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 {
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 ptr_ring *rx_ring;
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 = ptr_ring_consume_batched(nvq->rx_ring, 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_ring && !vhost_net_buf_is_empty(rxq)) {
171 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
172 vhost_net_buf_get_size(rxq),
174 rxq->head = rxq->tail = 0;
178 static int vhost_net_buf_peek_len(void *ptr)
180 if (tun_is_xdp_buff(ptr)) {
181 struct xdp_buff *xdp = tun_ptr_to_xdp(ptr);
183 return xdp->data_end - xdp->data;
186 return __skb_array_len_with_tag(ptr);
189 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
191 struct vhost_net_buf *rxq = &nvq->rxq;
193 if (!vhost_net_buf_is_empty(rxq))
196 if (!vhost_net_buf_produce(nvq))
200 return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
203 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
205 rxq->head = rxq->tail = 0;
208 static void vhost_net_enable_zcopy(int vq)
210 vhost_net_zcopy_mask |= 0x1 << vq;
213 static struct vhost_net_ubuf_ref *
214 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
216 struct vhost_net_ubuf_ref *ubufs;
217 /* No zero copy backend? Nothing to count. */
220 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
222 return ERR_PTR(-ENOMEM);
223 atomic_set(&ubufs->refcount, 1);
224 init_waitqueue_head(&ubufs->wait);
229 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
231 int r = atomic_sub_return(1, &ubufs->refcount);
233 wake_up(&ubufs->wait);
237 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
239 vhost_net_ubuf_put(ubufs);
240 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
243 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
245 vhost_net_ubuf_put_and_wait(ubufs);
249 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
253 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
254 kfree(n->vqs[i].ubuf_info);
255 n->vqs[i].ubuf_info = NULL;
259 static int vhost_net_set_ubuf_info(struct vhost_net *n)
264 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
265 zcopy = vhost_net_zcopy_mask & (0x1 << i);
268 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
269 UIO_MAXIOV, GFP_KERNEL);
270 if (!n->vqs[i].ubuf_info)
276 vhost_net_clear_ubuf_info(n);
280 static void vhost_net_vq_reset(struct vhost_net *n)
284 vhost_net_clear_ubuf_info(n);
286 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
287 n->vqs[i].done_idx = 0;
288 n->vqs[i].upend_idx = 0;
289 n->vqs[i].ubufs = NULL;
290 n->vqs[i].vhost_hlen = 0;
291 n->vqs[i].sock_hlen = 0;
292 vhost_net_buf_init(&n->vqs[i].rxq);
297 static void vhost_net_tx_packet(struct vhost_net *net)
300 if (net->tx_packets < 1024)
303 net->tx_zcopy_err = 0;
306 static void vhost_net_tx_err(struct vhost_net *net)
311 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
313 /* TX flush waits for outstanding DMAs to be done.
314 * Don't start new DMAs.
316 return !net->tx_flush &&
317 net->tx_packets / 64 >= net->tx_zcopy_err;
320 static bool vhost_sock_zcopy(struct socket *sock)
322 return unlikely(experimental_zcopytx) &&
323 sock_flag(sock->sk, SOCK_ZEROCOPY);
326 /* In case of DMA done not in order in lower device driver for some reason.
327 * upend_idx is used to track end of used idx, done_idx is used to track head
328 * of used idx. Once lower device DMA done contiguously, we will signal KVM
331 static void vhost_zerocopy_signal_used(struct vhost_net *net,
332 struct vhost_virtqueue *vq)
334 struct vhost_net_virtqueue *nvq =
335 container_of(vq, struct vhost_net_virtqueue, vq);
339 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
340 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
341 vhost_net_tx_err(net);
342 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
343 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
349 add = min(UIO_MAXIOV - nvq->done_idx, j);
350 vhost_add_used_and_signal_n(vq->dev, vq,
351 &vq->heads[nvq->done_idx], add);
352 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
357 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
359 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
360 struct vhost_virtqueue *vq = ubufs->vq;
365 /* set len to mark this desc buffers done DMA */
366 vq->heads[ubuf->desc].len = success ?
367 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
368 cnt = vhost_net_ubuf_put(ubufs);
371 * Trigger polling thread if guest stopped submitting new buffers:
372 * in this case, the refcount after decrement will eventually reach 1.
373 * We also trigger polling periodically after each 16 packets
374 * (the value 16 here is more or less arbitrary, it's tuned to trigger
375 * less than 10% of times).
377 if (cnt <= 1 || !(cnt % 16))
378 vhost_poll_queue(&vq->poll);
380 rcu_read_unlock_bh();
383 static inline unsigned long busy_clock(void)
385 return local_clock() >> 10;
388 static bool vhost_can_busy_poll(struct vhost_dev *dev,
389 unsigned long endtime)
391 return likely(!need_resched()) &&
392 likely(!time_after(busy_clock(), endtime)) &&
393 likely(!signal_pending(current)) &&
394 !vhost_has_work(dev);
397 static void vhost_net_disable_vq(struct vhost_net *n,
398 struct vhost_virtqueue *vq)
400 struct vhost_net_virtqueue *nvq =
401 container_of(vq, struct vhost_net_virtqueue, vq);
402 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
403 if (!vq->private_data)
405 vhost_poll_stop(poll);
408 static int vhost_net_enable_vq(struct vhost_net *n,
409 struct vhost_virtqueue *vq)
411 struct vhost_net_virtqueue *nvq =
412 container_of(vq, struct vhost_net_virtqueue, vq);
413 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
416 sock = vq->private_data;
420 return vhost_poll_start(poll, sock->file);
423 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
424 struct vhost_virtqueue *vq,
425 struct iovec iov[], unsigned int iov_size,
426 unsigned int *out_num, unsigned int *in_num)
428 unsigned long uninitialized_var(endtime);
429 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
430 out_num, in_num, NULL, NULL);
432 if (r == vq->num && vq->busyloop_timeout) {
434 endtime = busy_clock() + vq->busyloop_timeout;
435 while (vhost_can_busy_poll(vq->dev, endtime) &&
436 vhost_vq_avail_empty(vq->dev, vq))
439 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
440 out_num, in_num, NULL, NULL);
446 static bool vhost_exceeds_maxpend(struct vhost_net *net)
448 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
449 struct vhost_virtqueue *vq = &nvq->vq;
451 return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
452 min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
455 /* Expects to be always run from workqueue - which acts as
456 * read-size critical section for our kind of RCU. */
457 static void handle_tx(struct vhost_net *net)
459 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
460 struct vhost_virtqueue *vq = &nvq->vq;
463 struct msghdr msg = {
468 .msg_flags = MSG_DONTWAIT,
470 size_t len, total_len = 0;
474 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
475 bool zcopy, zcopy_used;
477 mutex_lock(&vq->mutex);
478 sock = vq->private_data;
482 if (!vq_iotlb_prefetch(vq))
485 vhost_disable_notify(&net->dev, vq);
486 vhost_net_disable_vq(net, vq);
488 hdr_size = nvq->vhost_hlen;
492 /* Release DMAs done buffers first */
494 vhost_zerocopy_signal_used(net, vq);
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 && !vhost_exceeds_maxpend(net)
531 && vhost_net_tx_select_zcopy(net);
533 /* use msg_control to pass vhost zerocopy ubuf info to skb */
535 struct ubuf_info *ubuf;
536 ubuf = nvq->ubuf_info + nvq->upend_idx;
538 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
539 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
540 ubuf->callback = vhost_zerocopy_callback;
541 ubuf->ctx = nvq->ubufs;
542 ubuf->desc = nvq->upend_idx;
543 refcount_set(&ubuf->refcnt, 1);
544 msg.msg_control = ubuf;
545 msg.msg_controllen = sizeof(ubuf);
547 atomic_inc(&ubufs->refcount);
548 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
550 msg.msg_control = NULL;
555 if (total_len < VHOST_NET_WEIGHT &&
556 !vhost_vq_avail_empty(&net->dev, vq) &&
557 likely(!vhost_exceeds_maxpend(net))) {
558 msg.msg_flags |= MSG_MORE;
560 msg.msg_flags &= ~MSG_MORE;
563 /* TODO: Check specific error and bomb out unless ENOBUFS? */
564 err = sock->ops->sendmsg(sock, &msg, len);
565 if (unlikely(err < 0)) {
567 vhost_net_ubuf_put(ubufs);
568 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
571 vhost_discard_vq_desc(vq, 1);
572 vhost_net_enable_vq(net, vq);
576 pr_debug("Truncated TX packet: "
577 " len %d != %zd\n", err, len);
579 vhost_add_used_and_signal(&net->dev, vq, head, 0);
581 vhost_zerocopy_signal_used(net, vq);
582 vhost_net_tx_packet(net);
583 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
584 vhost_poll_queue(&vq->poll);
589 mutex_unlock(&vq->mutex);
592 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
594 struct sk_buff *head;
599 return vhost_net_buf_peek(rvq);
601 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
602 head = skb_peek(&sk->sk_receive_queue);
605 if (skb_vlan_tag_present(head))
609 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
613 static int sk_has_rx_data(struct sock *sk)
615 struct socket *sock = sk->sk_socket;
617 if (sock->ops->peek_len)
618 return sock->ops->peek_len(sock);
620 return skb_queue_empty(&sk->sk_receive_queue);
623 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
625 struct vhost_net_virtqueue *rvq = &net->vqs[VHOST_NET_VQ_RX];
626 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
627 struct vhost_virtqueue *vq = &nvq->vq;
628 unsigned long uninitialized_var(endtime);
629 int len = peek_head_len(rvq, sk);
631 if (!len && vq->busyloop_timeout) {
632 /* Both tx vq and rx socket were polled here */
633 mutex_lock(&vq->mutex);
634 vhost_disable_notify(&net->dev, vq);
637 endtime = busy_clock() + vq->busyloop_timeout;
639 while (vhost_can_busy_poll(&net->dev, endtime) &&
640 !sk_has_rx_data(sk) &&
641 vhost_vq_avail_empty(&net->dev, vq))
646 if (!vhost_vq_avail_empty(&net->dev, vq))
647 vhost_poll_queue(&vq->poll);
648 else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
649 vhost_disable_notify(&net->dev, vq);
650 vhost_poll_queue(&vq->poll);
653 mutex_unlock(&vq->mutex);
655 len = peek_head_len(rvq, sk);
661 /* This is a multi-buffer version of vhost_get_desc, that works if
662 * vq has read descriptors only.
663 * @vq - the relevant virtqueue
664 * @datalen - data length we'll be reading
665 * @iovcount - returned count of io vectors we fill
667 * @log_num - log offset
668 * @quota - headcount quota, 1 for big buffer
669 * returns number of buffer heads allocated, negative on error
671 static int get_rx_bufs(struct vhost_virtqueue *vq,
672 struct vring_used_elem *heads,
675 struct vhost_log *log,
679 unsigned int out, in;
684 /* len is always initialized before use since we are always called with
687 u32 uninitialized_var(len);
689 while (datalen > 0 && headcount < quota) {
690 if (unlikely(seg >= UIO_MAXIOV)) {
694 r = vhost_get_vq_desc(vq, vq->iov + seg,
695 ARRAY_SIZE(vq->iov) - seg, &out,
705 if (unlikely(out || in <= 0)) {
706 vq_err(vq, "unexpected descriptor format for RX: "
707 "out %d, in %d\n", out, in);
715 heads[headcount].id = cpu_to_vhost32(vq, d);
716 len = iov_length(vq->iov + seg, in);
717 heads[headcount].len = cpu_to_vhost32(vq, len);
722 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
728 if (unlikely(datalen > 0)) {
734 vhost_discard_vq_desc(vq, headcount);
738 /* Expects to be always run from workqueue - which acts as
739 * read-size critical section for our kind of RCU. */
740 static void handle_rx(struct vhost_net *net)
742 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
743 struct vhost_virtqueue *vq = &nvq->vq;
744 unsigned uninitialized_var(in), log;
745 struct vhost_log *vq_log;
746 struct msghdr msg = {
749 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
751 .msg_flags = MSG_DONTWAIT,
753 struct virtio_net_hdr hdr = {
755 .gso_type = VIRTIO_NET_HDR_GSO_NONE
757 size_t total_len = 0;
759 s16 headcount, nheads = 0;
760 size_t vhost_hlen, sock_hlen;
761 size_t vhost_len, sock_len;
763 struct iov_iter fixup;
764 __virtio16 num_buffers;
766 mutex_lock(&vq->mutex);
767 sock = vq->private_data;
771 if (!vq_iotlb_prefetch(vq))
774 vhost_disable_notify(&net->dev, vq);
775 vhost_net_disable_vq(net, vq);
777 vhost_hlen = nvq->vhost_hlen;
778 sock_hlen = nvq->sock_hlen;
780 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
782 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
784 while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
785 sock_len += sock_hlen;
786 vhost_len = sock_len + vhost_hlen;
787 headcount = get_rx_bufs(vq, vq->heads + nheads, vhost_len,
789 likely(mergeable) ? UIO_MAXIOV : 1);
790 /* On error, stop handling until the next kick. */
791 if (unlikely(headcount < 0))
793 /* OK, now we need to know about added descriptors. */
795 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
796 /* They have slipped one in as we were
797 * doing that: check again. */
798 vhost_disable_notify(&net->dev, vq);
801 /* Nothing new? Wait for eventfd to tell us
806 msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
807 /* On overrun, truncate and discard */
808 if (unlikely(headcount > UIO_MAXIOV)) {
809 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
810 err = sock->ops->recvmsg(sock, &msg,
811 1, MSG_DONTWAIT | MSG_TRUNC);
812 pr_debug("Discarded rx packet: len %zd\n", sock_len);
815 /* We don't need to be notified again. */
816 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
817 fixup = msg.msg_iter;
818 if (unlikely((vhost_hlen))) {
819 /* We will supply the header ourselves
822 iov_iter_advance(&msg.msg_iter, vhost_hlen);
824 err = sock->ops->recvmsg(sock, &msg,
825 sock_len, MSG_DONTWAIT | MSG_TRUNC);
826 /* Userspace might have consumed the packet meanwhile:
827 * it's not supposed to do this usually, but might be hard
828 * to prevent. Discard data we got (if any) and keep going. */
829 if (unlikely(err != sock_len)) {
830 pr_debug("Discarded rx packet: "
831 " len %d, expected %zd\n", err, sock_len);
832 vhost_discard_vq_desc(vq, headcount);
835 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
836 if (unlikely(vhost_hlen)) {
837 if (copy_to_iter(&hdr, sizeof(hdr),
838 &fixup) != sizeof(hdr)) {
839 vq_err(vq, "Unable to write vnet_hdr "
840 "at addr %p\n", vq->iov->iov_base);
844 /* Header came from socket; we'll need to patch
845 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
847 iov_iter_advance(&fixup, sizeof(hdr));
849 /* TODO: Should check and handle checksum. */
851 num_buffers = cpu_to_vhost16(vq, headcount);
852 if (likely(mergeable) &&
853 copy_to_iter(&num_buffers, sizeof num_buffers,
854 &fixup) != sizeof num_buffers) {
855 vq_err(vq, "Failed num_buffers write");
856 vhost_discard_vq_desc(vq, headcount);
860 if (nheads > VHOST_RX_BATCH) {
861 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
865 if (unlikely(vq_log))
866 vhost_log_write(vq, vq_log, log, vhost_len);
867 total_len += vhost_len;
868 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
869 vhost_poll_queue(&vq->poll);
873 vhost_net_enable_vq(net, vq);
876 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
878 mutex_unlock(&vq->mutex);
881 static void handle_tx_kick(struct vhost_work *work)
883 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
885 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
890 static void handle_rx_kick(struct vhost_work *work)
892 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
894 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
899 static void handle_tx_net(struct vhost_work *work)
901 struct vhost_net *net = container_of(work, struct vhost_net,
902 poll[VHOST_NET_VQ_TX].work);
906 static void handle_rx_net(struct vhost_work *work)
908 struct vhost_net *net = container_of(work, struct vhost_net,
909 poll[VHOST_NET_VQ_RX].work);
913 static int vhost_net_open(struct inode *inode, struct file *f)
916 struct vhost_dev *dev;
917 struct vhost_virtqueue **vqs;
921 n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
924 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
930 queue = kmalloc_array(VHOST_RX_BATCH, sizeof(void *),
937 n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
940 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
941 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
942 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
943 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
944 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
945 n->vqs[i].ubufs = NULL;
946 n->vqs[i].ubuf_info = NULL;
947 n->vqs[i].upend_idx = 0;
948 n->vqs[i].done_idx = 0;
949 n->vqs[i].vhost_hlen = 0;
950 n->vqs[i].sock_hlen = 0;
951 n->vqs[i].rx_ring = NULL;
952 vhost_net_buf_init(&n->vqs[i].rxq);
954 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
956 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
957 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
964 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
965 struct vhost_virtqueue *vq)
968 struct vhost_net_virtqueue *nvq =
969 container_of(vq, struct vhost_net_virtqueue, vq);
971 mutex_lock(&vq->mutex);
972 sock = vq->private_data;
973 vhost_net_disable_vq(n, vq);
974 vq->private_data = NULL;
975 vhost_net_buf_unproduce(nvq);
977 mutex_unlock(&vq->mutex);
981 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
982 struct socket **rx_sock)
984 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
985 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
988 static void vhost_net_flush_vq(struct vhost_net *n, int index)
990 vhost_poll_flush(n->poll + index);
991 vhost_poll_flush(&n->vqs[index].vq.poll);
994 static void vhost_net_flush(struct vhost_net *n)
996 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
997 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
998 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
999 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1001 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1002 /* Wait for all lower device DMAs done. */
1003 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1004 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1005 n->tx_flush = false;
1006 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1007 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1011 static int vhost_net_release(struct inode *inode, struct file *f)
1013 struct vhost_net *n = f->private_data;
1014 struct socket *tx_sock;
1015 struct socket *rx_sock;
1017 vhost_net_stop(n, &tx_sock, &rx_sock);
1019 vhost_dev_stop(&n->dev);
1020 vhost_dev_cleanup(&n->dev);
1021 vhost_net_vq_reset(n);
1023 sockfd_put(tx_sock);
1025 sockfd_put(rx_sock);
1026 /* Make sure no callbacks are outstanding */
1027 synchronize_rcu_bh();
1028 /* We do an extra flush before freeing memory,
1029 * since jobs can re-queue themselves. */
1031 kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1037 static struct socket *get_raw_socket(int fd)
1040 struct sockaddr_ll sa;
1041 char buf[MAX_ADDR_LEN];
1043 int uaddr_len = sizeof uaddr, r;
1044 struct socket *sock = sockfd_lookup(fd, &r);
1047 return ERR_PTR(-ENOTSOCK);
1049 /* Parameter checking */
1050 if (sock->sk->sk_type != SOCK_RAW) {
1051 r = -ESOCKTNOSUPPORT;
1055 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
1060 if (uaddr.sa.sll_family != AF_PACKET) {
1070 static struct ptr_ring *get_tap_ptr_ring(int fd)
1072 struct ptr_ring *ring;
1073 struct file *file = fget(fd);
1077 ring = tun_get_tx_ring(file);
1080 ring = tap_get_ptr_ring(file);
1089 static struct socket *get_tap_socket(int fd)
1091 struct file *file = fget(fd);
1092 struct socket *sock;
1095 return ERR_PTR(-EBADF);
1096 sock = tun_get_socket(file);
1099 sock = tap_get_socket(file);
1105 static struct socket *get_socket(int fd)
1107 struct socket *sock;
1109 /* special case to disable backend */
1112 sock = get_raw_socket(fd);
1115 sock = get_tap_socket(fd);
1118 return ERR_PTR(-ENOTSOCK);
1121 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1123 struct socket *sock, *oldsock;
1124 struct vhost_virtqueue *vq;
1125 struct vhost_net_virtqueue *nvq;
1126 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1129 mutex_lock(&n->dev.mutex);
1130 r = vhost_dev_check_owner(&n->dev);
1134 if (index >= VHOST_NET_VQ_MAX) {
1138 vq = &n->vqs[index].vq;
1139 nvq = &n->vqs[index];
1140 mutex_lock(&vq->mutex);
1142 /* Verify that ring has been setup correctly. */
1143 if (!vhost_vq_access_ok(vq)) {
1147 sock = get_socket(fd);
1153 /* start polling new socket */
1154 oldsock = vq->private_data;
1155 if (sock != oldsock) {
1156 ubufs = vhost_net_ubuf_alloc(vq,
1157 sock && vhost_sock_zcopy(sock));
1158 if (IS_ERR(ubufs)) {
1163 vhost_net_disable_vq(n, vq);
1164 vq->private_data = sock;
1165 vhost_net_buf_unproduce(nvq);
1166 r = vhost_vq_init_access(vq);
1169 r = vhost_net_enable_vq(n, vq);
1172 if (index == VHOST_NET_VQ_RX)
1173 nvq->rx_ring = get_tap_ptr_ring(fd);
1175 oldubufs = nvq->ubufs;
1179 n->tx_zcopy_err = 0;
1180 n->tx_flush = false;
1183 mutex_unlock(&vq->mutex);
1186 vhost_net_ubuf_put_wait_and_free(oldubufs);
1187 mutex_lock(&vq->mutex);
1188 vhost_zerocopy_signal_used(n, vq);
1189 mutex_unlock(&vq->mutex);
1193 vhost_net_flush_vq(n, index);
1194 sockfd_put(oldsock);
1197 mutex_unlock(&n->dev.mutex);
1201 vq->private_data = oldsock;
1202 vhost_net_enable_vq(n, vq);
1204 vhost_net_ubuf_put_wait_and_free(ubufs);
1208 mutex_unlock(&vq->mutex);
1210 mutex_unlock(&n->dev.mutex);
1214 static long vhost_net_reset_owner(struct vhost_net *n)
1216 struct socket *tx_sock = NULL;
1217 struct socket *rx_sock = NULL;
1219 struct vhost_umem *umem;
1221 mutex_lock(&n->dev.mutex);
1222 err = vhost_dev_check_owner(&n->dev);
1225 umem = vhost_dev_reset_owner_prepare();
1230 vhost_net_stop(n, &tx_sock, &rx_sock);
1232 vhost_dev_stop(&n->dev);
1233 vhost_dev_reset_owner(&n->dev, umem);
1234 vhost_net_vq_reset(n);
1236 mutex_unlock(&n->dev.mutex);
1238 sockfd_put(tx_sock);
1240 sockfd_put(rx_sock);
1244 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1246 size_t vhost_hlen, sock_hlen, hdr_len;
1249 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1250 (1ULL << VIRTIO_F_VERSION_1))) ?
1251 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1252 sizeof(struct virtio_net_hdr);
1253 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1254 /* vhost provides vnet_hdr */
1255 vhost_hlen = hdr_len;
1258 /* socket provides vnet_hdr */
1260 sock_hlen = hdr_len;
1262 mutex_lock(&n->dev.mutex);
1263 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1264 !vhost_log_access_ok(&n->dev))
1267 if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1268 if (vhost_init_device_iotlb(&n->dev, true))
1272 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1273 mutex_lock(&n->vqs[i].vq.mutex);
1274 n->vqs[i].vq.acked_features = features;
1275 n->vqs[i].vhost_hlen = vhost_hlen;
1276 n->vqs[i].sock_hlen = sock_hlen;
1277 mutex_unlock(&n->vqs[i].vq.mutex);
1279 mutex_unlock(&n->dev.mutex);
1283 mutex_unlock(&n->dev.mutex);
1287 static long vhost_net_set_owner(struct vhost_net *n)
1291 mutex_lock(&n->dev.mutex);
1292 if (vhost_dev_has_owner(&n->dev)) {
1296 r = vhost_net_set_ubuf_info(n);
1299 r = vhost_dev_set_owner(&n->dev);
1301 vhost_net_clear_ubuf_info(n);
1304 mutex_unlock(&n->dev.mutex);
1308 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1311 struct vhost_net *n = f->private_data;
1312 void __user *argp = (void __user *)arg;
1313 u64 __user *featurep = argp;
1314 struct vhost_vring_file backend;
1319 case VHOST_NET_SET_BACKEND:
1320 if (copy_from_user(&backend, argp, sizeof backend))
1322 return vhost_net_set_backend(n, backend.index, backend.fd);
1323 case VHOST_GET_FEATURES:
1324 features = VHOST_NET_FEATURES;
1325 if (copy_to_user(featurep, &features, sizeof features))
1328 case VHOST_SET_FEATURES:
1329 if (copy_from_user(&features, featurep, sizeof features))
1331 if (features & ~VHOST_NET_FEATURES)
1333 return vhost_net_set_features(n, features);
1334 case VHOST_RESET_OWNER:
1335 return vhost_net_reset_owner(n);
1336 case VHOST_SET_OWNER:
1337 return vhost_net_set_owner(n);
1339 mutex_lock(&n->dev.mutex);
1340 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1341 if (r == -ENOIOCTLCMD)
1342 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1345 mutex_unlock(&n->dev.mutex);
1350 #ifdef CONFIG_COMPAT
1351 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1354 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1358 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1360 struct file *file = iocb->ki_filp;
1361 struct vhost_net *n = file->private_data;
1362 struct vhost_dev *dev = &n->dev;
1363 int noblock = file->f_flags & O_NONBLOCK;
1365 return vhost_chr_read_iter(dev, to, noblock);
1368 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1369 struct iov_iter *from)
1371 struct file *file = iocb->ki_filp;
1372 struct vhost_net *n = file->private_data;
1373 struct vhost_dev *dev = &n->dev;
1375 return vhost_chr_write_iter(dev, from);
1378 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1380 struct vhost_net *n = file->private_data;
1381 struct vhost_dev *dev = &n->dev;
1383 return vhost_chr_poll(file, dev, wait);
1386 static const struct file_operations vhost_net_fops = {
1387 .owner = THIS_MODULE,
1388 .release = vhost_net_release,
1389 .read_iter = vhost_net_chr_read_iter,
1390 .write_iter = vhost_net_chr_write_iter,
1391 .poll = vhost_net_chr_poll,
1392 .unlocked_ioctl = vhost_net_ioctl,
1393 #ifdef CONFIG_COMPAT
1394 .compat_ioctl = vhost_net_compat_ioctl,
1396 .open = vhost_net_open,
1397 .llseek = noop_llseek,
1400 static struct miscdevice vhost_net_misc = {
1401 .minor = VHOST_NET_MINOR,
1402 .name = "vhost-net",
1403 .fops = &vhost_net_fops,
1406 static int vhost_net_init(void)
1408 if (experimental_zcopytx)
1409 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1410 return misc_register(&vhost_net_misc);
1412 module_init(vhost_net_init);
1414 static void vhost_net_exit(void)
1416 misc_deregister(&vhost_net_misc);
1418 module_exit(vhost_net_exit);
1420 MODULE_VERSION("0.0.1");
1421 MODULE_LICENSE("GPL v2");
1422 MODULE_AUTHOR("Michael S. Tsirkin");
1423 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1424 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1425 MODULE_ALIAS("devname:vhost-net");