = VHOST_NET_MINOR,
870 .fops = &vhost_net_fops,
873 static int vhost_net_init(void)
875 if (experimental_zcopytx)
876 vhost_enable_zcopy(VHOST_NET_VQ_TX);
877 return misc_register(&vhost_net_misc);
879 module_init(vhost_net_init);
881 static void vhost_net_exit(void)
883 misc_deregister(&vhost_net_misc);
885 module_exit(vhost_net_exit);
887 MODULE_VERSION("0.0.1");
888 MODULE_LICENSE("GPL v2");
889 MODULE_AUTHOR("Michael S. Tsirkin");
890 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
891 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
892 MODULE_ALIAS("devname:vhost-net");