1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Copyright (C) 2009 Red Hat, Inc.
3 * Author: Michael S. Tsirkin <mst@redhat.com>
5 * virtio-net server in host kernel.
8 #include <linux/compat.h>
9 #include <linux/eventfd.h>
10 #include <linux/vhost.h>
11 #include <linux/virtio_net.h>
12 #include <linux/miscdevice.h>
13 #include <linux/module.h>
14 #include <linux/moduleparam.h>
15 #include <linux/mutex.h>
16 #include <linux/workqueue.h>
17 #include <linux/file.h>
18 #include <linux/slab.h>
19 #include <linux/sched/clock.h>
20 #include <linux/sched/signal.h>
21 #include <linux/vmalloc.h>
23 #include <linux/net.h>
24 #include <linux/if_packet.h>
25 #include <linux/if_arp.h>
26 #include <linux/if_tun.h>
27 #include <linux/if_macvlan.h>
28 #include <linux/if_tap.h>
29 #include <linux/if_vlan.h>
30 #include <linux/skb_array.h>
31 #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_ACCESS_PLATFORM)
80 VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
89 struct vhost_net_ubuf_ref {
90 /* refcount follows semantics similar to kref:
91 * 0: object is released
92 * 1: no outstanding ubufs
93 * >1: outstanding ubufs
96 wait_queue_head_t wait;
97 struct vhost_virtqueue *vq;
100 #define VHOST_NET_BATCH 64
101 struct vhost_net_buf {
107 struct vhost_net_virtqueue {
108 struct vhost_virtqueue vq;
111 /* vhost zerocopy support fields below: */
112 /* last used idx for outstanding DMA zerocopy buffers */
114 /* For TX, first used idx for DMA done zerocopy buffers
115 * For RX, number of batched heads
118 /* Number of XDP frames batched */
120 /* an array of userspace buffers info */
121 struct ubuf_info *ubuf_info;
122 /* Reference counting for outstanding ubufs.
123 * Protected by vq mutex. Writers must also take device mutex. */
124 struct vhost_net_ubuf_ref *ubufs;
125 struct ptr_ring *rx_ring;
126 struct vhost_net_buf rxq;
127 /* Batched XDP buffs */
128 struct xdp_buff *xdp;
132 struct vhost_dev dev;
133 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
134 struct vhost_poll poll[VHOST_NET_VQ_MAX];
135 /* Number of TX recently submitted.
136 * Protected by tx vq lock. */
138 /* Number of times zerocopy TX recently failed.
139 * Protected by tx vq lock. */
140 unsigned tx_zcopy_err;
141 /* Flush in progress. Protected by tx vq lock. */
143 /* Private page frag */
144 struct page_frag page_frag;
145 /* Refcount bias of page frag */
149 static unsigned vhost_net_zcopy_mask __read_mostly;
151 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
153 if (rxq->tail != rxq->head)
154 return rxq->queue[rxq->head];
159 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
161 return rxq->tail - rxq->head;
164 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
166 return rxq->tail == rxq->head;
169 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
171 void *ret = vhost_net_buf_get_ptr(rxq);
176 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
178 struct vhost_net_buf *rxq = &nvq->rxq;
181 rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
186 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
188 struct vhost_net_buf *rxq = &nvq->rxq;
190 if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
191 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
192 vhost_net_buf_get_size(rxq),
194 rxq->head = rxq->tail = 0;
198 static int vhost_net_buf_peek_len(void *ptr)
200 if (tun_is_xdp_frame(ptr)) {
201 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
206 return __skb_array_len_with_tag(ptr);
209 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
211 struct vhost_net_buf *rxq = &nvq->rxq;
213 if (!vhost_net_buf_is_empty(rxq))
216 if (!vhost_net_buf_produce(nvq))
220 return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
223 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
225 rxq->head = rxq->tail = 0;
228 static void vhost_net_enable_zcopy(int vq)
230 vhost_net_zcopy_mask |= 0x1 << vq;
233 static struct vhost_net_ubuf_ref *
234 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
236 struct vhost_net_ubuf_ref *ubufs;
237 /* No zero copy backend? Nothing to count. */
240 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
242 return ERR_PTR(-ENOMEM);
243 atomic_set(&ubufs->refcount, 1);
244 init_waitqueue_head(&ubufs->wait);
249 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
251 int r = atomic_sub_return(1, &ubufs->refcount);
253 wake_up(&ubufs->wait);
257 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
259 vhost_net_ubuf_put(ubufs);
260 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
263 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
265 vhost_net_ubuf_put_and_wait(ubufs);
269 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
273 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
274 kfree(n->vqs[i].ubuf_info);
275 n->vqs[i].ubuf_info = NULL;
279 static int vhost_net_set_ubuf_info(struct vhost_net *n)
284 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
285 zcopy = vhost_net_zcopy_mask & (0x1 << i);
288 n->vqs[i].ubuf_info =
289 kmalloc_array(UIO_MAXIOV,
290 sizeof(*n->vqs[i].ubuf_info),
292 if (!n->vqs[i].ubuf_info)
298 vhost_net_clear_ubuf_info(n);
302 static void vhost_net_vq_reset(struct vhost_net *n)
306 vhost_net_clear_ubuf_info(n);
308 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
309 n->vqs[i].done_idx = 0;
310 n->vqs[i].upend_idx = 0;
311 n->vqs[i].ubufs = NULL;
312 n->vqs[i].vhost_hlen = 0;
313 n->vqs[i].sock_hlen = 0;
314 vhost_net_buf_init(&n->vqs[i].rxq);
319 static void vhost_net_tx_packet(struct vhost_net *net)
322 if (net->tx_packets < 1024)
325 net->tx_zcopy_err = 0;
328 static void vhost_net_tx_err(struct vhost_net *net)
333 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
335 /* TX flush waits for outstanding DMAs to be done.
336 * Don't start new DMAs.
338 return !net->tx_flush &&
339 net->tx_packets / 64 >= net->tx_zcopy_err;
342 static bool vhost_sock_zcopy(struct socket *sock)
344 return unlikely(experimental_zcopytx) &&
345 sock_flag(sock->sk, SOCK_ZEROCOPY);
348 static bool vhost_sock_xdp(struct socket *sock)
350 return sock_flag(sock->sk, SOCK_XDP);
353 /* In case of DMA done not in order in lower device driver for some reason.
354 * upend_idx is used to track end of used idx, done_idx is used to track head
355 * of used idx. Once lower device DMA done contiguously, we will signal KVM
358 static void vhost_zerocopy_signal_used(struct vhost_net *net,
359 struct vhost_virtqueue *vq)
361 struct vhost_net_virtqueue *nvq =
362 container_of(vq, struct vhost_net_virtqueue, vq);
366 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
367 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
368 vhost_net_tx_err(net);
369 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
370 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
376 add = min(UIO_MAXIOV - nvq->done_idx, j);
377 vhost_add_used_and_signal_n(vq->dev, vq,
378 &vq->heads[nvq->done_idx], add);
379 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
384 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
386 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
387 struct vhost_virtqueue *vq = ubufs->vq;
392 /* set len to mark this desc buffers done DMA */
393 vq->heads[ubuf->desc].len = success ?
394 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
395 cnt = vhost_net_ubuf_put(ubufs);
398 * Trigger polling thread if guest stopped submitting new buffers:
399 * in this case, the refcount after decrement will eventually reach 1.
400 * We also trigger polling periodically after each 16 packets
401 * (the value 16 here is more or less arbitrary, it's tuned to trigger
402 * less than 10% of times).
404 if (cnt <= 1 || !(cnt % 16))
405 vhost_poll_queue(&vq->poll);
407 rcu_read_unlock_bh();
410 static inline unsigned long busy_clock(void)
412 return local_clock() >> 10;
415 static bool vhost_can_busy_poll(unsigned long endtime)
417 return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
418 !signal_pending(current));
421 static void vhost_net_disable_vq(struct vhost_net *n,
422 struct vhost_virtqueue *vq)
424 struct vhost_net_virtqueue *nvq =
425 container_of(vq, struct vhost_net_virtqueue, vq);
426 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
427 if (!vhost_vq_get_backend(vq))
429 vhost_poll_stop(poll);
432 static int vhost_net_enable_vq(struct vhost_net *n,
433 struct vhost_virtqueue *vq)
435 struct vhost_net_virtqueue *nvq =
436 container_of(vq, struct vhost_net_virtqueue, vq);
437 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
440 sock = vhost_vq_get_backend(vq);
444 return vhost_poll_start(poll, sock->file);
447 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
449 struct vhost_virtqueue *vq = &nvq->vq;
450 struct vhost_dev *dev = vq->dev;
455 vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
459 static void vhost_tx_batch(struct vhost_net *net,
460 struct vhost_net_virtqueue *nvq,
462 struct msghdr *msghdr)
464 struct tun_msg_ctl ctl = {
466 .num = nvq->batched_xdp,
471 if (nvq->batched_xdp == 0)
474 msghdr->msg_control = &ctl;
475 msghdr->msg_controllen = sizeof(ctl);
476 err = sock->ops->sendmsg(sock, msghdr, 0);
477 if (unlikely(err < 0)) {
478 vq_err(&nvq->vq, "Fail to batch sending packets\n");
480 /* free pages owned by XDP; since this is an unlikely error path,
481 * keep it simple and avoid more complex bulk update for the
484 for (i = 0; i < nvq->batched_xdp; ++i)
485 put_page(virt_to_head_page(nvq->xdp[i].data));
486 nvq->batched_xdp = 0;
492 vhost_net_signal_used(nvq);
493 nvq->batched_xdp = 0;
496 static int sock_has_rx_data(struct socket *sock)
501 if (sock->ops->peek_len)
502 return sock->ops->peek_len(sock);
504 return skb_queue_empty(&sock->sk->sk_receive_queue);
507 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
508 struct vhost_virtqueue *vq)
510 if (!vhost_vq_avail_empty(&net->dev, vq)) {
511 vhost_poll_queue(&vq->poll);
512 } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
513 vhost_disable_notify(&net->dev, vq);
514 vhost_poll_queue(&vq->poll);
518 static void vhost_net_busy_poll(struct vhost_net *net,
519 struct vhost_virtqueue *rvq,
520 struct vhost_virtqueue *tvq,
524 unsigned long busyloop_timeout;
525 unsigned long endtime;
527 struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
529 /* Try to hold the vq mutex of the paired virtqueue. We can't
530 * use mutex_lock() here since we could not guarantee a
531 * consistenet lock ordering.
533 if (!mutex_trylock(&vq->mutex))
536 vhost_disable_notify(&net->dev, vq);
537 sock = vhost_vq_get_backend(rvq);
539 busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
540 tvq->busyloop_timeout;
543 endtime = busy_clock() + busyloop_timeout;
545 while (vhost_can_busy_poll(endtime)) {
546 if (vhost_has_work(&net->dev)) {
547 *busyloop_intr = true;
551 if ((sock_has_rx_data(sock) &&
552 !vhost_vq_avail_empty(&net->dev, rvq)) ||
553 !vhost_vq_avail_empty(&net->dev, tvq))
561 if (poll_rx || sock_has_rx_data(sock))
562 vhost_net_busy_poll_try_queue(net, vq);
563 else if (!poll_rx) /* On tx here, sock has no rx data. */
564 vhost_enable_notify(&net->dev, rvq);
566 mutex_unlock(&vq->mutex);
569 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
570 struct vhost_net_virtqueue *tnvq,
571 unsigned int *out_num, unsigned int *in_num,
572 struct msghdr *msghdr, bool *busyloop_intr)
574 struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
575 struct vhost_virtqueue *rvq = &rnvq->vq;
576 struct vhost_virtqueue *tvq = &tnvq->vq;
578 int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
579 out_num, in_num, NULL, NULL);
581 if (r == tvq->num && tvq->busyloop_timeout) {
582 /* Flush batched packets first */
583 if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
584 vhost_tx_batch(net, tnvq,
585 vhost_vq_get_backend(tvq),
588 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
590 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
591 out_num, in_num, NULL, NULL);
597 static bool vhost_exceeds_maxpend(struct vhost_net *net)
599 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
600 struct vhost_virtqueue *vq = &nvq->vq;
602 return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
603 min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
606 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
607 size_t hdr_size, int out)
609 /* Skip header. TODO: support TSO. */
610 size_t len = iov_length(vq->iov, out);
612 iov_iter_init(iter, WRITE, vq->iov, out, len);
613 iov_iter_advance(iter, hdr_size);
615 return iov_iter_count(iter);
618 static int get_tx_bufs(struct vhost_net *net,
619 struct vhost_net_virtqueue *nvq,
621 unsigned int *out, unsigned int *in,
622 size_t *len, bool *busyloop_intr)
624 struct vhost_virtqueue *vq = &nvq->vq;
627 ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
629 if (ret < 0 || ret == vq->num)
633 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
639 *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
641 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
642 *len, nvq->vhost_hlen);
649 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
651 return total_len < VHOST_NET_WEIGHT &&
652 !vhost_vq_avail_empty(vq->dev, vq);
655 #define SKB_FRAG_PAGE_ORDER get_order(32768)
657 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
658 struct page_frag *pfrag, gfp_t gfp)
661 if (pfrag->offset + sz <= pfrag->size)
663 __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
667 net->refcnt_bias = 0;
668 if (SKB_FRAG_PAGE_ORDER) {
669 /* Avoid direct reclaim but allow kswapd to wake */
670 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
671 __GFP_COMP | __GFP_NOWARN |
673 SKB_FRAG_PAGE_ORDER);
674 if (likely(pfrag->page)) {
675 pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
679 pfrag->page = alloc_page(gfp);
680 if (likely(pfrag->page)) {
681 pfrag->size = PAGE_SIZE;
687 net->refcnt_bias = USHRT_MAX;
688 page_ref_add(pfrag->page, USHRT_MAX - 1);
692 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
694 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
695 struct iov_iter *from)
697 struct vhost_virtqueue *vq = &nvq->vq;
698 struct vhost_net *net = container_of(vq->dev, struct vhost_net,
700 struct socket *sock = vhost_vq_get_backend(vq);
701 struct page_frag *alloc_frag = &net->page_frag;
702 struct virtio_net_hdr *gso;
703 struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
704 struct tun_xdp_hdr *hdr;
705 size_t len = iov_iter_count(from);
706 int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
707 int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
708 int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
709 int sock_hlen = nvq->sock_hlen;
713 if (unlikely(len < nvq->sock_hlen))
716 if (SKB_DATA_ALIGN(len + pad) +
717 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
720 buflen += SKB_DATA_ALIGN(len + pad);
721 alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
722 if (unlikely(!vhost_net_page_frag_refill(net, buflen,
723 alloc_frag, GFP_KERNEL)))
726 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
727 copied = copy_page_from_iter(alloc_frag->page,
729 offsetof(struct tun_xdp_hdr, gso),
731 if (copied != sock_hlen)
737 if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
738 vhost16_to_cpu(vq, gso->csum_start) +
739 vhost16_to_cpu(vq, gso->csum_offset) + 2 >
740 vhost16_to_cpu(vq, gso->hdr_len)) {
741 gso->hdr_len = cpu_to_vhost16(vq,
742 vhost16_to_cpu(vq, gso->csum_start) +
743 vhost16_to_cpu(vq, gso->csum_offset) + 2);
745 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
750 copied = copy_page_from_iter(alloc_frag->page,
751 alloc_frag->offset + pad,
756 xdp->data_hard_start = buf;
757 xdp->data = buf + pad;
758 xdp->data_end = xdp->data + len;
759 hdr->buflen = buflen;
760 xdp->frame_sz = buflen;
763 alloc_frag->offset += buflen;
770 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
772 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
773 struct vhost_virtqueue *vq = &nvq->vq;
776 struct msghdr msg = {
781 .msg_flags = MSG_DONTWAIT,
783 size_t len, total_len = 0;
786 bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
789 bool busyloop_intr = false;
791 if (nvq->done_idx == VHOST_NET_BATCH)
792 vhost_tx_batch(net, nvq, sock, &msg);
794 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
796 /* On error, stop handling until the next kick. */
797 if (unlikely(head < 0))
799 /* Nothing new? Wait for eventfd to tell us they refilled. */
800 if (head == vq->num) {
801 if (unlikely(busyloop_intr)) {
802 vhost_poll_queue(&vq->poll);
803 } else if (unlikely(vhost_enable_notify(&net->dev,
805 vhost_disable_notify(&net->dev, vq);
813 /* For simplicity, TX batching is only enabled if
814 * sndbuf is unlimited.
816 if (sock_can_batch) {
817 err = vhost_net_build_xdp(nvq, &msg.msg_iter);
820 } else if (unlikely(err != -ENOSPC)) {
821 vhost_tx_batch(net, nvq, sock, &msg);
822 vhost_discard_vq_desc(vq, 1);
823 vhost_net_enable_vq(net, vq);
827 /* We can't build XDP buff, go for single
828 * packet path but let's flush batched
831 vhost_tx_batch(net, nvq, sock, &msg);
832 msg.msg_control = NULL;
834 if (tx_can_batch(vq, total_len))
835 msg.msg_flags |= MSG_MORE;
837 msg.msg_flags &= ~MSG_MORE;
840 /* TODO: Check specific error and bomb out unless ENOBUFS? */
841 err = sock->ops->sendmsg(sock, &msg, len);
842 if (unlikely(err < 0)) {
843 vhost_discard_vq_desc(vq, 1);
844 vhost_net_enable_vq(net, vq);
848 pr_debug("Truncated TX packet: len %d != %zd\n",
851 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
852 vq->heads[nvq->done_idx].len = 0;
854 } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
856 vhost_tx_batch(net, nvq, sock, &msg);
859 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
861 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
862 struct vhost_virtqueue *vq = &nvq->vq;
865 struct msghdr msg = {
870 .msg_flags = MSG_DONTWAIT,
872 struct tun_msg_ctl ctl;
873 size_t len, total_len = 0;
875 struct vhost_net_ubuf_ref *ubufs;
876 struct ubuf_info *ubuf;
883 /* Release DMAs done buffers first */
884 vhost_zerocopy_signal_used(net, vq);
886 busyloop_intr = false;
887 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
889 /* On error, stop handling until the next kick. */
890 if (unlikely(head < 0))
892 /* Nothing new? Wait for eventfd to tell us they refilled. */
893 if (head == vq->num) {
894 if (unlikely(busyloop_intr)) {
895 vhost_poll_queue(&vq->poll);
896 } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
897 vhost_disable_notify(&net->dev, vq);
903 zcopy_used = len >= VHOST_GOODCOPY_LEN
904 && !vhost_exceeds_maxpend(net)
905 && vhost_net_tx_select_zcopy(net);
907 /* use msg_control to pass vhost zerocopy ubuf info to skb */
909 ubuf = nvq->ubuf_info + nvq->upend_idx;
910 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
911 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
912 ubuf->callback = vhost_zerocopy_callback;
913 ubuf->ctx = nvq->ubufs;
914 ubuf->desc = nvq->upend_idx;
915 refcount_set(&ubuf->refcnt, 1);
916 msg.msg_control = &ctl;
917 ctl.type = TUN_MSG_UBUF;
919 msg.msg_controllen = sizeof(ctl);
921 atomic_inc(&ubufs->refcount);
922 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
924 msg.msg_control = NULL;
928 if (tx_can_batch(vq, total_len) &&
929 likely(!vhost_exceeds_maxpend(net))) {
930 msg.msg_flags |= MSG_MORE;
932 msg.msg_flags &= ~MSG_MORE;
935 /* TODO: Check specific error and bomb out unless ENOBUFS? */
936 err = sock->ops->sendmsg(sock, &msg, len);
937 if (unlikely(err < 0)) {
939 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
940 vhost_net_ubuf_put(ubufs);
941 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
944 vhost_discard_vq_desc(vq, 1);
945 vhost_net_enable_vq(net, vq);
949 pr_debug("Truncated TX packet: "
950 " len %d != %zd\n", err, len);
952 vhost_add_used_and_signal(&net->dev, vq, head, 0);
954 vhost_zerocopy_signal_used(net, vq);
955 vhost_net_tx_packet(net);
956 } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
959 /* Expects to be always run from workqueue - which acts as
960 * read-size critical section for our kind of RCU. */
961 static void handle_tx(struct vhost_net *net)
963 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
964 struct vhost_virtqueue *vq = &nvq->vq;
967 mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
968 sock = vhost_vq_get_backend(vq);
972 if (!vq_meta_prefetch(vq))
975 vhost_disable_notify(&net->dev, vq);
976 vhost_net_disable_vq(net, vq);
978 if (vhost_sock_zcopy(sock))
979 handle_tx_zerocopy(net, sock);
981 handle_tx_copy(net, sock);
984 mutex_unlock(&vq->mutex);
987 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
989 struct sk_buff *head;
994 return vhost_net_buf_peek(rvq);
996 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
997 head = skb_peek(&sk->sk_receive_queue);
1000 if (skb_vlan_tag_present(head))
1004 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
1008 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1009 bool *busyloop_intr)
1011 struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1012 struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1013 struct vhost_virtqueue *rvq = &rnvq->vq;
1014 struct vhost_virtqueue *tvq = &tnvq->vq;
1015 int len = peek_head_len(rnvq, sk);
1017 if (!len && rvq->busyloop_timeout) {
1018 /* Flush batched heads first */
1019 vhost_net_signal_used(rnvq);
1020 /* Both tx vq and rx socket were polled here */
1021 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1023 len = peek_head_len(rnvq, sk);
1029 /* This is a multi-buffer version of vhost_get_desc, that works if
1030 * vq has read descriptors only.
1031 * @vq - the relevant virtqueue
1032 * @datalen - data length we'll be reading
1033 * @iovcount - returned count of io vectors we fill
1035 * @log_num - log offset
1036 * @quota - headcount quota, 1 for big buffer
1037 * returns number of buffer heads allocated, negative on error
1039 static int get_rx_bufs(struct vhost_virtqueue *vq,
1040 struct vring_used_elem *heads,
1043 struct vhost_log *log,
1047 unsigned int out, in;
1052 /* len is always initialized before use since we are always called with
1057 while (datalen > 0 && headcount < quota) {
1058 if (unlikely(seg >= UIO_MAXIOV)) {
1062 r = vhost_get_vq_desc(vq, vq->iov + seg,
1063 ARRAY_SIZE(vq->iov) - seg, &out,
1065 if (unlikely(r < 0))
1073 if (unlikely(out || in <= 0)) {
1074 vq_err(vq, "unexpected descriptor format for RX: "
1075 "out %d, in %d\n", out, in);
1079 if (unlikely(log)) {
1083 heads[headcount].id = cpu_to_vhost32(vq, d);
1084 len = iov_length(vq->iov + seg, in);
1085 heads[headcount].len = cpu_to_vhost32(vq, len);
1090 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1095 /* Detect overrun */
1096 if (unlikely(datalen > 0)) {
1102 vhost_discard_vq_desc(vq, headcount);
1106 /* Expects to be always run from workqueue - which acts as
1107 * read-size critical section for our kind of RCU. */
1108 static void handle_rx(struct vhost_net *net)
1110 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1111 struct vhost_virtqueue *vq = &nvq->vq;
1113 struct vhost_log *vq_log;
1114 struct msghdr msg = {
1117 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1118 .msg_controllen = 0,
1119 .msg_flags = MSG_DONTWAIT,
1121 struct virtio_net_hdr hdr = {
1123 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1125 size_t total_len = 0;
1128 size_t vhost_hlen, sock_hlen;
1129 size_t vhost_len, sock_len;
1130 bool busyloop_intr = false;
1131 struct socket *sock;
1132 struct iov_iter fixup;
1133 __virtio16 num_buffers;
1136 mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1137 sock = vhost_vq_get_backend(vq);
1141 if (!vq_meta_prefetch(vq))
1144 vhost_disable_notify(&net->dev, vq);
1145 vhost_net_disable_vq(net, vq);
1147 vhost_hlen = nvq->vhost_hlen;
1148 sock_hlen = nvq->sock_hlen;
1150 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1152 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1155 sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1159 sock_len += sock_hlen;
1160 vhost_len = sock_len + vhost_hlen;
1161 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1162 vhost_len, &in, vq_log, &log,
1163 likely(mergeable) ? UIO_MAXIOV : 1);
1164 /* On error, stop handling until the next kick. */
1165 if (unlikely(headcount < 0))
1167 /* OK, now we need to know about added descriptors. */
1169 if (unlikely(busyloop_intr)) {
1170 vhost_poll_queue(&vq->poll);
1171 } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1172 /* They have slipped one in as we were
1173 * doing that: check again. */
1174 vhost_disable_notify(&net->dev, vq);
1177 /* Nothing new? Wait for eventfd to tell us
1181 busyloop_intr = false;
1183 msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1184 /* On overrun, truncate and discard */
1185 if (unlikely(headcount > UIO_MAXIOV)) {
1186 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
1187 err = sock->ops->recvmsg(sock, &msg,
1188 1, MSG_DONTWAIT | MSG_TRUNC);
1189 pr_debug("Discarded rx packet: len %zd\n", sock_len);
1192 /* We don't need to be notified again. */
1193 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
1194 fixup = msg.msg_iter;
1195 if (unlikely((vhost_hlen))) {
1196 /* We will supply the header ourselves
1197 * TODO: support TSO.
1199 iov_iter_advance(&msg.msg_iter, vhost_hlen);
1201 err = sock->ops->recvmsg(sock, &msg,
1202 sock_len, MSG_DONTWAIT | MSG_TRUNC);
1203 /* Userspace might have consumed the packet meanwhile:
1204 * it's not supposed to do this usually, but might be hard
1205 * to prevent. Discard data we got (if any) and keep going. */
1206 if (unlikely(err != sock_len)) {
1207 pr_debug("Discarded rx packet: "
1208 " len %d, expected %zd\n", err, sock_len);
1209 vhost_discard_vq_desc(vq, headcount);
1212 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1213 if (unlikely(vhost_hlen)) {
1214 if (copy_to_iter(&hdr, sizeof(hdr),
1215 &fixup) != sizeof(hdr)) {
1216 vq_err(vq, "Unable to write vnet_hdr "
1217 "at addr %p\n", vq->iov->iov_base);
1221 /* Header came from socket; we'll need to patch
1222 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1224 iov_iter_advance(&fixup, sizeof(hdr));
1226 /* TODO: Should check and handle checksum. */
1228 num_buffers = cpu_to_vhost16(vq, headcount);
1229 if (likely(mergeable) &&
1230 copy_to_iter(&num_buffers, sizeof num_buffers,
1231 &fixup) != sizeof num_buffers) {
1232 vq_err(vq, "Failed num_buffers write");
1233 vhost_discard_vq_desc(vq, headcount);
1236 nvq->done_idx += headcount;
1237 if (nvq->done_idx > VHOST_NET_BATCH)
1238 vhost_net_signal_used(nvq);
1239 if (unlikely(vq_log))
1240 vhost_log_write(vq, vq_log, log, vhost_len,
1242 total_len += vhost_len;
1243 } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1245 if (unlikely(busyloop_intr))
1246 vhost_poll_queue(&vq->poll);
1248 vhost_net_enable_vq(net, vq);
1250 vhost_net_signal_used(nvq);
1251 mutex_unlock(&vq->mutex);
1254 static void handle_tx_kick(struct vhost_work *work)
1256 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1258 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1263 static void handle_rx_kick(struct vhost_work *work)
1265 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1267 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1272 static void handle_tx_net(struct vhost_work *work)
1274 struct vhost_net *net = container_of(work, struct vhost_net,
1275 poll[VHOST_NET_VQ_TX].work);
1279 static void handle_rx_net(struct vhost_work *work)
1281 struct vhost_net *net = container_of(work, struct vhost_net,
1282 poll[VHOST_NET_VQ_RX].work);
1286 static int vhost_net_open(struct inode *inode, struct file *f)
1288 struct vhost_net *n;
1289 struct vhost_dev *dev;
1290 struct vhost_virtqueue **vqs;
1292 struct xdp_buff *xdp;
1295 n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1298 vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1304 queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1311 n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1313 xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1320 n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1323 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1324 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1325 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1326 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1327 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1328 n->vqs[i].ubufs = NULL;
1329 n->vqs[i].ubuf_info = NULL;
1330 n->vqs[i].upend_idx = 0;
1331 n->vqs[i].done_idx = 0;
1332 n->vqs[i].batched_xdp = 0;
1333 n->vqs[i].vhost_hlen = 0;
1334 n->vqs[i].sock_hlen = 0;
1335 n->vqs[i].rx_ring = NULL;
1336 vhost_net_buf_init(&n->vqs[i].rxq);
1338 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1339 UIO_MAXIOV + VHOST_NET_BATCH,
1340 VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1343 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1344 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1346 f->private_data = n;
1347 n->page_frag.page = NULL;
1353 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1354 struct vhost_virtqueue *vq)
1356 struct socket *sock;
1357 struct vhost_net_virtqueue *nvq =
1358 container_of(vq, struct vhost_net_virtqueue, vq);
1360 mutex_lock(&vq->mutex);
1361 sock = vhost_vq_get_backend(vq);
1362 vhost_net_disable_vq(n, vq);
1363 vhost_vq_set_backend(vq, NULL);
1364 vhost_net_buf_unproduce(nvq);
1365 nvq->rx_ring = NULL;
1366 mutex_unlock(&vq->mutex);
1370 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1371 struct socket **rx_sock)
1373 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1374 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1377 static void vhost_net_flush_vq(struct vhost_net *n, int index)
1379 vhost_poll_flush(n->poll + index);
1380 vhost_poll_flush(&n->vqs[index].vq.poll);
1383 static void vhost_net_flush(struct vhost_net *n)
1385 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1386 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1387 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1388 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1390 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1391 /* Wait for all lower device DMAs done. */
1392 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1393 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1394 n->tx_flush = false;
1395 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1396 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1400 static int vhost_net_release(struct inode *inode, struct file *f)
1402 struct vhost_net *n = f->private_data;
1403 struct socket *tx_sock;
1404 struct socket *rx_sock;
1406 vhost_net_stop(n, &tx_sock, &rx_sock);
1408 vhost_dev_stop(&n->dev);
1409 vhost_dev_cleanup(&n->dev);
1410 vhost_net_vq_reset(n);
1412 sockfd_put(tx_sock);
1414 sockfd_put(rx_sock);
1415 /* Make sure no callbacks are outstanding */
1417 /* We do an extra flush before freeing memory,
1418 * since jobs can re-queue themselves. */
1420 kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1421 kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1423 if (n->page_frag.page)
1424 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1429 static struct socket *get_raw_socket(int fd)
1432 struct socket *sock = sockfd_lookup(fd, &r);
1435 return ERR_PTR(-ENOTSOCK);
1437 /* Parameter checking */
1438 if (sock->sk->sk_type != SOCK_RAW) {
1439 r = -ESOCKTNOSUPPORT;
1443 if (sock->sk->sk_family != AF_PACKET) {
1453 static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1455 struct ptr_ring *ring;
1456 ring = tun_get_tx_ring(file);
1459 ring = tap_get_ptr_ring(file);
1467 static struct socket *get_tap_socket(int fd)
1469 struct file *file = fget(fd);
1470 struct socket *sock;
1473 return ERR_PTR(-EBADF);
1474 sock = tun_get_socket(file);
1477 sock = tap_get_socket(file);
1483 static struct socket *get_socket(int fd)
1485 struct socket *sock;
1487 /* special case to disable backend */
1490 sock = get_raw_socket(fd);
1493 sock = get_tap_socket(fd);
1496 return ERR_PTR(-ENOTSOCK);
1499 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1501 struct socket *sock, *oldsock;
1502 struct vhost_virtqueue *vq;
1503 struct vhost_net_virtqueue *nvq;
1504 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1507 mutex_lock(&n->dev.mutex);
1508 r = vhost_dev_check_owner(&n->dev);
1512 if (index >= VHOST_NET_VQ_MAX) {
1516 vq = &n->vqs[index].vq;
1517 nvq = &n->vqs[index];
1518 mutex_lock(&vq->mutex);
1521 vhost_clear_msg(&n->dev);
1523 /* Verify that ring has been setup correctly. */
1524 if (!vhost_vq_access_ok(vq)) {
1528 sock = get_socket(fd);
1534 /* start polling new socket */
1535 oldsock = vhost_vq_get_backend(vq);
1536 if (sock != oldsock) {
1537 ubufs = vhost_net_ubuf_alloc(vq,
1538 sock && vhost_sock_zcopy(sock));
1539 if (IS_ERR(ubufs)) {
1544 vhost_net_disable_vq(n, vq);
1545 vhost_vq_set_backend(vq, sock);
1546 vhost_net_buf_unproduce(nvq);
1547 r = vhost_vq_init_access(vq);
1550 r = vhost_net_enable_vq(n, vq);
1553 if (index == VHOST_NET_VQ_RX) {
1555 nvq->rx_ring = get_tap_ptr_ring(sock->file);
1557 nvq->rx_ring = NULL;
1560 oldubufs = nvq->ubufs;
1564 n->tx_zcopy_err = 0;
1565 n->tx_flush = false;
1568 mutex_unlock(&vq->mutex);
1571 vhost_net_ubuf_put_wait_and_free(oldubufs);
1572 mutex_lock(&vq->mutex);
1573 vhost_zerocopy_signal_used(n, vq);
1574 mutex_unlock(&vq->mutex);
1578 vhost_net_flush_vq(n, index);
1579 sockfd_put(oldsock);
1582 mutex_unlock(&n->dev.mutex);
1586 vhost_vq_set_backend(vq, oldsock);
1587 vhost_net_enable_vq(n, vq);
1589 vhost_net_ubuf_put_wait_and_free(ubufs);
1594 mutex_unlock(&vq->mutex);
1596 mutex_unlock(&n->dev.mutex);
1600 static long vhost_net_reset_owner(struct vhost_net *n)
1602 struct socket *tx_sock = NULL;
1603 struct socket *rx_sock = NULL;
1605 struct vhost_iotlb *umem;
1607 mutex_lock(&n->dev.mutex);
1608 err = vhost_dev_check_owner(&n->dev);
1611 umem = vhost_dev_reset_owner_prepare();
1616 vhost_net_stop(n, &tx_sock, &rx_sock);
1618 vhost_dev_stop(&n->dev);
1619 vhost_dev_reset_owner(&n->dev, umem);
1620 vhost_net_vq_reset(n);
1622 mutex_unlock(&n->dev.mutex);
1624 sockfd_put(tx_sock);
1626 sockfd_put(rx_sock);
1630 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1632 size_t vhost_hlen, sock_hlen, hdr_len;
1635 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1636 (1ULL << VIRTIO_F_VERSION_1))) ?
1637 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1638 sizeof(struct virtio_net_hdr);
1639 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1640 /* vhost provides vnet_hdr */
1641 vhost_hlen = hdr_len;
1644 /* socket provides vnet_hdr */
1646 sock_hlen = hdr_len;
1648 mutex_lock(&n->dev.mutex);
1649 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1650 !vhost_log_access_ok(&n->dev))
1653 if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1654 if (vhost_init_device_iotlb(&n->dev, true))
1658 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1659 mutex_lock(&n->vqs[i].vq.mutex);
1660 n->vqs[i].vq.acked_features = features;
1661 n->vqs[i].vhost_hlen = vhost_hlen;
1662 n->vqs[i].sock_hlen = sock_hlen;
1663 mutex_unlock(&n->vqs[i].vq.mutex);
1665 mutex_unlock(&n->dev.mutex);
1669 mutex_unlock(&n->dev.mutex);
1673 static long vhost_net_set_owner(struct vhost_net *n)
1677 mutex_lock(&n->dev.mutex);
1678 if (vhost_dev_has_owner(&n->dev)) {
1682 r = vhost_net_set_ubuf_info(n);
1685 r = vhost_dev_set_owner(&n->dev);
1687 vhost_net_clear_ubuf_info(n);
1690 mutex_unlock(&n->dev.mutex);
1694 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1697 struct vhost_net *n = f->private_data;
1698 void __user *argp = (void __user *)arg;
1699 u64 __user *featurep = argp;
1700 struct vhost_vring_file backend;
1705 case VHOST_NET_SET_BACKEND:
1706 if (copy_from_user(&backend, argp, sizeof backend))
1708 return vhost_net_set_backend(n, backend.index, backend.fd);
1709 case VHOST_GET_FEATURES:
1710 features = VHOST_NET_FEATURES;
1711 if (copy_to_user(featurep, &features, sizeof features))
1714 case VHOST_SET_FEATURES:
1715 if (copy_from_user(&features, featurep, sizeof features))
1717 if (features & ~VHOST_NET_FEATURES)
1719 return vhost_net_set_features(n, features);
1720 case VHOST_GET_BACKEND_FEATURES:
1721 features = VHOST_NET_BACKEND_FEATURES;
1722 if (copy_to_user(featurep, &features, sizeof(features)))
1725 case VHOST_SET_BACKEND_FEATURES:
1726 if (copy_from_user(&features, featurep, sizeof(features)))
1728 if (features & ~VHOST_NET_BACKEND_FEATURES)
1730 vhost_set_backend_features(&n->dev, features);
1732 case VHOST_RESET_OWNER:
1733 return vhost_net_reset_owner(n);
1734 case VHOST_SET_OWNER:
1735 return vhost_net_set_owner(n);
1737 mutex_lock(&n->dev.mutex);
1738 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1739 if (r == -ENOIOCTLCMD)
1740 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1743 mutex_unlock(&n->dev.mutex);
1748 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1750 struct file *file = iocb->ki_filp;
1751 struct vhost_net *n = file->private_data;
1752 struct vhost_dev *dev = &n->dev;
1753 int noblock = file->f_flags & O_NONBLOCK;
1755 return vhost_chr_read_iter(dev, to, noblock);
1758 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1759 struct iov_iter *from)
1761 struct file *file = iocb->ki_filp;
1762 struct vhost_net *n = file->private_data;
1763 struct vhost_dev *dev = &n->dev;
1765 return vhost_chr_write_iter(dev, from);
1768 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1770 struct vhost_net *n = file->private_data;
1771 struct vhost_dev *dev = &n->dev;
1773 return vhost_chr_poll(file, dev, wait);
1776 static const struct file_operations vhost_net_fops = {
1777 .owner = THIS_MODULE,
1778 .release = vhost_net_release,
1779 .read_iter = vhost_net_chr_read_iter,
1780 .write_iter = vhost_net_chr_write_iter,
1781 .poll = vhost_net_chr_poll,
1782 .unlocked_ioctl = vhost_net_ioctl,
1783 .compat_ioctl = compat_ptr_ioctl,
1784 .open = vhost_net_open,
1785 .llseek = noop_llseek,
1788 static struct miscdevice vhost_net_misc = {
1789 .minor = VHOST_NET_MINOR,
1790 .name = "vhost-net",
1791 .fops = &vhost_net_fops,
1794 static int vhost_net_init(void)
1796 if (experimental_zcopytx)
1797 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1798 return misc_register(&vhost_net_misc);
1800 module_init(vhost_net_init);
1802 static void vhost_net_exit(void)
1804 misc_deregister(&vhost_net_misc);
1806 module_exit(vhost_net_exit);
1808 MODULE_VERSION("0.0.1");
1809 MODULE_LICENSE("GPL v2");
1810 MODULE_AUTHOR("Michael S. Tsirkin");
1811 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1812 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1813 MODULE_ALIAS("devname:vhost-net");