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>
20 #include <linux/vmalloc.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 = 0;
34 module_param(experimental_zcopytx, int, 0444);
35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
36 " 1 -Enable; 0 - Disable");
38 /* Max number of bytes transferred before requeueing the job.
39 * Using this limit prevents one virtqueue from starving others. */
40 #define VHOST_NET_WEIGHT 0x80000
42 /* Max number of packets transferred before requeueing the job.
43 * Using this limit prevents one virtqueue from starving others with small
46 #define VHOST_NET_PKT_WEIGHT 256
48 /* MAX number of TX used buffers for outstanding zerocopy */
49 #define VHOST_MAX_PEND 128
50 #define VHOST_GOODCOPY_LEN 256
53 * For transmit, used buffer len is unused; we override it to track buffer
54 * status internally; used for zerocopy tx only.
56 /* Lower device DMA failed */
57 #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3)
58 /* Lower device DMA done */
59 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
60 /* Lower device DMA in progress */
61 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
63 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
65 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
68 VHOST_NET_FEATURES = VHOST_FEATURES |
69 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
70 (1ULL << VIRTIO_NET_F_MRG_RXBUF)
79 struct vhost_net_ubuf_ref {
80 /* refcount follows semantics similar to kref:
81 * 0: object is released
82 * 1: no outstanding ubufs
83 * >1: outstanding ubufs
86 wait_queue_head_t wait;
87 struct vhost_virtqueue *vq;
90 struct vhost_net_virtqueue {
91 struct vhost_virtqueue vq;
94 /* vhost zerocopy support fields below: */
95 /* last used idx for outstanding DMA zerocopy buffers */
97 /* first used idx for DMA done zerocopy buffers */
99 /* an array of userspace buffers info */
100 struct ubuf_info *ubuf_info;
101 /* Reference counting for outstanding ubufs.
102 * Protected by vq mutex. Writers must also take device mutex. */
103 struct vhost_net_ubuf_ref *ubufs;
107 struct vhost_dev dev;
108 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
109 struct vhost_poll poll[VHOST_NET_VQ_MAX];
110 /* Number of TX recently submitted.
111 * Protected by tx vq lock. */
113 /* Number of times zerocopy TX recently failed.
114 * Protected by tx vq lock. */
115 unsigned tx_zcopy_err;
116 /* Flush in progress. Protected by tx vq lock. */
120 static unsigned vhost_net_zcopy_mask __read_mostly;
122 static void vhost_net_enable_zcopy(int vq)
124 vhost_net_zcopy_mask |= 0x1 << vq;
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 atomic_set(&ubufs->refcount, 1);
138 init_waitqueue_head(&ubufs->wait);
143 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
145 int r = atomic_sub_return(1, &ubufs->refcount);
147 wake_up(&ubufs->wait);
151 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
153 vhost_net_ubuf_put(ubufs);
154 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
157 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
159 vhost_net_ubuf_put_and_wait(ubufs);
163 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
167 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
168 kfree(n->vqs[i].ubuf_info);
169 n->vqs[i].ubuf_info = NULL;
173 static int vhost_net_set_ubuf_info(struct vhost_net *n)
178 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
179 zcopy = vhost_net_zcopy_mask & (0x1 << i);
182 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
183 UIO_MAXIOV, GFP_KERNEL);
184 if (!n->vqs[i].ubuf_info)
190 vhost_net_clear_ubuf_info(n);
194 static void vhost_net_vq_reset(struct vhost_net *n)
198 vhost_net_clear_ubuf_info(n);
200 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
201 n->vqs[i].done_idx = 0;
202 n->vqs[i].upend_idx = 0;
203 n->vqs[i].ubufs = NULL;
204 n->vqs[i].vhost_hlen = 0;
205 n->vqs[i].sock_hlen = 0;
210 static void vhost_net_tx_packet(struct vhost_net *net)
213 if (net->tx_packets < 1024)
216 net->tx_zcopy_err = 0;
219 static void vhost_net_tx_err(struct vhost_net *net)
224 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
226 /* TX flush waits for outstanding DMAs to be done.
227 * Don't start new DMAs.
229 return !net->tx_flush &&
230 net->tx_packets / 64 >= net->tx_zcopy_err;
233 static bool vhost_sock_zcopy(struct socket *sock)
235 return unlikely(experimental_zcopytx) &&
236 sock_flag(sock->sk, SOCK_ZEROCOPY);
239 /* In case of DMA done not in order in lower device driver for some reason.
240 * upend_idx is used to track end of used idx, done_idx is used to track head
241 * of used idx. Once lower device DMA done contiguously, we will signal KVM
244 static void vhost_zerocopy_signal_used(struct vhost_net *net,
245 struct vhost_virtqueue *vq)
247 struct vhost_net_virtqueue *nvq =
248 container_of(vq, struct vhost_net_virtqueue, vq);
252 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
253 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
254 vhost_net_tx_err(net);
255 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
256 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
262 add = min(UIO_MAXIOV - nvq->done_idx, j);
263 vhost_add_used_and_signal_n(vq->dev, vq,
264 &vq->heads[nvq->done_idx], add);
265 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
270 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
272 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
273 struct vhost_virtqueue *vq = ubufs->vq;
278 /* set len to mark this desc buffers done DMA */
279 vq->heads[ubuf->desc].len = success ?
280 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
281 cnt = vhost_net_ubuf_put(ubufs);
284 * Trigger polling thread if guest stopped submitting new buffers:
285 * in this case, the refcount after decrement will eventually reach 1.
286 * We also trigger polling periodically after each 16 packets
287 * (the value 16 here is more or less arbitrary, it's tuned to trigger
288 * less than 10% of times).
290 if (cnt <= 1 || !(cnt % 16))
291 vhost_poll_queue(&vq->poll);
293 rcu_read_unlock_bh();
296 /* Expects to be always run from workqueue - which acts as
297 * read-size critical section for our kind of RCU. */
298 static void handle_tx(struct vhost_net *net)
300 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
301 struct vhost_virtqueue *vq = &nvq->vq;
304 struct msghdr msg = {
309 .msg_flags = MSG_DONTWAIT,
311 size_t len, total_len = 0;
315 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
316 struct ubuf_info *ubuf;
317 bool zcopy, zcopy_used;
320 mutex_lock(&vq->mutex);
321 sock = vq->private_data;
325 vhost_disable_notify(&net->dev, vq);
327 hdr_size = nvq->vhost_hlen;
331 /* Release DMAs done buffers first */
333 vhost_zerocopy_signal_used(net, vq);
335 /* If more outstanding DMAs, queue the work.
336 * Handle upend_idx wrap around
338 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
339 % UIO_MAXIOV == nvq->done_idx))
342 head = vhost_get_vq_desc(vq, vq->iov,
346 /* On error, stop handling until the next kick. */
347 if (unlikely(head < 0))
349 /* Nothing new? Wait for eventfd to tell us they refilled. */
350 if (head == vq->num) {
351 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
352 vhost_disable_notify(&net->dev, vq);
358 vq_err(vq, "Unexpected descriptor format for TX: "
359 "out %d, int %d\n", out, in);
362 /* Skip header. TODO: support TSO. */
363 len = iov_length(vq->iov, out);
364 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
365 iov_iter_advance(&msg.msg_iter, hdr_size);
367 if (!msg_data_left(&msg)) {
368 vq_err(vq, "Unexpected header len for TX: "
369 "%zd expected %zd\n",
373 len = msg_data_left(&msg);
375 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
376 && (nvq->upend_idx + 1) % UIO_MAXIOV !=
378 && vhost_net_tx_select_zcopy(net);
380 /* use msg_control to pass vhost zerocopy ubuf info to skb */
382 ubuf = nvq->ubuf_info + nvq->upend_idx;
383 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
384 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
385 ubuf->callback = vhost_zerocopy_callback;
386 ubuf->ctx = nvq->ubufs;
387 ubuf->desc = nvq->upend_idx;
388 msg.msg_control = ubuf;
389 msg.msg_controllen = sizeof(ubuf);
391 atomic_inc(&ubufs->refcount);
392 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
394 msg.msg_control = NULL;
397 /* TODO: Check specific error and bomb out unless ENOBUFS? */
398 err = sock->ops->sendmsg(sock, &msg, len);
399 if (unlikely(err < 0)) {
401 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
402 vhost_net_ubuf_put(ubufs);
403 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
406 vhost_discard_vq_desc(vq, 1);
410 pr_debug("Truncated TX packet: "
411 " len %d != %zd\n", err, len);
413 vhost_add_used_and_signal(&net->dev, vq, head, 0);
415 vhost_zerocopy_signal_used(net, vq);
417 vhost_net_tx_packet(net);
418 } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
420 mutex_unlock(&vq->mutex);
423 static int peek_head_len(struct sock *sk)
425 struct sk_buff *head;
429 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
430 head = skb_peek(&sk->sk_receive_queue);
433 if (skb_vlan_tag_present(head))
437 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
441 /* This is a multi-buffer version of vhost_get_desc, that works if
442 * vq has read descriptors only.
443 * @vq - the relevant virtqueue
444 * @datalen - data length we'll be reading
445 * @iovcount - returned count of io vectors we fill
447 * @log_num - log offset
448 * @quota - headcount quota, 1 for big buffer
449 * returns number of buffer heads allocated, negative on error
451 static int get_rx_bufs(struct vhost_virtqueue *vq,
452 struct vring_used_elem *heads,
455 struct vhost_log *log,
459 unsigned int out, in;
464 /* len is always initialized before use since we are always called with
467 u32 uninitialized_var(len);
469 while (datalen > 0 && headcount < quota) {
470 if (unlikely(seg >= UIO_MAXIOV)) {
474 r = vhost_get_vq_desc(vq, vq->iov + seg,
475 ARRAY_SIZE(vq->iov) - seg, &out,
485 if (unlikely(out || in <= 0)) {
486 vq_err(vq, "unexpected descriptor format for RX: "
487 "out %d, in %d\n", out, in);
495 heads[headcount].id = cpu_to_vhost32(vq, d);
496 len = iov_length(vq->iov + seg, in);
497 heads[headcount].len = cpu_to_vhost32(vq, len);
502 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
508 if (unlikely(datalen > 0)) {
514 vhost_discard_vq_desc(vq, headcount);
518 /* Expects to be always run from workqueue - which acts as
519 * read-size critical section for our kind of RCU. */
520 static void handle_rx(struct vhost_net *net)
522 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
523 struct vhost_virtqueue *vq = &nvq->vq;
524 unsigned uninitialized_var(in), log;
525 struct vhost_log *vq_log;
526 struct msghdr msg = {
529 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
531 .msg_flags = MSG_DONTWAIT,
533 struct virtio_net_hdr hdr = {
535 .gso_type = VIRTIO_NET_HDR_GSO_NONE
537 size_t total_len = 0;
540 size_t vhost_hlen, sock_hlen;
541 size_t vhost_len, sock_len;
543 struct iov_iter fixup;
544 __virtio16 num_buffers;
547 mutex_lock(&vq->mutex);
548 sock = vq->private_data;
551 vhost_disable_notify(&net->dev, vq);
553 vhost_hlen = nvq->vhost_hlen;
554 sock_hlen = nvq->sock_hlen;
556 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
558 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
561 sock_len = peek_head_len(sock->sk);
564 sock_len += sock_hlen;
565 vhost_len = sock_len + vhost_hlen;
566 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
568 likely(mergeable) ? UIO_MAXIOV : 1);
569 /* On error, stop handling until the next kick. */
570 if (unlikely(headcount < 0))
572 /* On overrun, truncate and discard */
573 if (unlikely(headcount > UIO_MAXIOV)) {
574 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
575 err = sock->ops->recvmsg(sock, &msg,
576 1, MSG_DONTWAIT | MSG_TRUNC);
577 pr_debug("Discarded rx packet: len %zd\n", sock_len);
580 /* OK, now we need to know about added descriptors. */
582 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
583 /* They have slipped one in as we were
584 * doing that: check again. */
585 vhost_disable_notify(&net->dev, vq);
588 /* Nothing new? Wait for eventfd to tell us
592 /* We don't need to be notified again. */
593 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
594 fixup = msg.msg_iter;
595 if (unlikely((vhost_hlen))) {
596 /* We will supply the header ourselves
599 iov_iter_advance(&msg.msg_iter, vhost_hlen);
601 err = sock->ops->recvmsg(sock, &msg,
602 sock_len, MSG_DONTWAIT | MSG_TRUNC);
603 /* Userspace might have consumed the packet meanwhile:
604 * it's not supposed to do this usually, but might be hard
605 * to prevent. Discard data we got (if any) and keep going. */
606 if (unlikely(err != sock_len)) {
607 pr_debug("Discarded rx packet: "
608 " len %d, expected %zd\n", err, sock_len);
609 vhost_discard_vq_desc(vq, headcount);
612 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
613 if (unlikely(vhost_hlen)) {
614 if (copy_to_iter(&hdr, sizeof(hdr),
615 &fixup) != sizeof(hdr)) {
616 vq_err(vq, "Unable to write vnet_hdr "
617 "at addr %p\n", vq->iov->iov_base);
621 /* Header came from socket; we'll need to patch
622 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
624 iov_iter_advance(&fixup, sizeof(hdr));
626 /* TODO: Should check and handle checksum. */
628 num_buffers = cpu_to_vhost16(vq, headcount);
629 if (likely(mergeable) &&
630 copy_to_iter(&num_buffers, sizeof num_buffers,
631 &fixup) != sizeof num_buffers) {
632 vq_err(vq, "Failed num_buffers write");
633 vhost_discard_vq_desc(vq, headcount);
636 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
638 if (unlikely(vq_log))
639 vhost_log_write(vq, vq_log, log, vhost_len);
640 total_len += vhost_len;
641 } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
644 mutex_unlock(&vq->mutex);
647 static void handle_tx_kick(struct vhost_work *work)
649 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
651 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
656 static void handle_rx_kick(struct vhost_work *work)
658 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
660 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
665 static void handle_tx_net(struct vhost_work *work)
667 struct vhost_net *net = container_of(work, struct vhost_net,
668 poll[VHOST_NET_VQ_TX].work);
672 static void handle_rx_net(struct vhost_work *work)
674 struct vhost_net *net = container_of(work, struct vhost_net,
675 poll[VHOST_NET_VQ_RX].work);
679 static int vhost_net_open(struct inode *inode, struct file *f)
682 struct vhost_dev *dev;
683 struct vhost_virtqueue **vqs;
686 n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
688 n = vmalloc(sizeof *n);
692 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
699 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
700 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
701 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
702 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
703 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
704 n->vqs[i].ubufs = NULL;
705 n->vqs[i].ubuf_info = NULL;
706 n->vqs[i].upend_idx = 0;
707 n->vqs[i].done_idx = 0;
708 n->vqs[i].vhost_hlen = 0;
709 n->vqs[i].sock_hlen = 0;
711 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
712 VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT);
714 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
715 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
722 static void vhost_net_disable_vq(struct vhost_net *n,
723 struct vhost_virtqueue *vq)
725 struct vhost_net_virtqueue *nvq =
726 container_of(vq, struct vhost_net_virtqueue, vq);
727 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
728 if (!vq->private_data)
730 vhost_poll_stop(poll);
733 static int vhost_net_enable_vq(struct vhost_net *n,
734 struct vhost_virtqueue *vq)
736 struct vhost_net_virtqueue *nvq =
737 container_of(vq, struct vhost_net_virtqueue, vq);
738 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
741 sock = vq->private_data;
745 return vhost_poll_start(poll, sock->file);
748 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
749 struct vhost_virtqueue *vq)
753 mutex_lock(&vq->mutex);
754 sock = vq->private_data;
755 vhost_net_disable_vq(n, vq);
756 vq->private_data = NULL;
757 mutex_unlock(&vq->mutex);
761 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
762 struct socket **rx_sock)
764 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
765 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
768 static void vhost_net_flush_vq(struct vhost_net *n, int index)
770 vhost_poll_flush(n->poll + index);
771 vhost_poll_flush(&n->vqs[index].vq.poll);
774 static void vhost_net_flush(struct vhost_net *n)
776 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
777 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
778 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
779 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
781 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
782 /* Wait for all lower device DMAs done. */
783 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
784 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
786 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
787 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
791 static int vhost_net_release(struct inode *inode, struct file *f)
793 struct vhost_net *n = f->private_data;
794 struct socket *tx_sock;
795 struct socket *rx_sock;
797 vhost_net_stop(n, &tx_sock, &rx_sock);
799 vhost_dev_stop(&n->dev);
800 vhost_dev_cleanup(&n->dev, false);
801 vhost_net_vq_reset(n);
806 /* Make sure no callbacks are outstanding */
807 synchronize_rcu_bh();
808 /* We do an extra flush before freeing memory,
809 * since jobs can re-queue themselves. */
816 static struct socket *get_raw_socket(int fd)
819 struct socket *sock = sockfd_lookup(fd, &r);
822 return ERR_PTR(-ENOTSOCK);
824 /* Parameter checking */
825 if (sock->sk->sk_type != SOCK_RAW) {
826 r = -ESOCKTNOSUPPORT;
830 if (sock->sk->sk_family != AF_PACKET) {
840 static struct socket *get_tap_socket(int fd)
842 struct file *file = fget(fd);
846 return ERR_PTR(-EBADF);
847 sock = tun_get_socket(file);
850 sock = macvtap_get_socket(file);
856 static struct socket *get_socket(int fd)
860 /* special case to disable backend */
863 sock = get_raw_socket(fd);
866 sock = get_tap_socket(fd);
869 return ERR_PTR(-ENOTSOCK);
872 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
874 struct socket *sock, *oldsock;
875 struct vhost_virtqueue *vq;
876 struct vhost_net_virtqueue *nvq;
877 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
880 mutex_lock(&n->dev.mutex);
881 r = vhost_dev_check_owner(&n->dev);
885 if (index >= VHOST_NET_VQ_MAX) {
889 vq = &n->vqs[index].vq;
890 nvq = &n->vqs[index];
891 mutex_lock(&vq->mutex);
893 /* Verify that ring has been setup correctly. */
894 if (!vhost_vq_access_ok(vq)) {
898 sock = get_socket(fd);
904 /* start polling new socket */
905 oldsock = vq->private_data;
906 if (sock != oldsock) {
907 ubufs = vhost_net_ubuf_alloc(vq,
908 sock && vhost_sock_zcopy(sock));
914 vhost_net_disable_vq(n, vq);
915 vq->private_data = sock;
916 r = vhost_init_used(vq);
919 r = vhost_net_enable_vq(n, vq);
923 oldubufs = nvq->ubufs;
931 mutex_unlock(&vq->mutex);
934 vhost_net_ubuf_put_wait_and_free(oldubufs);
935 mutex_lock(&vq->mutex);
936 vhost_zerocopy_signal_used(n, vq);
937 mutex_unlock(&vq->mutex);
941 vhost_net_flush_vq(n, index);
945 mutex_unlock(&n->dev.mutex);
949 vq->private_data = oldsock;
950 vhost_net_enable_vq(n, vq);
952 vhost_net_ubuf_put_wait_and_free(ubufs);
957 mutex_unlock(&vq->mutex);
959 mutex_unlock(&n->dev.mutex);
963 static long vhost_net_reset_owner(struct vhost_net *n)
965 struct socket *tx_sock = NULL;
966 struct socket *rx_sock = NULL;
968 struct vhost_memory *memory;
970 mutex_lock(&n->dev.mutex);
971 err = vhost_dev_check_owner(&n->dev);
974 memory = vhost_dev_reset_owner_prepare();
979 vhost_net_stop(n, &tx_sock, &rx_sock);
981 vhost_dev_stop(&n->dev);
982 vhost_dev_reset_owner(&n->dev, memory);
983 vhost_net_vq_reset(n);
985 mutex_unlock(&n->dev.mutex);
993 static int vhost_net_set_features(struct vhost_net *n, u64 features)
995 size_t vhost_hlen, sock_hlen, hdr_len;
998 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
999 (1ULL << VIRTIO_F_VERSION_1))) ?
1000 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1001 sizeof(struct virtio_net_hdr);
1002 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1003 /* vhost provides vnet_hdr */
1004 vhost_hlen = hdr_len;
1007 /* socket provides vnet_hdr */
1009 sock_hlen = hdr_len;
1011 mutex_lock(&n->dev.mutex);
1012 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1013 !vhost_log_access_ok(&n->dev)) {
1014 mutex_unlock(&n->dev.mutex);
1017 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1018 mutex_lock(&n->vqs[i].vq.mutex);
1019 n->vqs[i].vq.acked_features = features;
1020 n->vqs[i].vhost_hlen = vhost_hlen;
1021 n->vqs[i].sock_hlen = sock_hlen;
1022 mutex_unlock(&n->vqs[i].vq.mutex);
1024 mutex_unlock(&n->dev.mutex);
1028 static long vhost_net_set_owner(struct vhost_net *n)
1032 mutex_lock(&n->dev.mutex);
1033 if (vhost_dev_has_owner(&n->dev)) {
1037 r = vhost_net_set_ubuf_info(n);
1040 r = vhost_dev_set_owner(&n->dev);
1042 vhost_net_clear_ubuf_info(n);
1045 mutex_unlock(&n->dev.mutex);
1049 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1052 struct vhost_net *n = f->private_data;
1053 void __user *argp = (void __user *)arg;
1054 u64 __user *featurep = argp;
1055 struct vhost_vring_file backend;
1060 case VHOST_NET_SET_BACKEND:
1061 if (copy_from_user(&backend, argp, sizeof backend))
1063 return vhost_net_set_backend(n, backend.index, backend.fd);
1064 case VHOST_GET_FEATURES:
1065 features = VHOST_NET_FEATURES;
1066 if (copy_to_user(featurep, &features, sizeof features))
1069 case VHOST_SET_FEATURES:
1070 if (copy_from_user(&features, featurep, sizeof features))
1072 if (features & ~VHOST_NET_FEATURES)
1074 return vhost_net_set_features(n, features);
1075 case VHOST_RESET_OWNER:
1076 return vhost_net_reset_owner(n);
1077 case VHOST_SET_OWNER:
1078 return vhost_net_set_owner(n);
1080 mutex_lock(&n->dev.mutex);
1081 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1082 if (r == -ENOIOCTLCMD)
1083 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1086 mutex_unlock(&n->dev.mutex);
1091 #ifdef CONFIG_COMPAT
1092 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1095 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1099 static const struct file_operations vhost_net_fops = {
1100 .owner = THIS_MODULE,
1101 .release = vhost_net_release,
1102 .unlocked_ioctl = vhost_net_ioctl,
1103 #ifdef CONFIG_COMPAT
1104 .compat_ioctl = vhost_net_compat_ioctl,
1106 .open = vhost_net_open,
1107 .llseek = noop_llseek,
1110 static struct miscdevice vhost_net_misc = {
1111 .minor = VHOST_NET_MINOR,
1112 .name = "vhost-net",
1113 .fops = &vhost_net_fops,
1116 static int vhost_net_init(void)
1118 if (experimental_zcopytx)
1119 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1120 return misc_register(&vhost_net_misc);
1122 module_init(vhost_net_init);
1124 static void vhost_net_exit(void)
1126 misc_deregister(&vhost_net_misc);
1128 module_exit(vhost_net_exit);
1130 MODULE_VERSION("0.0.1");
1131 MODULE_LICENSE("GPL v2");
1132 MODULE_AUTHOR("Michael S. Tsirkin");
1133 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1134 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1135 MODULE_ALIAS("devname:vhost-net");