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/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>
26 #include <linux/if_vlan.h>
32 static int experimental_zcopytx = 1;
33 module_param(experimental_zcopytx, int, 0444);
34 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
35 " 1 -Enable; 0 - Disable");
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
46 * For transmit, used buffer len is unused; we override it to track buffer
47 * status internally; used for zerocopy tx only.
49 /* Lower device DMA failed */
50 #define VHOST_DMA_FAILED_LEN 3
51 /* Lower device DMA done */
52 #define VHOST_DMA_DONE_LEN 2
53 /* Lower device DMA in progress */
54 #define VHOST_DMA_IN_PROGRESS 1
56 #define VHOST_DMA_CLEAR_LEN 0
58 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
61 VHOST_NET_FEATURES = VHOST_FEATURES |
62 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
63 (1ULL << VIRTIO_NET_F_MRG_RXBUF),
72 struct vhost_net_ubuf_ref {
74 wait_queue_head_t wait;
75 struct vhost_virtqueue *vq;
78 struct vhost_net_virtqueue {
79 struct vhost_virtqueue vq;
80 /* hdr is used to store the virtio header.
81 * Since each iovec has >= 1 byte length, we never need more than
82 * header length entries to store the header. */
83 struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
86 /* vhost zerocopy support fields below: */
87 /* last used idx for outstanding DMA zerocopy buffers */
89 /* first used idx for DMA done zerocopy buffers */
91 /* an array of userspace buffers info */
92 struct ubuf_info *ubuf_info;
93 /* Reference counting for outstanding ubufs.
94 * Protected by vq mutex. Writers must also take device mutex. */
95 struct vhost_net_ubuf_ref *ubufs;
100 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
101 struct vhost_poll poll[VHOST_NET_VQ_MAX];
102 /* Number of TX recently submitted.
103 * Protected by tx vq lock. */
105 /* Number of times zerocopy TX recently failed.
106 * Protected by tx vq lock. */
107 unsigned tx_zcopy_err;
108 /* Flush in progress. Protected by tx vq lock. */
112 static unsigned vhost_net_zcopy_mask __read_mostly;
114 static void vhost_net_enable_zcopy(int vq)
116 vhost_net_zcopy_mask |= 0x1 << vq;
119 static void vhost_net_zerocopy_done_signal(struct kref *kref)
121 struct vhost_net_ubuf_ref *ubufs;
123 ubufs = container_of(kref, struct vhost_net_ubuf_ref, kref);
124 wake_up(&ubufs->wait);
127 static struct vhost_net_ubuf_ref *
128 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
130 struct vhost_net_ubuf_ref *ubufs;
131 /* No zero copy backend? Nothing to count. */
134 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
136 return ERR_PTR(-ENOMEM);
137 kref_init(&ubufs->kref);
138 init_waitqueue_head(&ubufs->wait);
143 static void vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
145 kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
148 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
150 kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
151 wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
154 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
156 vhost_net_ubuf_put_and_wait(ubufs);
160 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
164 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
165 kfree(n->vqs[i].ubuf_info);
166 n->vqs[i].ubuf_info = NULL;
170 static int vhost_net_set_ubuf_info(struct vhost_net *n)
175 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
176 zcopy = vhost_net_zcopy_mask & (0x1 << i);
179 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
180 UIO_MAXIOV, GFP_KERNEL);
181 if (!n->vqs[i].ubuf_info)
187 vhost_net_clear_ubuf_info(n);
191 static void vhost_net_vq_reset(struct vhost_net *n)
195 vhost_net_clear_ubuf_info(n);
197 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
198 n->vqs[i].done_idx = 0;
199 n->vqs[i].upend_idx = 0;
200 n->vqs[i].ubufs = NULL;
201 n->vqs[i].vhost_hlen = 0;
202 n->vqs[i].sock_hlen = 0;
207 static void vhost_net_tx_packet(struct vhost_net *net)
210 if (net->tx_packets < 1024)
213 net->tx_zcopy_err = 0;
216 static void vhost_net_tx_err(struct vhost_net *net)
221 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
223 /* TX flush waits for outstanding DMAs to be done.
224 * Don't start new DMAs.
226 return !net->tx_flush &&
227 net->tx_packets / 64 >= net->tx_zcopy_err;
230 static bool vhost_sock_zcopy(struct socket *sock)
232 return unlikely(experimental_zcopytx) &&
233 sock_flag(sock->sk, SOCK_ZEROCOPY);
236 /* Pop first len bytes from iovec. Return number of segments used. */
237 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
238 size_t len, int iov_count)
243 while (len && seg < iov_count) {
244 size = min(from->iov_len, len);
245 to->iov_base = from->iov_base;
247 from->iov_len -= size;
248 from->iov_base += size;
256 /* Copy iovec entries for len bytes from iovec. */
257 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
258 size_t len, int iovcount)
263 while (len && seg < iovcount) {
264 size = min(from->iov_len, len);
265 to->iov_base = from->iov_base;
274 /* In case of DMA done not in order in lower device driver for some reason.
275 * upend_idx is used to track end of used idx, done_idx is used to track head
276 * of used idx. Once lower device DMA done contiguously, we will signal KVM
279 static void vhost_zerocopy_signal_used(struct vhost_net *net,
280 struct vhost_virtqueue *vq)
282 struct vhost_net_virtqueue *nvq =
283 container_of(vq, struct vhost_net_virtqueue, vq);
287 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
288 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
289 vhost_net_tx_err(net);
290 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
291 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
297 add = min(UIO_MAXIOV - nvq->done_idx, j);
298 vhost_add_used_and_signal_n(vq->dev, vq,
299 &vq->heads[nvq->done_idx], add);
300 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
305 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
307 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
308 struct vhost_virtqueue *vq = ubufs->vq;
309 int cnt = atomic_read(&ubufs->kref.refcount);
312 * Trigger polling thread if guest stopped submitting new buffers:
313 * in this case, the refcount after decrement will eventually reach 1
315 * We also trigger polling periodically after each 16 packets
316 * (the value 16 here is more or less arbitrary, it's tuned to trigger
317 * less than 10% of times).
319 if (cnt <= 2 || !(cnt % 16))
320 vhost_poll_queue(&vq->poll);
321 /* set len to mark this desc buffers done DMA */
322 vq->heads[ubuf->desc].len = success ?
323 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
324 vhost_net_ubuf_put(ubufs);
327 /* Expects to be always run from workqueue - which acts as
328 * read-size critical section for our kind of RCU. */
329 static void handle_tx(struct vhost_net *net)
331 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
332 struct vhost_virtqueue *vq = &nvq->vq;
335 struct msghdr msg = {
341 .msg_flags = MSG_DONTWAIT,
343 size_t len, total_len = 0;
347 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
348 bool zcopy, zcopy_used;
350 mutex_lock(&vq->mutex);
351 sock = vq->private_data;
355 vhost_disable_notify(&net->dev, vq);
357 hdr_size = nvq->vhost_hlen;
361 /* Release DMAs done buffers first */
363 vhost_zerocopy_signal_used(net, vq);
365 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
369 /* On error, stop handling until the next kick. */
370 if (unlikely(head < 0))
372 /* Nothing new? Wait for eventfd to tell us they refilled. */
373 if (head == vq->num) {
376 /* If more outstanding DMAs, queue the work.
377 * Handle upend_idx wrap around
379 num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
380 (nvq->upend_idx - nvq->done_idx) :
381 (nvq->upend_idx + UIO_MAXIOV -
383 if (unlikely(num_pends > VHOST_MAX_PEND))
385 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
386 vhost_disable_notify(&net->dev, vq);
392 vq_err(vq, "Unexpected descriptor format for TX: "
393 "out %d, int %d\n", out, in);
396 /* Skip header. TODO: support TSO. */
397 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
398 msg.msg_iovlen = out;
399 len = iov_length(vq->iov, out);
402 vq_err(vq, "Unexpected header len for TX: "
403 "%zd expected %zd\n",
404 iov_length(nvq->hdr, s), hdr_size);
407 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
408 nvq->upend_idx != nvq->done_idx);
410 /* use msg_control to pass vhost zerocopy ubuf info to skb */
412 vq->heads[nvq->upend_idx].id = head;
413 if (!vhost_net_tx_select_zcopy(net) ||
414 len < VHOST_GOODCOPY_LEN) {
415 /* copy don't need to wait for DMA done */
416 vq->heads[nvq->upend_idx].len =
418 msg.msg_control = NULL;
419 msg.msg_controllen = 0;
422 struct ubuf_info *ubuf;
423 ubuf = nvq->ubuf_info + nvq->upend_idx;
425 vq->heads[nvq->upend_idx].len =
426 VHOST_DMA_IN_PROGRESS;
427 ubuf->callback = vhost_zerocopy_callback;
428 ubuf->ctx = nvq->ubufs;
429 ubuf->desc = nvq->upend_idx;
430 msg.msg_control = ubuf;
431 msg.msg_controllen = sizeof(ubuf);
433 kref_get(&ubufs->kref);
435 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
437 msg.msg_control = NULL;
438 /* TODO: Check specific error and bomb out unless ENOBUFS? */
439 err = sock->ops->sendmsg(NULL, sock, &msg, len);
440 if (unlikely(err < 0)) {
443 vhost_net_ubuf_put(ubufs);
444 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
447 vhost_discard_vq_desc(vq, 1);
451 pr_debug("Truncated TX packet: "
452 " len %d != %zd\n", err, len);
454 vhost_add_used_and_signal(&net->dev, vq, head, 0);
456 vhost_zerocopy_signal_used(net, vq);
458 vhost_net_tx_packet(net);
459 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
460 vhost_poll_queue(&vq->poll);
465 mutex_unlock(&vq->mutex);
468 static int peek_head_len(struct sock *sk)
470 struct sk_buff *head;
474 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
475 head = skb_peek(&sk->sk_receive_queue);
478 if (vlan_tx_tag_present(head))
482 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
486 /* This is a multi-buffer version of vhost_get_desc, that works if
487 * vq has read descriptors only.
488 * @vq - the relevant virtqueue
489 * @datalen - data length we'll be reading
490 * @iovcount - returned count of io vectors we fill
492 * @log_num - log offset
493 * @quota - headcount quota, 1 for big buffer
494 * returns number of buffer heads allocated, negative on error
496 static int get_rx_bufs(struct vhost_virtqueue *vq,
497 struct vring_used_elem *heads,
500 struct vhost_log *log,
504 unsigned int out, in;
510 while (datalen > 0 && headcount < quota) {
511 if (unlikely(seg >= UIO_MAXIOV)) {
515 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
516 ARRAY_SIZE(vq->iov) - seg, &out,
522 if (unlikely(out || in <= 0)) {
523 vq_err(vq, "unexpected descriptor format for RX: "
524 "out %d, in %d\n", out, in);
532 heads[headcount].id = d;
533 heads[headcount].len = iov_length(vq->iov + seg, in);
534 datalen -= heads[headcount].len;
538 heads[headcount - 1].len += datalen;
544 vhost_discard_vq_desc(vq, headcount);
548 /* Expects to be always run from workqueue - which acts as
549 * read-size critical section for our kind of RCU. */
550 static void handle_rx(struct vhost_net *net)
552 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
553 struct vhost_virtqueue *vq = &nvq->vq;
554 unsigned uninitialized_var(in), log;
555 struct vhost_log *vq_log;
556 struct msghdr msg = {
559 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
562 .msg_flags = MSG_DONTWAIT,
564 struct virtio_net_hdr_mrg_rxbuf hdr = {
566 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
568 size_t total_len = 0;
571 size_t vhost_hlen, sock_hlen;
572 size_t vhost_len, sock_len;
575 mutex_lock(&vq->mutex);
576 sock = vq->private_data;
579 vhost_disable_notify(&net->dev, vq);
581 vhost_hlen = nvq->vhost_hlen;
582 sock_hlen = nvq->sock_hlen;
584 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
586 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
588 while ((sock_len = peek_head_len(sock->sk))) {
589 sock_len += sock_hlen;
590 vhost_len = sock_len + vhost_hlen;
591 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
593 likely(mergeable) ? UIO_MAXIOV : 1);
594 /* On error, stop handling until the next kick. */
595 if (unlikely(headcount < 0))
597 /* OK, now we need to know about added descriptors. */
599 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
600 /* They have slipped one in as we were
601 * doing that: check again. */
602 vhost_disable_notify(&net->dev, vq);
605 /* Nothing new? Wait for eventfd to tell us
609 /* We don't need to be notified again. */
610 if (unlikely((vhost_hlen)))
611 /* Skip header. TODO: support TSO. */
612 move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
614 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
615 * needed because recvmsg can modify msg_iov. */
616 copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
618 err = sock->ops->recvmsg(NULL, sock, &msg,
619 sock_len, MSG_DONTWAIT | MSG_TRUNC);
620 /* Userspace might have consumed the packet meanwhile:
621 * it's not supposed to do this usually, but might be hard
622 * to prevent. Discard data we got (if any) and keep going. */
623 if (unlikely(err != sock_len)) {
624 pr_debug("Discarded rx packet: "
625 " len %d, expected %zd\n", err, sock_len);
626 vhost_discard_vq_desc(vq, headcount);
629 if (unlikely(vhost_hlen) &&
630 memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
632 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
636 /* TODO: Should check and handle checksum. */
637 if (likely(mergeable) &&
638 memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
639 offsetof(typeof(hdr), num_buffers),
640 sizeof hdr.num_buffers)) {
641 vq_err(vq, "Failed num_buffers write");
642 vhost_discard_vq_desc(vq, headcount);
645 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
647 if (unlikely(vq_log))
648 vhost_log_write(vq, vq_log, log, vhost_len);
649 total_len += vhost_len;
650 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
651 vhost_poll_queue(&vq->poll);
656 mutex_unlock(&vq->mutex);
659 static void handle_tx_kick(struct vhost_work *work)
661 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
663 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
668 static void handle_rx_kick(struct vhost_work *work)
670 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
672 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
677 static void handle_tx_net(struct vhost_work *work)
679 struct vhost_net *net = container_of(work, struct vhost_net,
680 poll[VHOST_NET_VQ_TX].work);
684 static void handle_rx_net(struct vhost_work *work)
686 struct vhost_net *net = container_of(work, struct vhost_net,
687 poll[VHOST_NET_VQ_RX].work);
691 static int vhost_net_open(struct inode *inode, struct file *f)
693 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
694 struct vhost_dev *dev;
695 struct vhost_virtqueue **vqs;
700 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
707 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
708 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
709 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
710 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
711 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
712 n->vqs[i].ubufs = NULL;
713 n->vqs[i].ubuf_info = NULL;
714 n->vqs[i].upend_idx = 0;
715 n->vqs[i].done_idx = 0;
716 n->vqs[i].vhost_hlen = 0;
717 n->vqs[i].sock_hlen = 0;
719 r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
726 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
727 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
734 static void vhost_net_disable_vq(struct vhost_net *n,
735 struct vhost_virtqueue *vq)
737 struct vhost_net_virtqueue *nvq =
738 container_of(vq, struct vhost_net_virtqueue, vq);
739 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
740 if (!vq->private_data)
742 vhost_poll_stop(poll);
745 static int vhost_net_enable_vq(struct vhost_net *n,
746 struct vhost_virtqueue *vq)
748 struct vhost_net_virtqueue *nvq =
749 container_of(vq, struct vhost_net_virtqueue, vq);
750 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
753 sock = vq->private_data;
757 return vhost_poll_start(poll, sock->file);
760 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
761 struct vhost_virtqueue *vq)
765 mutex_lock(&vq->mutex);
766 sock = vq->private_data;
767 vhost_net_disable_vq(n, vq);
768 vq->private_data = NULL;
769 mutex_unlock(&vq->mutex);
773 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
774 struct socket **rx_sock)
776 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
777 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
780 static void vhost_net_flush_vq(struct vhost_net *n, int index)
782 vhost_poll_flush(n->poll + index);
783 vhost_poll_flush(&n->vqs[index].vq.poll);
786 static void vhost_net_flush(struct vhost_net *n)
788 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
789 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
790 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
791 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
793 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
794 /* Wait for all lower device DMAs done. */
795 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
796 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
798 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
799 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
803 static int vhost_net_release(struct inode *inode, struct file *f)
805 struct vhost_net *n = f->private_data;
806 struct socket *tx_sock;
807 struct socket *rx_sock;
809 vhost_net_stop(n, &tx_sock, &rx_sock);
811 vhost_dev_stop(&n->dev);
812 vhost_dev_cleanup(&n->dev, false);
813 vhost_net_vq_reset(n);
818 /* We do an extra flush before freeing memory,
819 * since jobs can re-queue themselves. */
826 static struct socket *get_raw_socket(int fd)
829 struct sockaddr_ll sa;
830 char buf[MAX_ADDR_LEN];
832 int uaddr_len = sizeof uaddr, r;
833 struct socket *sock = sockfd_lookup(fd, &r);
836 return ERR_PTR(-ENOTSOCK);
838 /* Parameter checking */
839 if (sock->sk->sk_type != SOCK_RAW) {
840 r = -ESOCKTNOSUPPORT;
844 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
849 if (uaddr.sa.sll_family != AF_PACKET) {
859 static struct socket *get_tap_socket(int fd)
861 struct file *file = fget(fd);
865 return ERR_PTR(-EBADF);
866 sock = tun_get_socket(file);
869 sock = macvtap_get_socket(file);
875 static struct socket *get_socket(int fd)
879 /* special case to disable backend */
882 sock = get_raw_socket(fd);
885 sock = get_tap_socket(fd);
888 return ERR_PTR(-ENOTSOCK);
891 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
893 struct socket *sock, *oldsock;
894 struct vhost_virtqueue *vq;
895 struct vhost_net_virtqueue *nvq;
896 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
899 mutex_lock(&n->dev.mutex);
900 r = vhost_dev_check_owner(&n->dev);
904 if (index >= VHOST_NET_VQ_MAX) {
908 vq = &n->vqs[index].vq;
909 nvq = &n->vqs[index];
910 mutex_lock(&vq->mutex);
912 /* Verify that ring has been setup correctly. */
913 if (!vhost_vq_access_ok(vq)) {
917 sock = get_socket(fd);
923 /* start polling new socket */
924 oldsock = vq->private_data;
925 if (sock != oldsock) {
926 ubufs = vhost_net_ubuf_alloc(vq,
927 sock && vhost_sock_zcopy(sock));
933 vhost_net_disable_vq(n, vq);
934 vq->private_data = sock;
935 r = vhost_init_used(vq);
938 r = vhost_net_enable_vq(n, vq);
942 oldubufs = nvq->ubufs;
950 mutex_unlock(&vq->mutex);
953 vhost_net_ubuf_put_wait_and_free(oldubufs);
954 mutex_lock(&vq->mutex);
955 vhost_zerocopy_signal_used(n, vq);
956 mutex_unlock(&vq->mutex);
960 vhost_net_flush_vq(n, index);
964 mutex_unlock(&n->dev.mutex);
968 vq->private_data = oldsock;
969 vhost_net_enable_vq(n, vq);
971 vhost_net_ubuf_put_wait_and_free(ubufs);
975 mutex_unlock(&vq->mutex);
977 mutex_unlock(&n->dev.mutex);
981 static long vhost_net_reset_owner(struct vhost_net *n)
983 struct socket *tx_sock = NULL;
984 struct socket *rx_sock = NULL;
986 struct vhost_memory *memory;
988 mutex_lock(&n->dev.mutex);
989 err = vhost_dev_check_owner(&n->dev);
992 memory = vhost_dev_reset_owner_prepare();
997 vhost_net_stop(n, &tx_sock, &rx_sock);
999 vhost_dev_reset_owner(&n->dev, memory);
1000 vhost_net_vq_reset(n);
1002 mutex_unlock(&n->dev.mutex);
1004 fput(tx_sock->file);
1006 fput(rx_sock->file);
1010 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1012 size_t vhost_hlen, sock_hlen, hdr_len;
1015 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1016 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1017 sizeof(struct virtio_net_hdr);
1018 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1019 /* vhost provides vnet_hdr */
1020 vhost_hlen = hdr_len;
1023 /* socket provides vnet_hdr */
1025 sock_hlen = hdr_len;
1027 mutex_lock(&n->dev.mutex);
1028 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1029 !vhost_log_access_ok(&n->dev)) {
1030 mutex_unlock(&n->dev.mutex);
1033 n->dev.acked_features = features;
1035 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1036 mutex_lock(&n->vqs[i].vq.mutex);
1037 n->vqs[i].vhost_hlen = vhost_hlen;
1038 n->vqs[i].sock_hlen = sock_hlen;
1039 mutex_unlock(&n->vqs[i].vq.mutex);
1042 mutex_unlock(&n->dev.mutex);
1046 static long vhost_net_set_owner(struct vhost_net *n)
1050 mutex_lock(&n->dev.mutex);
1051 if (vhost_dev_has_owner(&n->dev)) {
1055 r = vhost_net_set_ubuf_info(n);
1058 r = vhost_dev_set_owner(&n->dev);
1060 vhost_net_clear_ubuf_info(n);
1063 mutex_unlock(&n->dev.mutex);
1067 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1070 struct vhost_net *n = f->private_data;
1071 void __user *argp = (void __user *)arg;
1072 u64 __user *featurep = argp;
1073 struct vhost_vring_file backend;
1078 case VHOST_NET_SET_BACKEND:
1079 if (copy_from_user(&backend, argp, sizeof backend))
1081 return vhost_net_set_backend(n, backend.index, backend.fd);
1082 case VHOST_GET_FEATURES:
1083 features = VHOST_NET_FEATURES;
1084 if (copy_to_user(featurep, &features, sizeof features))
1087 case VHOST_SET_FEATURES:
1088 if (copy_from_user(&features, featurep, sizeof features))
1090 if (features & ~VHOST_NET_FEATURES)
1092 return vhost_net_set_features(n, features);
1093 case VHOST_RESET_OWNER:
1094 return vhost_net_reset_owner(n);
1095 case VHOST_SET_OWNER:
1096 return vhost_net_set_owner(n);
1098 mutex_lock(&n->dev.mutex);
1099 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1100 if (r == -ENOIOCTLCMD)
1101 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1104 mutex_unlock(&n->dev.mutex);
1109 #ifdef CONFIG_COMPAT
1110 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1113 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1117 static const struct file_operations vhost_net_fops = {
1118 .owner = THIS_MODULE,
1119 .release = vhost_net_release,
1120 .unlocked_ioctl = vhost_net_ioctl,
1121 #ifdef CONFIG_COMPAT
1122 .compat_ioctl = vhost_net_compat_ioctl,
1124 .open = vhost_net_open,
1125 .llseek = noop_llseek,
1128 static struct miscdevice vhost_net_misc = {
1129 .minor = VHOST_NET_MINOR,
1130 .name = "vhost-net",
1131 .fops = &vhost_net_fops,
1134 static int vhost_net_init(void)
1136 if (experimental_zcopytx)
1137 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1138 return misc_register(&vhost_net_misc);
1140 module_init(vhost_net_init);
1142 static void vhost_net_exit(void)
1144 misc_deregister(&vhost_net_misc);
1146 module_exit(vhost_net_exit);
1148 MODULE_VERSION("0.0.1");
1149 MODULE_LICENSE("GPL v2");
1150 MODULE_AUTHOR("Michael S. Tsirkin");
1151 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1152 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1153 MODULE_ALIAS("devname:vhost-net");