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);
311 /* set len to mark this desc buffers done DMA */
312 vq->heads[ubuf->desc].len = success ?
313 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
314 vhost_net_ubuf_put(ubufs);
317 * Trigger polling thread if guest stopped submitting new buffers:
318 * in this case, the refcount after decrement will eventually reach 1
320 * We also trigger polling periodically after each 16 packets
321 * (the value 16 here is more or less arbitrary, it's tuned to trigger
322 * less than 10% of times).
324 if (cnt <= 2 || !(cnt % 16))
325 vhost_poll_queue(&vq->poll);
328 /* Expects to be always run from workqueue - which acts as
329 * read-size critical section for our kind of RCU. */
330 static void handle_tx(struct vhost_net *net)
332 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
333 struct vhost_virtqueue *vq = &nvq->vq;
336 struct msghdr msg = {
342 .msg_flags = MSG_DONTWAIT,
344 size_t len, total_len = 0;
348 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
349 bool zcopy, zcopy_used;
351 mutex_lock(&vq->mutex);
352 sock = vq->private_data;
356 vhost_disable_notify(&net->dev, vq);
358 hdr_size = nvq->vhost_hlen;
362 /* Release DMAs done buffers first */
364 vhost_zerocopy_signal_used(net, vq);
366 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
370 /* On error, stop handling until the next kick. */
371 if (unlikely(head < 0))
373 /* Nothing new? Wait for eventfd to tell us they refilled. */
374 if (head == vq->num) {
377 /* If more outstanding DMAs, queue the work.
378 * Handle upend_idx wrap around
380 num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
381 (nvq->upend_idx - nvq->done_idx) :
382 (nvq->upend_idx + UIO_MAXIOV -
384 if (unlikely(num_pends > VHOST_MAX_PEND))
386 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
387 vhost_disable_notify(&net->dev, vq);
393 vq_err(vq, "Unexpected descriptor format for TX: "
394 "out %d, int %d\n", out, in);
397 /* Skip header. TODO: support TSO. */
398 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
399 msg.msg_iovlen = out;
400 len = iov_length(vq->iov, out);
403 vq_err(vq, "Unexpected header len for TX: "
404 "%zd expected %zd\n",
405 iov_length(nvq->hdr, s), hdr_size);
409 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
410 && (nvq->upend_idx + 1) % UIO_MAXIOV !=
412 && vhost_net_tx_select_zcopy(net);
414 /* use msg_control to pass vhost zerocopy ubuf info to skb */
416 struct ubuf_info *ubuf;
417 ubuf = nvq->ubuf_info + nvq->upend_idx;
419 vq->heads[nvq->upend_idx].id = head;
420 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
421 ubuf->callback = vhost_zerocopy_callback;
422 ubuf->ctx = nvq->ubufs;
423 ubuf->desc = nvq->upend_idx;
424 msg.msg_control = ubuf;
425 msg.msg_controllen = sizeof(ubuf);
427 kref_get(&ubufs->kref);
428 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
430 msg.msg_control = NULL;
433 /* TODO: Check specific error and bomb out unless ENOBUFS? */
434 err = sock->ops->sendmsg(NULL, sock, &msg, len);
435 if (unlikely(err < 0)) {
437 vhost_net_ubuf_put(ubufs);
438 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
441 vhost_discard_vq_desc(vq, 1);
445 pr_debug("Truncated TX packet: "
446 " len %d != %zd\n", err, len);
448 vhost_add_used_and_signal(&net->dev, vq, head, 0);
450 vhost_zerocopy_signal_used(net, vq);
452 vhost_net_tx_packet(net);
453 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
454 vhost_poll_queue(&vq->poll);
459 mutex_unlock(&vq->mutex);
462 static int peek_head_len(struct sock *sk)
464 struct sk_buff *head;
468 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
469 head = skb_peek(&sk->sk_receive_queue);
472 if (vlan_tx_tag_present(head))
476 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
480 /* This is a multi-buffer version of vhost_get_desc, that works if
481 * vq has read descriptors only.
482 * @vq - the relevant virtqueue
483 * @datalen - data length we'll be reading
484 * @iovcount - returned count of io vectors we fill
486 * @log_num - log offset
487 * @quota - headcount quota, 1 for big buffer
488 * returns number of buffer heads allocated, negative on error
490 static int get_rx_bufs(struct vhost_virtqueue *vq,
491 struct vring_used_elem *heads,
494 struct vhost_log *log,
498 unsigned int out, in;
504 while (datalen > 0 && headcount < quota) {
505 if (unlikely(seg >= UIO_MAXIOV)) {
509 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
510 ARRAY_SIZE(vq->iov) - seg, &out,
516 if (unlikely(out || in <= 0)) {
517 vq_err(vq, "unexpected descriptor format for RX: "
518 "out %d, in %d\n", out, in);
526 heads[headcount].id = d;
527 heads[headcount].len = iov_length(vq->iov + seg, in);
528 datalen -= heads[headcount].len;
532 heads[headcount - 1].len += datalen;
538 vhost_discard_vq_desc(vq, headcount);
542 /* Expects to be always run from workqueue - which acts as
543 * read-size critical section for our kind of RCU. */
544 static void handle_rx(struct vhost_net *net)
546 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
547 struct vhost_virtqueue *vq = &nvq->vq;
548 unsigned uninitialized_var(in), log;
549 struct vhost_log *vq_log;
550 struct msghdr msg = {
553 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
556 .msg_flags = MSG_DONTWAIT,
558 struct virtio_net_hdr_mrg_rxbuf hdr = {
560 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
562 size_t total_len = 0;
565 size_t vhost_hlen, sock_hlen;
566 size_t vhost_len, sock_len;
569 mutex_lock(&vq->mutex);
570 sock = vq->private_data;
573 vhost_disable_notify(&net->dev, vq);
575 vhost_hlen = nvq->vhost_hlen;
576 sock_hlen = nvq->sock_hlen;
578 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
580 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
582 while ((sock_len = peek_head_len(sock->sk))) {
583 sock_len += sock_hlen;
584 vhost_len = sock_len + vhost_hlen;
585 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
587 likely(mergeable) ? UIO_MAXIOV : 1);
588 /* On error, stop handling until the next kick. */
589 if (unlikely(headcount < 0))
591 /* OK, now we need to know about added descriptors. */
593 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
594 /* They have slipped one in as we were
595 * doing that: check again. */
596 vhost_disable_notify(&net->dev, vq);
599 /* Nothing new? Wait for eventfd to tell us
603 /* We don't need to be notified again. */
604 if (unlikely((vhost_hlen)))
605 /* Skip header. TODO: support TSO. */
606 move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
608 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
609 * needed because recvmsg can modify msg_iov. */
610 copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
612 err = sock->ops->recvmsg(NULL, sock, &msg,
613 sock_len, MSG_DONTWAIT | MSG_TRUNC);
614 /* Userspace might have consumed the packet meanwhile:
615 * it's not supposed to do this usually, but might be hard
616 * to prevent. Discard data we got (if any) and keep going. */
617 if (unlikely(err != sock_len)) {
618 pr_debug("Discarded rx packet: "
619 " len %d, expected %zd\n", err, sock_len);
620 vhost_discard_vq_desc(vq, headcount);
623 if (unlikely(vhost_hlen) &&
624 memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
626 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
630 /* TODO: Should check and handle checksum. */
631 if (likely(mergeable) &&
632 memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
633 offsetof(typeof(hdr), num_buffers),
634 sizeof hdr.num_buffers)) {
635 vq_err(vq, "Failed num_buffers write");
636 vhost_discard_vq_desc(vq, headcount);
639 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
641 if (unlikely(vq_log))
642 vhost_log_write(vq, vq_log, log, vhost_len);
643 total_len += vhost_len;
644 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
645 vhost_poll_queue(&vq->poll);
650 mutex_unlock(&vq->mutex);
653 static void handle_tx_kick(struct vhost_work *work)
655 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
657 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
662 static void handle_rx_kick(struct vhost_work *work)
664 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
666 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
671 static void handle_tx_net(struct vhost_work *work)
673 struct vhost_net *net = container_of(work, struct vhost_net,
674 poll[VHOST_NET_VQ_TX].work);
678 static void handle_rx_net(struct vhost_work *work)
680 struct vhost_net *net = container_of(work, struct vhost_net,
681 poll[VHOST_NET_VQ_RX].work);
685 static int vhost_net_open(struct inode *inode, struct file *f)
687 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
688 struct vhost_dev *dev;
689 struct vhost_virtqueue **vqs;
694 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
701 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
702 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
703 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
704 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
705 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
706 n->vqs[i].ubufs = NULL;
707 n->vqs[i].ubuf_info = NULL;
708 n->vqs[i].upend_idx = 0;
709 n->vqs[i].done_idx = 0;
710 n->vqs[i].vhost_hlen = 0;
711 n->vqs[i].sock_hlen = 0;
713 r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
720 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
721 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
728 static void vhost_net_disable_vq(struct vhost_net *n,
729 struct vhost_virtqueue *vq)
731 struct vhost_net_virtqueue *nvq =
732 container_of(vq, struct vhost_net_virtqueue, vq);
733 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
734 if (!vq->private_data)
736 vhost_poll_stop(poll);
739 static int vhost_net_enable_vq(struct vhost_net *n,
740 struct vhost_virtqueue *vq)
742 struct vhost_net_virtqueue *nvq =
743 container_of(vq, struct vhost_net_virtqueue, vq);
744 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
747 sock = vq->private_data;
751 return vhost_poll_start(poll, sock->file);
754 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
755 struct vhost_virtqueue *vq)
759 mutex_lock(&vq->mutex);
760 sock = vq->private_data;
761 vhost_net_disable_vq(n, vq);
762 vq->private_data = NULL;
763 mutex_unlock(&vq->mutex);
767 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
768 struct socket **rx_sock)
770 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
771 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
774 static void vhost_net_flush_vq(struct vhost_net *n, int index)
776 vhost_poll_flush(n->poll + index);
777 vhost_poll_flush(&n->vqs[index].vq.poll);
780 static void vhost_net_flush(struct vhost_net *n)
782 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
783 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
784 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
785 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
787 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
788 /* Wait for all lower device DMAs done. */
789 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
790 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
792 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
793 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
797 static int vhost_net_release(struct inode *inode, struct file *f)
799 struct vhost_net *n = f->private_data;
800 struct socket *tx_sock;
801 struct socket *rx_sock;
803 vhost_net_stop(n, &tx_sock, &rx_sock);
805 vhost_dev_stop(&n->dev);
806 vhost_dev_cleanup(&n->dev, false);
807 vhost_net_vq_reset(n);
812 /* We do an extra flush before freeing memory,
813 * since jobs can re-queue themselves. */
820 static struct socket *get_raw_socket(int fd)
823 struct sockaddr_ll sa;
824 char buf[MAX_ADDR_LEN];
826 int uaddr_len = sizeof uaddr, r;
827 struct socket *sock = sockfd_lookup(fd, &r);
830 return ERR_PTR(-ENOTSOCK);
832 /* Parameter checking */
833 if (sock->sk->sk_type != SOCK_RAW) {
834 r = -ESOCKTNOSUPPORT;
838 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
843 if (uaddr.sa.sll_family != AF_PACKET) {
853 static struct socket *get_tap_socket(int fd)
855 struct file *file = fget(fd);
859 return ERR_PTR(-EBADF);
860 sock = tun_get_socket(file);
863 sock = macvtap_get_socket(file);
869 static struct socket *get_socket(int fd)
873 /* special case to disable backend */
876 sock = get_raw_socket(fd);
879 sock = get_tap_socket(fd);
882 return ERR_PTR(-ENOTSOCK);
885 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
887 struct socket *sock, *oldsock;
888 struct vhost_virtqueue *vq;
889 struct vhost_net_virtqueue *nvq;
890 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
893 mutex_lock(&n->dev.mutex);
894 r = vhost_dev_check_owner(&n->dev);
898 if (index >= VHOST_NET_VQ_MAX) {
902 vq = &n->vqs[index].vq;
903 nvq = &n->vqs[index];
904 mutex_lock(&vq->mutex);
906 /* Verify that ring has been setup correctly. */
907 if (!vhost_vq_access_ok(vq)) {
911 sock = get_socket(fd);
917 /* start polling new socket */
918 oldsock = vq->private_data;
919 if (sock != oldsock) {
920 ubufs = vhost_net_ubuf_alloc(vq,
921 sock && vhost_sock_zcopy(sock));
927 vhost_net_disable_vq(n, vq);
928 vq->private_data = sock;
929 r = vhost_init_used(vq);
932 r = vhost_net_enable_vq(n, vq);
936 oldubufs = nvq->ubufs;
944 mutex_unlock(&vq->mutex);
947 vhost_net_ubuf_put_wait_and_free(oldubufs);
948 mutex_lock(&vq->mutex);
949 vhost_zerocopy_signal_used(n, vq);
950 mutex_unlock(&vq->mutex);
954 vhost_net_flush_vq(n, index);
958 mutex_unlock(&n->dev.mutex);
962 vq->private_data = oldsock;
963 vhost_net_enable_vq(n, vq);
965 vhost_net_ubuf_put_wait_and_free(ubufs);
969 mutex_unlock(&vq->mutex);
971 mutex_unlock(&n->dev.mutex);
975 static long vhost_net_reset_owner(struct vhost_net *n)
977 struct socket *tx_sock = NULL;
978 struct socket *rx_sock = NULL;
980 struct vhost_memory *memory;
982 mutex_lock(&n->dev.mutex);
983 err = vhost_dev_check_owner(&n->dev);
986 memory = vhost_dev_reset_owner_prepare();
991 vhost_net_stop(n, &tx_sock, &rx_sock);
993 vhost_dev_reset_owner(&n->dev, memory);
994 vhost_net_vq_reset(n);
996 mutex_unlock(&n->dev.mutex);
1000 fput(rx_sock->file);
1004 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1006 size_t vhost_hlen, sock_hlen, hdr_len;
1009 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1010 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1011 sizeof(struct virtio_net_hdr);
1012 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1013 /* vhost provides vnet_hdr */
1014 vhost_hlen = hdr_len;
1017 /* socket provides vnet_hdr */
1019 sock_hlen = hdr_len;
1021 mutex_lock(&n->dev.mutex);
1022 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1023 !vhost_log_access_ok(&n->dev)) {
1024 mutex_unlock(&n->dev.mutex);
1027 n->dev.acked_features = features;
1029 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1030 mutex_lock(&n->vqs[i].vq.mutex);
1031 n->vqs[i].vhost_hlen = vhost_hlen;
1032 n->vqs[i].sock_hlen = sock_hlen;
1033 mutex_unlock(&n->vqs[i].vq.mutex);
1036 mutex_unlock(&n->dev.mutex);
1040 static long vhost_net_set_owner(struct vhost_net *n)
1044 mutex_lock(&n->dev.mutex);
1045 if (vhost_dev_has_owner(&n->dev)) {
1049 r = vhost_net_set_ubuf_info(n);
1052 r = vhost_dev_set_owner(&n->dev);
1054 vhost_net_clear_ubuf_info(n);
1057 mutex_unlock(&n->dev.mutex);
1061 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1064 struct vhost_net *n = f->private_data;
1065 void __user *argp = (void __user *)arg;
1066 u64 __user *featurep = argp;
1067 struct vhost_vring_file backend;
1072 case VHOST_NET_SET_BACKEND:
1073 if (copy_from_user(&backend, argp, sizeof backend))
1075 return vhost_net_set_backend(n, backend.index, backend.fd);
1076 case VHOST_GET_FEATURES:
1077 features = VHOST_NET_FEATURES;
1078 if (copy_to_user(featurep, &features, sizeof features))
1081 case VHOST_SET_FEATURES:
1082 if (copy_from_user(&features, featurep, sizeof features))
1084 if (features & ~VHOST_NET_FEATURES)
1086 return vhost_net_set_features(n, features);
1087 case VHOST_RESET_OWNER:
1088 return vhost_net_reset_owner(n);
1089 case VHOST_SET_OWNER:
1090 return vhost_net_set_owner(n);
1092 mutex_lock(&n->dev.mutex);
1093 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1094 if (r == -ENOIOCTLCMD)
1095 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1098 mutex_unlock(&n->dev.mutex);
1103 #ifdef CONFIG_COMPAT
1104 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1107 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1111 static const struct file_operations vhost_net_fops = {
1112 .owner = THIS_MODULE,
1113 .release = vhost_net_release,
1114 .unlocked_ioctl = vhost_net_ioctl,
1115 #ifdef CONFIG_COMPAT
1116 .compat_ioctl = vhost_net_compat_ioctl,
1118 .open = vhost_net_open,
1119 .llseek = noop_llseek,
1122 static struct miscdevice vhost_net_misc = {
1123 .minor = VHOST_NET_MINOR,
1124 .name = "vhost-net",
1125 .fops = &vhost_net_fops,
1128 static int vhost_net_init(void)
1130 if (experimental_zcopytx)
1131 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1132 return misc_register(&vhost_net_misc);
1134 module_init(vhost_net_init);
1136 static void vhost_net_exit(void)
1138 misc_deregister(&vhost_net_misc);
1140 module_exit(vhost_net_exit);
1142 MODULE_VERSION("0.0.1");
1143 MODULE_LICENSE("GPL v2");
1144 MODULE_AUTHOR("Michael S. Tsirkin");
1145 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1146 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1147 MODULE_ALIAS("devname:vhost-net");