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 = 0;
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 packets transferred before requeueing the job.
48 * Using this limit prevents one virtqueue from starving others with small
51 #define VHOST_NET_PKT_WEIGHT 256
53 /* MAX number of TX used buffers for outstanding zerocopy */
54 #define VHOST_MAX_PEND 128
55 #define VHOST_GOODCOPY_LEN 256
58 * For transmit, used buffer len is unused; we override it to track buffer
59 * status internally; used for zerocopy tx only.
61 /* Lower device DMA failed */
62 #define VHOST_DMA_FAILED_LEN ((__force __virtio32)3)
63 /* Lower device DMA done */
64 #define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
65 /* Lower device DMA in progress */
66 #define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
68 #define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
70 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
73 VHOST_NET_FEATURES = VHOST_FEATURES |
74 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
75 (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
76 (1ULL << VIRTIO_F_IOMMU_PLATFORM)
85 struct vhost_net_ubuf_ref {
86 /* refcount follows semantics similar to kref:
87 * 0: object is released
88 * 1: no outstanding ubufs
89 * >1: outstanding ubufs
92 wait_queue_head_t wait;
93 struct vhost_virtqueue *vq;
96 #define VHOST_RX_BATCH 64
97 struct vhost_net_buf {
98 struct sk_buff **queue;
103 struct vhost_net_virtqueue {
104 struct vhost_virtqueue vq;
107 /* vhost zerocopy support fields below: */
108 /* last used idx for outstanding DMA zerocopy buffers */
110 /* first used idx for DMA done zerocopy buffers */
112 /* an array of userspace buffers info */
113 struct ubuf_info *ubuf_info;
114 /* Reference counting for outstanding ubufs.
115 * Protected by vq mutex. Writers must also take device mutex. */
116 struct vhost_net_ubuf_ref *ubufs;
117 struct skb_array *rx_array;
118 struct vhost_net_buf rxq;
122 struct vhost_dev dev;
123 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
124 struct vhost_poll poll[VHOST_NET_VQ_MAX];
125 /* Number of TX recently submitted.
126 * Protected by tx vq lock. */
128 /* Number of times zerocopy TX recently failed.
129 * Protected by tx vq lock. */
130 unsigned tx_zcopy_err;
131 /* Flush in progress. Protected by tx vq lock. */
135 static unsigned vhost_net_zcopy_mask __read_mostly;
137 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
139 if (rxq->tail != rxq->head)
140 return rxq->queue[rxq->head];
145 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
147 return rxq->tail - rxq->head;
150 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
152 return rxq->tail == rxq->head;
155 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
157 void *ret = vhost_net_buf_get_ptr(rxq);
162 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
164 struct vhost_net_buf *rxq = &nvq->rxq;
167 rxq->tail = skb_array_consume_batched(nvq->rx_array, rxq->queue,
172 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
174 struct vhost_net_buf *rxq = &nvq->rxq;
176 if (nvq->rx_array && !vhost_net_buf_is_empty(rxq)) {
177 skb_array_unconsume(nvq->rx_array, rxq->queue + rxq->head,
178 vhost_net_buf_get_size(rxq));
179 rxq->head = rxq->tail = 0;
183 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
185 struct vhost_net_buf *rxq = &nvq->rxq;
187 if (!vhost_net_buf_is_empty(rxq))
190 if (!vhost_net_buf_produce(nvq))
194 return __skb_array_len_with_tag(vhost_net_buf_get_ptr(rxq));
197 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
199 rxq->head = rxq->tail = 0;
202 static void vhost_net_enable_zcopy(int vq)
204 vhost_net_zcopy_mask |= 0x1 << vq;
207 static struct vhost_net_ubuf_ref *
208 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
210 struct vhost_net_ubuf_ref *ubufs;
211 /* No zero copy backend? Nothing to count. */
214 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
216 return ERR_PTR(-ENOMEM);
217 atomic_set(&ubufs->refcount, 1);
218 init_waitqueue_head(&ubufs->wait);
223 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
225 int r = atomic_sub_return(1, &ubufs->refcount);
227 wake_up(&ubufs->wait);
231 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
233 vhost_net_ubuf_put(ubufs);
234 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
237 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
239 vhost_net_ubuf_put_and_wait(ubufs);
243 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
247 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
248 kfree(n->vqs[i].ubuf_info);
249 n->vqs[i].ubuf_info = NULL;
253 static int vhost_net_set_ubuf_info(struct vhost_net *n)
258 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
259 zcopy = vhost_net_zcopy_mask & (0x1 << i);
262 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
263 UIO_MAXIOV, GFP_KERNEL);
264 if (!n->vqs[i].ubuf_info)
270 vhost_net_clear_ubuf_info(n);
274 static void vhost_net_vq_reset(struct vhost_net *n)
278 vhost_net_clear_ubuf_info(n);
280 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
281 n->vqs[i].done_idx = 0;
282 n->vqs[i].upend_idx = 0;
283 n->vqs[i].ubufs = NULL;
284 n->vqs[i].vhost_hlen = 0;
285 n->vqs[i].sock_hlen = 0;
286 vhost_net_buf_init(&n->vqs[i].rxq);
291 static void vhost_net_tx_packet(struct vhost_net *net)
294 if (net->tx_packets < 1024)
297 net->tx_zcopy_err = 0;
300 static void vhost_net_tx_err(struct vhost_net *net)
305 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
307 /* TX flush waits for outstanding DMAs to be done.
308 * Don't start new DMAs.
310 return !net->tx_flush &&
311 net->tx_packets / 64 >= net->tx_zcopy_err;
314 static bool vhost_sock_zcopy(struct socket *sock)
316 return unlikely(experimental_zcopytx) &&
317 sock_flag(sock->sk, SOCK_ZEROCOPY);
320 /* In case of DMA done not in order in lower device driver for some reason.
321 * upend_idx is used to track end of used idx, done_idx is used to track head
322 * of used idx. Once lower device DMA done contiguously, we will signal KVM
325 static void vhost_zerocopy_signal_used(struct vhost_net *net,
326 struct vhost_virtqueue *vq)
328 struct vhost_net_virtqueue *nvq =
329 container_of(vq, struct vhost_net_virtqueue, vq);
333 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
334 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
335 vhost_net_tx_err(net);
336 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
337 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
343 add = min(UIO_MAXIOV - nvq->done_idx, j);
344 vhost_add_used_and_signal_n(vq->dev, vq,
345 &vq->heads[nvq->done_idx], add);
346 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
351 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
353 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
354 struct vhost_virtqueue *vq = ubufs->vq;
359 /* set len to mark this desc buffers done DMA */
360 vq->heads[ubuf->desc].len = success ?
361 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
362 cnt = vhost_net_ubuf_put(ubufs);
365 * Trigger polling thread if guest stopped submitting new buffers:
366 * in this case, the refcount after decrement will eventually reach 1.
367 * We also trigger polling periodically after each 16 packets
368 * (the value 16 here is more or less arbitrary, it's tuned to trigger
369 * less than 10% of times).
371 if (cnt <= 1 || !(cnt % 16))
372 vhost_poll_queue(&vq->poll);
374 rcu_read_unlock_bh();
377 static inline unsigned long busy_clock(void)
379 return local_clock() >> 10;
382 static bool vhost_can_busy_poll(struct vhost_dev *dev,
383 unsigned long endtime)
385 return likely(!need_resched()) &&
386 likely(!time_after(busy_clock(), endtime)) &&
387 likely(!signal_pending(current)) &&
388 !vhost_has_work(dev);
391 static void vhost_net_disable_vq(struct vhost_net *n,
392 struct vhost_virtqueue *vq)
394 struct vhost_net_virtqueue *nvq =
395 container_of(vq, struct vhost_net_virtqueue, vq);
396 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
397 if (!vq->private_data)
399 vhost_poll_stop(poll);
402 static int vhost_net_enable_vq(struct vhost_net *n,
403 struct vhost_virtqueue *vq)
405 struct vhost_net_virtqueue *nvq =
406 container_of(vq, struct vhost_net_virtqueue, vq);
407 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
410 sock = vq->private_data;
414 return vhost_poll_start(poll, sock->file);
417 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
418 struct vhost_virtqueue *vq,
419 struct iovec iov[], unsigned int iov_size,
420 unsigned int *out_num, unsigned int *in_num)
422 unsigned long uninitialized_var(endtime);
423 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
424 out_num, in_num, NULL, NULL);
426 if (r == vq->num && vq->busyloop_timeout) {
428 endtime = busy_clock() + vq->busyloop_timeout;
429 while (vhost_can_busy_poll(vq->dev, endtime) &&
430 vhost_vq_avail_empty(vq->dev, vq))
433 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
434 out_num, in_num, NULL, NULL);
440 static bool vhost_exceeds_maxpend(struct vhost_net *net)
442 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
443 struct vhost_virtqueue *vq = &nvq->vq;
445 return (nvq->upend_idx + vq->num - VHOST_MAX_PEND) % UIO_MAXIOV
449 /* Expects to be always run from workqueue - which acts as
450 * read-size critical section for our kind of RCU. */
451 static void handle_tx(struct vhost_net *net)
453 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
454 struct vhost_virtqueue *vq = &nvq->vq;
457 struct msghdr msg = {
462 .msg_flags = MSG_DONTWAIT,
464 size_t len, total_len = 0;
468 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
469 struct ubuf_info *ubuf;
470 bool zcopy, zcopy_used;
473 mutex_lock(&vq->mutex);
474 sock = vq->private_data;
478 if (!vq_iotlb_prefetch(vq))
481 vhost_disable_notify(&net->dev, vq);
483 hdr_size = nvq->vhost_hlen;
487 /* Release DMAs done buffers first */
489 vhost_zerocopy_signal_used(net, vq);
491 /* If more outstanding DMAs, queue the work.
492 * Handle upend_idx wrap around
494 if (unlikely(vhost_exceeds_maxpend(net)))
497 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
500 /* On error, stop handling until the next kick. */
501 if (unlikely(head < 0))
503 /* Nothing new? Wait for eventfd to tell us they refilled. */
504 if (head == vq->num) {
505 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
506 vhost_disable_notify(&net->dev, vq);
512 vq_err(vq, "Unexpected descriptor format for TX: "
513 "out %d, int %d\n", out, in);
516 /* Skip header. TODO: support TSO. */
517 len = iov_length(vq->iov, out);
518 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
519 iov_iter_advance(&msg.msg_iter, hdr_size);
521 if (!msg_data_left(&msg)) {
522 vq_err(vq, "Unexpected header len for TX: "
523 "%zd expected %zd\n",
527 len = msg_data_left(&msg);
529 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
530 && (nvq->upend_idx + 1) % UIO_MAXIOV !=
532 && vhost_net_tx_select_zcopy(net);
534 /* use msg_control to pass vhost zerocopy ubuf info to skb */
536 ubuf = nvq->ubuf_info + nvq->upend_idx;
537 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
538 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
539 ubuf->callback = vhost_zerocopy_callback;
540 ubuf->ctx = nvq->ubufs;
541 ubuf->desc = nvq->upend_idx;
542 refcount_set(&ubuf->refcnt, 1);
543 msg.msg_control = ubuf;
544 msg.msg_controllen = sizeof(ubuf);
546 atomic_inc(&ubufs->refcount);
547 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
549 msg.msg_control = NULL;
553 if (total_len < VHOST_NET_WEIGHT &&
554 !vhost_vq_avail_empty(&net->dev, vq) &&
555 likely(!vhost_exceeds_maxpend(net))) {
556 msg.msg_flags |= MSG_MORE;
558 msg.msg_flags &= ~MSG_MORE;
561 /* TODO: Check specific error and bomb out unless ENOBUFS? */
562 err = sock->ops->sendmsg(sock, &msg, len);
563 if (unlikely(err < 0)) {
565 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
566 vhost_net_ubuf_put(ubufs);
567 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
570 vhost_discard_vq_desc(vq, 1);
574 pr_debug("Truncated TX packet: "
575 " len %d != %zd\n", err, len);
577 vhost_add_used_and_signal(&net->dev, vq, head, 0);
579 vhost_zerocopy_signal_used(net, vq);
580 vhost_net_tx_packet(net);
581 } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
583 mutex_unlock(&vq->mutex);
586 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
588 struct sk_buff *head;
593 return vhost_net_buf_peek(rvq);
595 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
596 head = skb_peek(&sk->sk_receive_queue);
599 if (skb_vlan_tag_present(head))
603 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
607 static int sk_has_rx_data(struct sock *sk)
609 struct socket *sock = sk->sk_socket;
611 if (sock->ops->peek_len)
612 return sock->ops->peek_len(sock);
614 return skb_queue_empty(&sk->sk_receive_queue);
617 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
619 struct vhost_net_virtqueue *rvq = &net->vqs[VHOST_NET_VQ_RX];
620 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
621 struct vhost_virtqueue *vq = &nvq->vq;
622 unsigned long uninitialized_var(endtime);
623 int len = peek_head_len(rvq, sk);
625 if (!len && vq->busyloop_timeout) {
626 /* Both tx vq and rx socket were polled here */
627 mutex_lock_nested(&vq->mutex, 1);
628 vhost_disable_notify(&net->dev, vq);
631 endtime = busy_clock() + vq->busyloop_timeout;
633 while (vhost_can_busy_poll(&net->dev, endtime) &&
634 !sk_has_rx_data(sk) &&
635 vhost_vq_avail_empty(&net->dev, vq))
640 if (!vhost_vq_avail_empty(&net->dev, vq))
641 vhost_poll_queue(&vq->poll);
642 else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
643 vhost_disable_notify(&net->dev, vq);
644 vhost_poll_queue(&vq->poll);
647 mutex_unlock(&vq->mutex);
649 len = peek_head_len(rvq, sk);
655 /* This is a multi-buffer version of vhost_get_desc, that works if
656 * vq has read descriptors only.
657 * @vq - the relevant virtqueue
658 * @datalen - data length we'll be reading
659 * @iovcount - returned count of io vectors we fill
661 * @log_num - log offset
662 * @quota - headcount quota, 1 for big buffer
663 * returns number of buffer heads allocated, negative on error
665 static int get_rx_bufs(struct vhost_virtqueue *vq,
666 struct vring_used_elem *heads,
669 struct vhost_log *log,
673 unsigned int out, in;
678 /* len is always initialized before use since we are always called with
683 while (datalen > 0 && headcount < quota) {
684 if (unlikely(seg >= UIO_MAXIOV)) {
688 r = vhost_get_vq_desc(vq, vq->iov + seg,
689 ARRAY_SIZE(vq->iov) - seg, &out,
699 if (unlikely(out || in <= 0)) {
700 vq_err(vq, "unexpected descriptor format for RX: "
701 "out %d, in %d\n", out, in);
709 heads[headcount].id = cpu_to_vhost32(vq, d);
710 len = iov_length(vq->iov + seg, in);
711 heads[headcount].len = cpu_to_vhost32(vq, len);
716 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
722 if (unlikely(datalen > 0)) {
728 vhost_discard_vq_desc(vq, headcount);
732 /* Expects to be always run from workqueue - which acts as
733 * read-size critical section for our kind of RCU. */
734 static void handle_rx(struct vhost_net *net)
736 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
737 struct vhost_virtqueue *vq = &nvq->vq;
739 struct vhost_log *vq_log;
740 struct msghdr msg = {
743 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
745 .msg_flags = MSG_DONTWAIT,
747 struct virtio_net_hdr hdr = {
749 .gso_type = VIRTIO_NET_HDR_GSO_NONE
751 size_t total_len = 0;
754 size_t vhost_hlen, sock_hlen;
755 size_t vhost_len, sock_len;
757 struct iov_iter fixup;
758 __virtio16 num_buffers;
761 mutex_lock_nested(&vq->mutex, 0);
762 sock = vq->private_data;
766 if (!vq_iotlb_prefetch(vq))
769 vhost_disable_notify(&net->dev, vq);
770 vhost_net_disable_vq(net, vq);
772 vhost_hlen = nvq->vhost_hlen;
773 sock_hlen = nvq->sock_hlen;
775 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
777 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
780 sock_len = vhost_net_rx_peek_head_len(net, sock->sk);
784 sock_len += sock_hlen;
785 vhost_len = sock_len + vhost_hlen;
786 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
788 likely(mergeable) ? UIO_MAXIOV : 1);
789 /* On error, stop handling until the next kick. */
790 if (unlikely(headcount < 0))
792 /* OK, now we need to know about added descriptors. */
794 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
795 /* They have slipped one in as we were
796 * doing that: check again. */
797 vhost_disable_notify(&net->dev, vq);
800 /* Nothing new? Wait for eventfd to tell us
805 msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
806 /* On overrun, truncate and discard */
807 if (unlikely(headcount > UIO_MAXIOV)) {
808 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
809 err = sock->ops->recvmsg(sock, &msg,
810 1, MSG_DONTWAIT | MSG_TRUNC);
811 pr_debug("Discarded rx packet: len %zd\n", sock_len);
814 /* We don't need to be notified again. */
815 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
816 fixup = msg.msg_iter;
817 if (unlikely((vhost_hlen))) {
818 /* We will supply the header ourselves
821 iov_iter_advance(&msg.msg_iter, vhost_hlen);
823 err = sock->ops->recvmsg(sock, &msg,
824 sock_len, MSG_DONTWAIT | MSG_TRUNC);
825 /* Userspace might have consumed the packet meanwhile:
826 * it's not supposed to do this usually, but might be hard
827 * to prevent. Discard data we got (if any) and keep going. */
828 if (unlikely(err != sock_len)) {
829 pr_debug("Discarded rx packet: "
830 " len %d, expected %zd\n", err, sock_len);
831 vhost_discard_vq_desc(vq, headcount);
834 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
835 if (unlikely(vhost_hlen)) {
836 if (copy_to_iter(&hdr, sizeof(hdr),
837 &fixup) != sizeof(hdr)) {
838 vq_err(vq, "Unable to write vnet_hdr "
839 "at addr %p\n", vq->iov->iov_base);
843 /* Header came from socket; we'll need to patch
844 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
846 iov_iter_advance(&fixup, sizeof(hdr));
848 /* TODO: Should check and handle checksum. */
850 num_buffers = cpu_to_vhost16(vq, headcount);
851 if (likely(mergeable) &&
852 copy_to_iter(&num_buffers, sizeof num_buffers,
853 &fixup) != sizeof num_buffers) {
854 vq_err(vq, "Failed num_buffers write");
855 vhost_discard_vq_desc(vq, headcount);
858 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
860 if (unlikely(vq_log))
861 vhost_log_write(vq, vq_log, log, vhost_len,
863 total_len += vhost_len;
864 } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
866 vhost_net_enable_vq(net, vq);
868 mutex_unlock(&vq->mutex);
871 static void handle_tx_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_rx_kick(struct vhost_work *work)
882 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
884 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
889 static void handle_tx_net(struct vhost_work *work)
891 struct vhost_net *net = container_of(work, struct vhost_net,
892 poll[VHOST_NET_VQ_TX].work);
896 static void handle_rx_net(struct vhost_work *work)
898 struct vhost_net *net = container_of(work, struct vhost_net,
899 poll[VHOST_NET_VQ_RX].work);
903 static int vhost_net_open(struct inode *inode, struct file *f)
906 struct vhost_dev *dev;
907 struct vhost_virtqueue **vqs;
908 struct sk_buff **queue;
911 n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
914 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
920 queue = kmalloc_array(VHOST_RX_BATCH, sizeof(struct sk_buff *),
927 n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
930 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
931 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
932 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
933 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
934 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
935 n->vqs[i].ubufs = NULL;
936 n->vqs[i].ubuf_info = NULL;
937 n->vqs[i].upend_idx = 0;
938 n->vqs[i].done_idx = 0;
939 n->vqs[i].vhost_hlen = 0;
940 n->vqs[i].sock_hlen = 0;
941 vhost_net_buf_init(&n->vqs[i].rxq);
943 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
944 VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT);
946 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
947 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
954 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
955 struct vhost_virtqueue *vq)
958 struct vhost_net_virtqueue *nvq =
959 container_of(vq, struct vhost_net_virtqueue, vq);
961 mutex_lock(&vq->mutex);
962 sock = vq->private_data;
963 vhost_net_disable_vq(n, vq);
964 vq->private_data = NULL;
965 vhost_net_buf_unproduce(nvq);
966 mutex_unlock(&vq->mutex);
970 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
971 struct socket **rx_sock)
973 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
974 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
977 static void vhost_net_flush_vq(struct vhost_net *n, int index)
979 vhost_poll_flush(n->poll + index);
980 vhost_poll_flush(&n->vqs[index].vq.poll);
983 static void vhost_net_flush(struct vhost_net *n)
985 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
986 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
987 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
988 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
990 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
991 /* Wait for all lower device DMAs done. */
992 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
993 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
995 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
996 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1000 static int vhost_net_release(struct inode *inode, struct file *f)
1002 struct vhost_net *n = f->private_data;
1003 struct socket *tx_sock;
1004 struct socket *rx_sock;
1006 vhost_net_stop(n, &tx_sock, &rx_sock);
1008 vhost_dev_stop(&n->dev);
1009 vhost_dev_cleanup(&n->dev, false);
1010 vhost_net_vq_reset(n);
1012 sockfd_put(tx_sock);
1014 sockfd_put(rx_sock);
1015 /* Make sure no callbacks are outstanding */
1016 synchronize_rcu_bh();
1017 /* We do an extra flush before freeing memory,
1018 * since jobs can re-queue themselves. */
1020 kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1026 static struct socket *get_raw_socket(int fd)
1029 struct socket *sock = sockfd_lookup(fd, &r);
1032 return ERR_PTR(-ENOTSOCK);
1034 /* Parameter checking */
1035 if (sock->sk->sk_type != SOCK_RAW) {
1036 r = -ESOCKTNOSUPPORT;
1040 if (sock->sk->sk_family != AF_PACKET) {
1050 static struct skb_array *get_tap_skb_array(struct file *file)
1052 struct skb_array *array;
1053 array = tun_get_skb_array(file);
1056 array = tap_get_skb_array(file);
1064 static struct socket *get_tap_socket(int fd)
1066 struct file *file = fget(fd);
1067 struct socket *sock;
1070 return ERR_PTR(-EBADF);
1071 sock = tun_get_socket(file);
1074 sock = tap_get_socket(file);
1080 static struct socket *get_socket(int fd)
1082 struct socket *sock;
1084 /* special case to disable backend */
1087 sock = get_raw_socket(fd);
1090 sock = get_tap_socket(fd);
1093 return ERR_PTR(-ENOTSOCK);
1096 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1098 struct socket *sock, *oldsock;
1099 struct vhost_virtqueue *vq;
1100 struct vhost_net_virtqueue *nvq;
1101 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1104 mutex_lock(&n->dev.mutex);
1105 r = vhost_dev_check_owner(&n->dev);
1109 if (index >= VHOST_NET_VQ_MAX) {
1113 vq = &n->vqs[index].vq;
1114 nvq = &n->vqs[index];
1115 mutex_lock(&vq->mutex);
1117 /* Verify that ring has been setup correctly. */
1118 if (!vhost_vq_access_ok(vq)) {
1122 sock = get_socket(fd);
1128 /* start polling new socket */
1129 oldsock = vq->private_data;
1130 if (sock != oldsock) {
1131 ubufs = vhost_net_ubuf_alloc(vq,
1132 sock && vhost_sock_zcopy(sock));
1133 if (IS_ERR(ubufs)) {
1138 vhost_net_disable_vq(n, vq);
1139 vq->private_data = sock;
1140 vhost_net_buf_unproduce(nvq);
1141 if (index == VHOST_NET_VQ_RX) {
1143 nvq->rx_array = get_tap_skb_array(sock->file);
1145 nvq->rx_array = NULL;
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);
1188 mutex_unlock(&vq->mutex);
1190 mutex_unlock(&n->dev.mutex);
1194 static long vhost_net_reset_owner(struct vhost_net *n)
1196 struct socket *tx_sock = NULL;
1197 struct socket *rx_sock = NULL;
1199 struct vhost_umem *umem;
1201 mutex_lock(&n->dev.mutex);
1202 err = vhost_dev_check_owner(&n->dev);
1205 umem = vhost_dev_reset_owner_prepare();
1210 vhost_net_stop(n, &tx_sock, &rx_sock);
1212 vhost_dev_stop(&n->dev);
1213 vhost_dev_reset_owner(&n->dev, umem);
1214 vhost_net_vq_reset(n);
1216 mutex_unlock(&n->dev.mutex);
1218 sockfd_put(tx_sock);
1220 sockfd_put(rx_sock);
1224 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1226 size_t vhost_hlen, sock_hlen, hdr_len;
1229 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1230 (1ULL << VIRTIO_F_VERSION_1))) ?
1231 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1232 sizeof(struct virtio_net_hdr);
1233 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1234 /* vhost provides vnet_hdr */
1235 vhost_hlen = hdr_len;
1238 /* socket provides vnet_hdr */
1240 sock_hlen = hdr_len;
1242 mutex_lock(&n->dev.mutex);
1243 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1244 !vhost_log_access_ok(&n->dev))
1247 if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1248 if (vhost_init_device_iotlb(&n->dev, true))
1252 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1253 mutex_lock(&n->vqs[i].vq.mutex);
1254 n->vqs[i].vq.acked_features = features;
1255 n->vqs[i].vhost_hlen = vhost_hlen;
1256 n->vqs[i].sock_hlen = sock_hlen;
1257 mutex_unlock(&n->vqs[i].vq.mutex);
1259 mutex_unlock(&n->dev.mutex);
1263 mutex_unlock(&n->dev.mutex);
1267 static long vhost_net_set_owner(struct vhost_net *n)
1271 mutex_lock(&n->dev.mutex);
1272 if (vhost_dev_has_owner(&n->dev)) {
1276 r = vhost_net_set_ubuf_info(n);
1279 r = vhost_dev_set_owner(&n->dev);
1281 vhost_net_clear_ubuf_info(n);
1284 mutex_unlock(&n->dev.mutex);
1288 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1291 struct vhost_net *n = f->private_data;
1292 void __user *argp = (void __user *)arg;
1293 u64 __user *featurep = argp;
1294 struct vhost_vring_file backend;
1299 case VHOST_NET_SET_BACKEND:
1300 if (copy_from_user(&backend, argp, sizeof backend))
1302 return vhost_net_set_backend(n, backend.index, backend.fd);
1303 case VHOST_GET_FEATURES:
1304 features = VHOST_NET_FEATURES;
1305 if (copy_to_user(featurep, &features, sizeof features))
1308 case VHOST_SET_FEATURES:
1309 if (copy_from_user(&features, featurep, sizeof features))
1311 if (features & ~VHOST_NET_FEATURES)
1313 return vhost_net_set_features(n, features);
1314 case VHOST_RESET_OWNER:
1315 return vhost_net_reset_owner(n);
1316 case VHOST_SET_OWNER:
1317 return vhost_net_set_owner(n);
1319 mutex_lock(&n->dev.mutex);
1320 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1321 if (r == -ENOIOCTLCMD)
1322 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1325 mutex_unlock(&n->dev.mutex);
1330 #ifdef CONFIG_COMPAT
1331 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1334 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1338 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1340 struct file *file = iocb->ki_filp;
1341 struct vhost_net *n = file->private_data;
1342 struct vhost_dev *dev = &n->dev;
1343 int noblock = file->f_flags & O_NONBLOCK;
1345 return vhost_chr_read_iter(dev, to, noblock);
1348 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1349 struct iov_iter *from)
1351 struct file *file = iocb->ki_filp;
1352 struct vhost_net *n = file->private_data;
1353 struct vhost_dev *dev = &n->dev;
1355 return vhost_chr_write_iter(dev, from);
1358 static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait)
1360 struct vhost_net *n = file->private_data;
1361 struct vhost_dev *dev = &n->dev;
1363 return vhost_chr_poll(file, dev, wait);
1366 static const struct file_operations vhost_net_fops = {
1367 .owner = THIS_MODULE,
1368 .release = vhost_net_release,
1369 .read_iter = vhost_net_chr_read_iter,
1370 .write_iter = vhost_net_chr_write_iter,
1371 .poll = vhost_net_chr_poll,
1372 .unlocked_ioctl = vhost_net_ioctl,
1373 #ifdef CONFIG_COMPAT
1374 .compat_ioctl = vhost_net_compat_ioctl,
1376 .open = vhost_net_open,
1377 .llseek = noop_llseek,
1380 static struct miscdevice vhost_net_misc = {
1381 .minor = VHOST_NET_MINOR,
1382 .name = "vhost-net",
1383 .fops = &vhost_net_fops,
1386 static int vhost_net_init(void)
1388 if (experimental_zcopytx)
1389 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1390 return misc_register(&vhost_net_misc);
1392 module_init(vhost_net_init);
1394 static void vhost_net_exit(void)
1396 misc_deregister(&vhost_net_misc);
1398 module_exit(vhost_net_exit);
1400 MODULE_VERSION("0.0.1");
1401 MODULE_LICENSE("GPL v2");
1402 MODULE_AUTHOR("Michael S. Tsirkin");
1403 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1404 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1405 MODULE_ALIAS("devname:vhost-net");