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/sched/clock.h>
21 #include <linux/sched/signal.h>
22 #include <linux/vmalloc.h>
24 #include <linux/net.h>
25 #include <linux/if_packet.h>
26 #include <linux/if_arp.h>
27 #include <linux/if_tun.h>
28 #include <linux/if_macvlan.h>
29 #include <linux/if_tap.h>
30 #include <linux/if_vlan.h>
31 #include <linux/skb_array.h>
32 #include <linux/skbuff.h>
38 static int experimental_zcopytx = 1;
39 module_param(experimental_zcopytx, int, 0444);
40 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
41 " 1 -Enable; 0 - Disable");
43 /* Max number of bytes transferred before requeueing the job.
44 * Using this limit prevents one virtqueue from starving others. */
45 #define VHOST_NET_WEIGHT 0x80000
47 /* MAX number of TX used buffers for outstanding zerocopy */
48 #define VHOST_MAX_PEND 128
49 #define VHOST_GOODCOPY_LEN 256
52 * For transmit, used buffer len is unused; we override it to track buffer
53 * status internally; used for zerocopy tx only.
55 /* Lower device DMA failed */
56 #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3)
57 /* Lower device DMA done */
58 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
59 /* Lower device DMA in progress */
60 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
62 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
64 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
67 VHOST_NET_FEATURES = VHOST_FEATURES |
68 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
69 (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
70 (1ULL << VIRTIO_F_IOMMU_PLATFORM)
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 #define VHOST_RX_BATCH 64
91 struct vhost_net_buf {
92 struct sk_buff **queue;
97 struct vhost_net_virtqueue {
98 struct vhost_virtqueue vq;
101 /* vhost zerocopy support fields below: */
102 /* last used idx for outstanding DMA zerocopy buffers */
104 /* first used idx for DMA done zerocopy buffers */
106 /* an array of userspace buffers info */
107 struct ubuf_info *ubuf_info;
108 /* Reference counting for outstanding ubufs.
109 * Protected by vq mutex. Writers must also take device mutex. */
110 struct vhost_net_ubuf_ref *ubufs;
111 struct skb_array *rx_array;
112 struct vhost_net_buf rxq;
116 struct vhost_dev dev;
117 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
118 struct vhost_poll poll[VHOST_NET_VQ_MAX];
119 /* Number of TX recently submitted.
120 * Protected by tx vq lock. */
122 /* Number of times zerocopy TX recently failed.
123 * Protected by tx vq lock. */
124 unsigned tx_zcopy_err;
125 /* Flush in progress. Protected by tx vq lock. */
129 static unsigned vhost_net_zcopy_mask __read_mostly;
131 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
133 if (rxq->tail != rxq->head)
134 return rxq->queue[rxq->head];
139 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
141 return rxq->tail - rxq->head;
144 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
146 return rxq->tail == rxq->head;
149 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
151 void *ret = vhost_net_buf_get_ptr(rxq);
156 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
158 struct vhost_net_buf *rxq = &nvq->rxq;
161 rxq->tail = skb_array_consume_batched(nvq->rx_array, rxq->queue,
166 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
168 struct vhost_net_buf *rxq = &nvq->rxq;
170 if (nvq->rx_array && !vhost_net_buf_is_empty(rxq)) {
171 skb_array_unconsume(nvq->rx_array, rxq->queue + rxq->head,
172 vhost_net_buf_get_size(rxq));
173 rxq->head = rxq->tail = 0;
177 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
179 struct vhost_net_buf *rxq = &nvq->rxq;
181 if (!vhost_net_buf_is_empty(rxq))
184 if (!vhost_net_buf_produce(nvq))
188 return __skb_array_len_with_tag(vhost_net_buf_get_ptr(rxq));
191 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
193 rxq->head = rxq->tail = 0;
196 static void vhost_net_enable_zcopy(int vq)
198 vhost_net_zcopy_mask |= 0x1 << vq;
201 static struct vhost_net_ubuf_ref *
202 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
204 struct vhost_net_ubuf_ref *ubufs;
205 /* No zero copy backend? Nothing to count. */
208 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
210 return ERR_PTR(-ENOMEM);
211 atomic_set(&ubufs->refcount, 1);
212 init_waitqueue_head(&ubufs->wait);
217 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
219 int r = atomic_sub_return(1, &ubufs->refcount);
221 wake_up(&ubufs->wait);
225 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
227 vhost_net_ubuf_put(ubufs);
228 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
231 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
233 vhost_net_ubuf_put_and_wait(ubufs);
237 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
241 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
242 kfree(n->vqs[i].ubuf_info);
243 n->vqs[i].ubuf_info = NULL;
247 static int vhost_net_set_ubuf_info(struct vhost_net *n)
252 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
253 zcopy = vhost_net_zcopy_mask & (0x1 << i);
256 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
257 UIO_MAXIOV, GFP_KERNEL);
258 if (!n->vqs[i].ubuf_info)
264 vhost_net_clear_ubuf_info(n);
268 static void vhost_net_vq_reset(struct vhost_net *n)
272 vhost_net_clear_ubuf_info(n);
274 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
275 n->vqs[i].done_idx = 0;
276 n->vqs[i].upend_idx = 0;
277 n->vqs[i].ubufs = NULL;
278 n->vqs[i].vhost_hlen = 0;
279 n->vqs[i].sock_hlen = 0;
280 vhost_net_buf_init(&n->vqs[i].rxq);
285 static void vhost_net_tx_packet(struct vhost_net *net)
288 if (net->tx_packets < 1024)
291 net->tx_zcopy_err = 0;
294 static void vhost_net_tx_err(struct vhost_net *net)
299 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
301 /* TX flush waits for outstanding DMAs to be done.
302 * Don't start new DMAs.
304 return !net->tx_flush &&
305 net->tx_packets / 64 >= net->tx_zcopy_err;
308 static bool vhost_sock_zcopy(struct socket *sock)
310 return unlikely(experimental_zcopytx) &&
311 sock_flag(sock->sk, SOCK_ZEROCOPY);
314 /* In case of DMA done not in order in lower device driver for some reason.
315 * upend_idx is used to track end of used idx, done_idx is used to track head
316 * of used idx. Once lower device DMA done contiguously, we will signal KVM
319 static void vhost_zerocopy_signal_used(struct vhost_net *net,
320 struct vhost_virtqueue *vq)
322 struct vhost_net_virtqueue *nvq =
323 container_of(vq, struct vhost_net_virtqueue, vq);
327 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
328 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
329 vhost_net_tx_err(net);
330 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
331 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
337 add = min(UIO_MAXIOV - nvq->done_idx, j);
338 vhost_add_used_and_signal_n(vq->dev, vq,
339 &vq->heads[nvq->done_idx], add);
340 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
345 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
347 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
348 struct vhost_virtqueue *vq = ubufs->vq;
353 /* set len to mark this desc buffers done DMA */
354 vq->heads[ubuf->desc].len = success ?
355 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
356 cnt = vhost_net_ubuf_put(ubufs);
359 * Trigger polling thread if guest stopped submitting new buffers:
360 * in this case, the refcount after decrement will eventually reach 1.
361 * We also trigger polling periodically after each 16 packets
362 * (the value 16 here is more or less arbitrary, it's tuned to trigger
363 * less than 10% of times).
365 if (cnt <= 1 || !(cnt % 16))
366 vhost_poll_queue(&vq->poll);
368 rcu_read_unlock_bh();
371 static inline unsigned long busy_clock(void)
373 return local_clock() >> 10;
376 static bool vhost_can_busy_poll(struct vhost_dev *dev,
377 unsigned long endtime)
379 return likely(!need_resched()) &&
380 likely(!time_after(busy_clock(), endtime)) &&
381 likely(!signal_pending(current)) &&
382 !vhost_has_work(dev);
385 static void vhost_net_disable_vq(struct vhost_net *n,
386 struct vhost_virtqueue *vq)
388 struct vhost_net_virtqueue *nvq =
389 container_of(vq, struct vhost_net_virtqueue, vq);
390 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
391 if (!vq->private_data)
393 vhost_poll_stop(poll);
396 static int vhost_net_enable_vq(struct vhost_net *n,
397 struct vhost_virtqueue *vq)
399 struct vhost_net_virtqueue *nvq =
400 container_of(vq, struct vhost_net_virtqueue, vq);
401 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
404 sock = vq->private_data;
408 return vhost_poll_start(poll, sock->file);
411 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
412 struct vhost_virtqueue *vq,
413 struct iovec iov[], unsigned int iov_size,
414 unsigned int *out_num, unsigned int *in_num)
416 unsigned long uninitialized_var(endtime);
417 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
418 out_num, in_num, NULL, NULL);
420 if (r == vq->num && vq->busyloop_timeout) {
422 endtime = busy_clock() + vq->busyloop_timeout;
423 while (vhost_can_busy_poll(vq->dev, endtime) &&
424 vhost_vq_avail_empty(vq->dev, vq))
427 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
428 out_num, in_num, NULL, NULL);
434 static bool vhost_exceeds_maxpend(struct vhost_net *net)
436 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
437 struct vhost_virtqueue *vq = &nvq->vq;
439 return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
440 min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
443 /* Expects to be always run from workqueue - which acts as
444 * read-size critical section for our kind of RCU. */
445 static void handle_tx(struct vhost_net *net)
447 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
448 struct vhost_virtqueue *vq = &nvq->vq;
451 struct msghdr msg = {
456 .msg_flags = MSG_DONTWAIT,
458 size_t len, total_len = 0;
462 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
463 bool zcopy, zcopy_used;
465 mutex_lock(&vq->mutex);
466 sock = vq->private_data;
470 if (!vq_iotlb_prefetch(vq))
473 vhost_disable_notify(&net->dev, vq);
474 vhost_net_disable_vq(net, vq);
476 hdr_size = nvq->vhost_hlen;
480 /* Release DMAs done buffers first */
482 vhost_zerocopy_signal_used(net, vq);
485 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
488 /* On error, stop handling until the next kick. */
489 if (unlikely(head < 0))
491 /* Nothing new? Wait for eventfd to tell us they refilled. */
492 if (head == vq->num) {
493 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
494 vhost_disable_notify(&net->dev, vq);
500 vq_err(vq, "Unexpected descriptor format for TX: "
501 "out %d, int %d\n", out, in);
504 /* Skip header. TODO: support TSO. */
505 len = iov_length(vq->iov, out);
506 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
507 iov_iter_advance(&msg.msg_iter, hdr_size);
509 if (!msg_data_left(&msg)) {
510 vq_err(vq, "Unexpected header len for TX: "
511 "%zd expected %zd\n",
515 len = msg_data_left(&msg);
517 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
518 && !vhost_exceeds_maxpend(net)
519 && vhost_net_tx_select_zcopy(net);
521 /* use msg_control to pass vhost zerocopy ubuf info to skb */
523 struct ubuf_info *ubuf;
524 ubuf = nvq->ubuf_info + nvq->upend_idx;
526 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
527 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
528 ubuf->callback = vhost_zerocopy_callback;
529 ubuf->ctx = nvq->ubufs;
530 ubuf->desc = nvq->upend_idx;
531 refcount_set(&ubuf->refcnt, 1);
532 msg.msg_control = ubuf;
533 msg.msg_controllen = sizeof(ubuf);
535 atomic_inc(&ubufs->refcount);
536 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
538 msg.msg_control = NULL;
543 if (total_len < VHOST_NET_WEIGHT &&
544 !vhost_vq_avail_empty(&net->dev, vq) &&
545 likely(!vhost_exceeds_maxpend(net))) {
546 msg.msg_flags |= MSG_MORE;
548 msg.msg_flags &= ~MSG_MORE;
551 /* TODO: Check specific error and bomb out unless ENOBUFS? */
552 err = sock->ops->sendmsg(sock, &msg, len);
553 if (unlikely(err < 0)) {
555 vhost_net_ubuf_put(ubufs);
556 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
559 vhost_discard_vq_desc(vq, 1);
560 vhost_net_enable_vq(net, vq);
564 pr_debug("Truncated TX packet: "
565 " len %d != %zd\n", err, len);
567 vhost_add_used_and_signal(&net->dev, vq, head, 0);
569 vhost_zerocopy_signal_used(net, vq);
570 vhost_net_tx_packet(net);
571 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
572 vhost_poll_queue(&vq->poll);
577 mutex_unlock(&vq->mutex);
580 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
582 struct sk_buff *head;
587 return vhost_net_buf_peek(rvq);
589 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
590 head = skb_peek(&sk->sk_receive_queue);
593 if (skb_vlan_tag_present(head))
597 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
601 static int sk_has_rx_data(struct sock *sk)
603 struct socket *sock = sk->sk_socket;
605 if (sock->ops->peek_len)
606 return sock->ops->peek_len(sock);
608 return skb_queue_empty(&sk->sk_receive_queue);
611 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
613 struct vhost_net_virtqueue *rvq = &net->vqs[VHOST_NET_VQ_RX];
614 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
615 struct vhost_virtqueue *vq = &nvq->vq;
616 unsigned long uninitialized_var(endtime);
617 int len = peek_head_len(rvq, sk);
619 if (!len && vq->busyloop_timeout) {
620 /* Both tx vq and rx socket were polled here */
621 mutex_lock(&vq->mutex);
622 vhost_disable_notify(&net->dev, vq);
625 endtime = busy_clock() + vq->busyloop_timeout;
627 while (vhost_can_busy_poll(&net->dev, endtime) &&
628 !sk_has_rx_data(sk) &&
629 vhost_vq_avail_empty(&net->dev, vq))
634 if (!vhost_vq_avail_empty(&net->dev, vq))
635 vhost_poll_queue(&vq->poll);
636 else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
637 vhost_disable_notify(&net->dev, vq);
638 vhost_poll_queue(&vq->poll);
641 mutex_unlock(&vq->mutex);
643 len = peek_head_len(rvq, sk);
649 /* This is a multi-buffer version of vhost_get_desc, that works if
650 * vq has read descriptors only.
651 * @vq - the relevant virtqueue
652 * @datalen - data length we'll be reading
653 * @iovcount - returned count of io vectors we fill
655 * @log_num - log offset
656 * @quota - headcount quota, 1 for big buffer
657 * returns number of buffer heads allocated, negative on error
659 static int get_rx_bufs(struct vhost_virtqueue *vq,
660 struct vring_used_elem *heads,
663 struct vhost_log *log,
667 unsigned int out, in;
672 /* len is always initialized before use since we are always called with
675 u32 uninitialized_var(len);
677 while (datalen > 0 && headcount < quota) {
678 if (unlikely(seg >= UIO_MAXIOV)) {
682 r = vhost_get_vq_desc(vq, vq->iov + seg,
683 ARRAY_SIZE(vq->iov) - seg, &out,
693 if (unlikely(out || in <= 0)) {
694 vq_err(vq, "unexpected descriptor format for RX: "
695 "out %d, in %d\n", out, in);
703 heads[headcount].id = cpu_to_vhost32(vq, d);
704 len = iov_length(vq->iov + seg, in);
705 heads[headcount].len = cpu_to_vhost32(vq, len);
710 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
716 if (unlikely(datalen > 0)) {
722 vhost_discard_vq_desc(vq, headcount);
726 /* Expects to be always run from workqueue - which acts as
727 * read-size critical section for our kind of RCU. */
728 static void handle_rx(struct vhost_net *net)
730 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
731 struct vhost_virtqueue *vq = &nvq->vq;
732 unsigned uninitialized_var(in), log;
733 struct vhost_log *vq_log;
734 struct msghdr msg = {
737 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
739 .msg_flags = MSG_DONTWAIT,
741 struct virtio_net_hdr hdr = {
743 .gso_type = VIRTIO_NET_HDR_GSO_NONE
745 size_t total_len = 0;
748 size_t vhost_hlen, sock_hlen;
749 size_t vhost_len, sock_len;
751 struct iov_iter fixup;
752 __virtio16 num_buffers;
754 mutex_lock(&vq->mutex);
755 sock = vq->private_data;
759 if (!vq_iotlb_prefetch(vq))
762 vhost_disable_notify(&net->dev, vq);
763 vhost_net_disable_vq(net, vq);
765 vhost_hlen = nvq->vhost_hlen;
766 sock_hlen = nvq->sock_hlen;
768 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
770 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
772 while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
773 sock_len += sock_hlen;
774 vhost_len = sock_len + vhost_hlen;
775 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
777 likely(mergeable) ? UIO_MAXIOV : 1);
778 /* On error, stop handling until the next kick. */
779 if (unlikely(headcount < 0))
781 /* OK, now we need to know about added descriptors. */
783 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
784 /* They have slipped one in as we were
785 * doing that: check again. */
786 vhost_disable_notify(&net->dev, vq);
789 /* Nothing new? Wait for eventfd to tell us
794 msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
795 /* On overrun, truncate and discard */
796 if (unlikely(headcount > UIO_MAXIOV)) {
797 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
798 err = sock->ops->recvmsg(sock, &msg,
799 1, MSG_DONTWAIT | MSG_TRUNC);
800 pr_debug("Discarded rx packet: len %zd\n", sock_len);
803 /* We don't need to be notified again. */
804 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
805 fixup = msg.msg_iter;
806 if (unlikely((vhost_hlen))) {
807 /* We will supply the header ourselves
810 iov_iter_advance(&msg.msg_iter, vhost_hlen);
812 err = sock->ops->recvmsg(sock, &msg,
813 sock_len, MSG_DONTWAIT | MSG_TRUNC);
814 /* Userspace might have consumed the packet meanwhile:
815 * it's not supposed to do this usually, but might be hard
816 * to prevent. Discard data we got (if any) and keep going. */
817 if (unlikely(err != sock_len)) {
818 pr_debug("Discarded rx packet: "
819 " len %d, expected %zd\n", err, sock_len);
820 vhost_discard_vq_desc(vq, headcount);
823 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
824 if (unlikely(vhost_hlen)) {
825 if (copy_to_iter(&hdr, sizeof(hdr),
826 &fixup) != sizeof(hdr)) {
827 vq_err(vq, "Unable to write vnet_hdr "
828 "at addr %p\n", vq->iov->iov_base);
832 /* Header came from socket; we'll need to patch
833 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
835 iov_iter_advance(&fixup, sizeof(hdr));
837 /* TODO: Should check and handle checksum. */
839 num_buffers = cpu_to_vhost16(vq, headcount);
840 if (likely(mergeable) &&
841 copy_to_iter(&num_buffers, sizeof num_buffers,
842 &fixup) != sizeof num_buffers) {
843 vq_err(vq, "Failed num_buffers write");
844 vhost_discard_vq_desc(vq, headcount);
847 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
849 if (unlikely(vq_log))
850 vhost_log_write(vq, vq_log, log, vhost_len);
851 total_len += vhost_len;
852 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
853 vhost_poll_queue(&vq->poll);
857 vhost_net_enable_vq(net, vq);
859 mutex_unlock(&vq->mutex);
862 static void handle_tx_kick(struct vhost_work *work)
864 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
866 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
871 static void handle_rx_kick(struct vhost_work *work)
873 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
875 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
880 static void handle_tx_net(struct vhost_work *work)
882 struct vhost_net *net = container_of(work, struct vhost_net,
883 poll[VHOST_NET_VQ_TX].work);
887 static void handle_rx_net(struct vhost_work *work)
889 struct vhost_net *net = container_of(work, struct vhost_net,
890 poll[VHOST_NET_VQ_RX].work);
894 static int vhost_net_open(struct inode *inode, struct file *f)
897 struct vhost_dev *dev;
898 struct vhost_virtqueue **vqs;
899 struct sk_buff **queue;
902 n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
905 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
911 queue = kmalloc_array(VHOST_RX_BATCH, sizeof(struct sk_buff *),
918 n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
921 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
922 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
923 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
924 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
925 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
926 n->vqs[i].ubufs = NULL;
927 n->vqs[i].ubuf_info = NULL;
928 n->vqs[i].upend_idx = 0;
929 n->vqs[i].done_idx = 0;
930 n->vqs[i].vhost_hlen = 0;
931 n->vqs[i].sock_hlen = 0;
932 vhost_net_buf_init(&n->vqs[i].rxq);
934 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
936 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
937 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
944 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
945 struct vhost_virtqueue *vq)
948 struct vhost_net_virtqueue *nvq =
949 container_of(vq, struct vhost_net_virtqueue, vq);
951 mutex_lock(&vq->mutex);
952 sock = vq->private_data;
953 vhost_net_disable_vq(n, vq);
954 vq->private_data = NULL;
955 vhost_net_buf_unproduce(nvq);
956 mutex_unlock(&vq->mutex);
960 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
961 struct socket **rx_sock)
963 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
964 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
967 static void vhost_net_flush_vq(struct vhost_net *n, int index)
969 vhost_poll_flush(n->poll + index);
970 vhost_poll_flush(&n->vqs[index].vq.poll);
973 static void vhost_net_flush(struct vhost_net *n)
975 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
976 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
977 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
978 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
980 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
981 /* Wait for all lower device DMAs done. */
982 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
983 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
985 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
986 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
990 static int vhost_net_release(struct inode *inode, struct file *f)
992 struct vhost_net *n = f->private_data;
993 struct socket *tx_sock;
994 struct socket *rx_sock;
996 vhost_net_stop(n, &tx_sock, &rx_sock);
998 vhost_dev_stop(&n->dev);
999 vhost_dev_cleanup(&n->dev, false);
1000 vhost_net_vq_reset(n);
1002 sockfd_put(tx_sock);
1004 sockfd_put(rx_sock);
1005 /* Make sure no callbacks are outstanding */
1006 synchronize_rcu_bh();
1007 /* We do an extra flush before freeing memory,
1008 * since jobs can re-queue themselves. */
1010 kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1016 static struct socket *get_raw_socket(int fd)
1019 struct sockaddr_ll sa;
1020 char buf[MAX_ADDR_LEN];
1022 int uaddr_len = sizeof uaddr, r;
1023 struct socket *sock = sockfd_lookup(fd, &r);
1026 return ERR_PTR(-ENOTSOCK);
1028 /* Parameter checking */
1029 if (sock->sk->sk_type != SOCK_RAW) {
1030 r = -ESOCKTNOSUPPORT;
1034 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
1039 if (uaddr.sa.sll_family != AF_PACKET) {
1049 static struct skb_array *get_tap_skb_array(int fd)
1051 struct skb_array *array;
1052 struct file *file = fget(fd);
1056 array = tun_get_skb_array(file);
1059 array = tap_get_skb_array(file);
1068 static struct socket *get_tap_socket(int fd)
1070 struct file *file = fget(fd);
1071 struct socket *sock;
1074 return ERR_PTR(-EBADF);
1075 sock = tun_get_socket(file);
1078 sock = tap_get_socket(file);
1084 static struct socket *get_socket(int fd)
1086 struct socket *sock;
1088 /* special case to disable backend */
1091 sock = get_raw_socket(fd);
1094 sock = get_tap_socket(fd);
1097 return ERR_PTR(-ENOTSOCK);
1100 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1102 struct socket *sock, *oldsock;
1103 struct vhost_virtqueue *vq;
1104 struct vhost_net_virtqueue *nvq;
1105 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1108 mutex_lock(&n->dev.mutex);
1109 r = vhost_dev_check_owner(&n->dev);
1113 if (index >= VHOST_NET_VQ_MAX) {
1117 vq = &n->vqs[index].vq;
1118 nvq = &n->vqs[index];
1119 mutex_lock(&vq->mutex);
1121 /* Verify that ring has been setup correctly. */
1122 if (!vhost_vq_access_ok(vq)) {
1126 sock = get_socket(fd);
1132 /* start polling new socket */
1133 oldsock = vq->private_data;
1134 if (sock != oldsock) {
1135 ubufs = vhost_net_ubuf_alloc(vq,
1136 sock && vhost_sock_zcopy(sock));
1137 if (IS_ERR(ubufs)) {
1142 vhost_net_disable_vq(n, vq);
1143 vq->private_data = sock;
1144 vhost_net_buf_unproduce(nvq);
1145 if (index == VHOST_NET_VQ_RX)
1146 nvq->rx_array = get_tap_skb_array(fd);
1147 r = vhost_vq_init_access(vq);
1150 r = vhost_net_enable_vq(n, vq);
1154 oldubufs = nvq->ubufs;
1158 n->tx_zcopy_err = 0;
1159 n->tx_flush = false;
1162 mutex_unlock(&vq->mutex);
1165 vhost_net_ubuf_put_wait_and_free(oldubufs);
1166 mutex_lock(&vq->mutex);
1167 vhost_zerocopy_signal_used(n, vq);
1168 mutex_unlock(&vq->mutex);
1172 vhost_net_flush_vq(n, index);
1173 sockfd_put(oldsock);
1176 mutex_unlock(&n->dev.mutex);
1180 vq->private_data = oldsock;
1181 vhost_net_enable_vq(n, vq);
1183 vhost_net_ubuf_put_wait_and_free(ubufs);
1187 mutex_unlock(&vq->mutex);
1189 mutex_unlock(&n->dev.mutex);
1193 static long vhost_net_reset_owner(struct vhost_net *n)
1195 struct socket *tx_sock = NULL;
1196 struct socket *rx_sock = NULL;
1198 struct vhost_umem *umem;
1200 mutex_lock(&n->dev.mutex);
1201 err = vhost_dev_check_owner(&n->dev);
1204 umem = vhost_dev_reset_owner_prepare();
1209 vhost_net_stop(n, &tx_sock, &rx_sock);
1211 vhost_dev_reset_owner(&n->dev, umem);
1212 vhost_net_vq_reset(n);
1214 mutex_unlock(&n->dev.mutex);
1216 sockfd_put(tx_sock);
1218 sockfd_put(rx_sock);
1222 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1224 size_t vhost_hlen, sock_hlen, hdr_len;
1227 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1228 (1ULL << VIRTIO_F_VERSION_1))) ?
1229 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1230 sizeof(struct virtio_net_hdr);
1231 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1232 /* vhost provides vnet_hdr */
1233 vhost_hlen = hdr_len;
1236 /* socket provides vnet_hdr */
1238 sock_hlen = hdr_len;
1240 mutex_lock(&n->dev.mutex);
1241 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1242 !vhost_log_access_ok(&n->dev))
1245 if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1246 if (vhost_init_device_iotlb(&n->dev, true))
1250 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1251 mutex_lock(&n->vqs[i].vq.mutex);
1252 n->vqs[i].vq.acked_features = features;
1253 n->vqs[i].vhost_hlen = vhost_hlen;
1254 n->vqs[i].sock_hlen = sock_hlen;
1255 mutex_unlock(&n->vqs[i].vq.mutex);
1257 mutex_unlock(&n->dev.mutex);
1261 mutex_unlock(&n->dev.mutex);
1265 static long vhost_net_set_owner(struct vhost_net *n)
1269 mutex_lock(&n->dev.mutex);
1270 if (vhost_dev_has_owner(&n->dev)) {
1274 r = vhost_net_set_ubuf_info(n);
1277 r = vhost_dev_set_owner(&n->dev);
1279 vhost_net_clear_ubuf_info(n);
1282 mutex_unlock(&n->dev.mutex);
1286 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1289 struct vhost_net *n = f->private_data;
1290 void __user *argp = (void __user *)arg;
1291 u64 __user *featurep = argp;
1292 struct vhost_vring_file backend;
1297 case VHOST_NET_SET_BACKEND:
1298 if (copy_from_user(&backend, argp, sizeof backend))
1300 return vhost_net_set_backend(n, backend.index, backend.fd);
1301 case VHOST_GET_FEATURES:
1302 features = VHOST_NET_FEATURES;
1303 if (copy_to_user(featurep, &features, sizeof features))
1306 case VHOST_SET_FEATURES:
1307 if (copy_from_user(&features, featurep, sizeof features))
1309 if (features & ~VHOST_NET_FEATURES)
1311 return vhost_net_set_features(n, features);
1312 case VHOST_RESET_OWNER:
1313 return vhost_net_reset_owner(n);
1314 case VHOST_SET_OWNER:
1315 return vhost_net_set_owner(n);
1317 mutex_lock(&n->dev.mutex);
1318 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1319 if (r == -ENOIOCTLCMD)
1320 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1323 mutex_unlock(&n->dev.mutex);
1328 #ifdef CONFIG_COMPAT
1329 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1332 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1336 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1338 struct file *file = iocb->ki_filp;
1339 struct vhost_net *n = file->private_data;
1340 struct vhost_dev *dev = &n->dev;
1341 int noblock = file->f_flags & O_NONBLOCK;
1343 return vhost_chr_read_iter(dev, to, noblock);
1346 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1347 struct iov_iter *from)
1349 struct file *file = iocb->ki_filp;
1350 struct vhost_net *n = file->private_data;
1351 struct vhost_dev *dev = &n->dev;
1353 return vhost_chr_write_iter(dev, from);
1356 static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait)
1358 struct vhost_net *n = file->private_data;
1359 struct vhost_dev *dev = &n->dev;
1361 return vhost_chr_poll(file, dev, wait);
1364 static const struct file_operations vhost_net_fops = {
1365 .owner = THIS_MODULE,
1366 .release = vhost_net_release,
1367 .read_iter = vhost_net_chr_read_iter,
1368 .write_iter = vhost_net_chr_write_iter,
1369 .poll = vhost_net_chr_poll,
1370 .unlocked_ioctl = vhost_net_ioctl,
1371 #ifdef CONFIG_COMPAT
1372 .compat_ioctl = vhost_net_compat_ioctl,
1374 .open = vhost_net_open,
1375 .llseek = noop_llseek,
1378 static struct miscdevice vhost_net_misc = {
1379 .minor = VHOST_NET_MINOR,
1380 .name = "vhost-net",
1381 .fops = &vhost_net_fops,
1384 static int vhost_net_init(void)
1386 if (experimental_zcopytx)
1387 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1388 return misc_register(&vhost_net_misc);
1390 module_init(vhost_net_init);
1392 static void vhost_net_exit(void)
1394 misc_deregister(&vhost_net_misc);
1396 module_exit(vhost_net_exit);
1398 MODULE_VERSION("0.0.1");
1399 MODULE_LICENSE("GPL v2");
1400 MODULE_AUTHOR("Michael S. Tsirkin");
1401 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1402 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1403 MODULE_ALIAS("devname:vhost-net");