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) |
77 (1ULL << VIRTIO_F_RING_RESET)
81 VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
90 struct vhost_net_ubuf_ref {
91 /* refcount follows semantics similar to kref:
92 * 0: object is released
93 * 1: no outstanding ubufs
94 * >1: outstanding ubufs
97 wait_queue_head_t wait;
98 struct vhost_virtqueue *vq;
101 #define VHOST_NET_BATCH 64
102 struct vhost_net_buf {
108 struct vhost_net_virtqueue {
109 struct vhost_virtqueue vq;
112 /* vhost zerocopy support fields below: */
113 /* last used idx for outstanding DMA zerocopy buffers */
115 /* For TX, first used idx for DMA done zerocopy buffers
116 * For RX, number of batched heads
119 /* Number of XDP frames batched */
121 /* an array of userspace buffers info */
122 struct ubuf_info_msgzc *ubuf_info;
123 /* Reference counting for outstanding ubufs.
124 * Protected by vq mutex. Writers must also take device mutex. */
125 struct vhost_net_ubuf_ref *ubufs;
126 struct ptr_ring *rx_ring;
127 struct vhost_net_buf rxq;
128 /* Batched XDP buffs */
129 struct xdp_buff *xdp;
133 struct vhost_dev dev;
134 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
135 struct vhost_poll poll[VHOST_NET_VQ_MAX];
136 /* Number of TX recently submitted.
137 * Protected by tx vq lock. */
139 /* Number of times zerocopy TX recently failed.
140 * Protected by tx vq lock. */
141 unsigned tx_zcopy_err;
142 /* Flush in progress. Protected by tx vq lock. */
144 /* Private page frag */
145 struct page_frag page_frag;
146 /* Refcount bias of page frag */
150 static unsigned vhost_net_zcopy_mask __read_mostly;
152 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
154 if (rxq->tail != rxq->head)
155 return rxq->queue[rxq->head];
160 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
162 return rxq->tail - rxq->head;
165 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
167 return rxq->tail == rxq->head;
170 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
172 void *ret = vhost_net_buf_get_ptr(rxq);
177 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
179 struct vhost_net_buf *rxq = &nvq->rxq;
182 rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
187 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
189 struct vhost_net_buf *rxq = &nvq->rxq;
191 if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
192 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
193 vhost_net_buf_get_size(rxq),
195 rxq->head = rxq->tail = 0;
199 static int vhost_net_buf_peek_len(void *ptr)
201 if (tun_is_xdp_frame(ptr)) {
202 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
207 return __skb_array_len_with_tag(ptr);
210 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
212 struct vhost_net_buf *rxq = &nvq->rxq;
214 if (!vhost_net_buf_is_empty(rxq))
217 if (!vhost_net_buf_produce(nvq))
221 return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
224 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
226 rxq->head = rxq->tail = 0;
229 static void vhost_net_enable_zcopy(int vq)
231 vhost_net_zcopy_mask |= 0x1 << vq;
234 static struct vhost_net_ubuf_ref *
235 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
237 struct vhost_net_ubuf_ref *ubufs;
238 /* No zero copy backend? Nothing to count. */
241 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
243 return ERR_PTR(-ENOMEM);
244 atomic_set(&ubufs->refcount, 1);
245 init_waitqueue_head(&ubufs->wait);
250 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
252 int r = atomic_sub_return(1, &ubufs->refcount);
254 wake_up(&ubufs->wait);
258 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
260 vhost_net_ubuf_put(ubufs);
261 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
264 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
266 vhost_net_ubuf_put_and_wait(ubufs);
270 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
274 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
275 kfree(n->vqs[i].ubuf_info);
276 n->vqs[i].ubuf_info = NULL;
280 static int vhost_net_set_ubuf_info(struct vhost_net *n)
285 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
286 zcopy = vhost_net_zcopy_mask & (0x1 << i);
289 n->vqs[i].ubuf_info =
290 kmalloc_array(UIO_MAXIOV,
291 sizeof(*n->vqs[i].ubuf_info),
293 if (!n->vqs[i].ubuf_info)
299 vhost_net_clear_ubuf_info(n);
303 static void vhost_net_vq_reset(struct vhost_net *n)
307 vhost_net_clear_ubuf_info(n);
309 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
310 n->vqs[i].done_idx = 0;
311 n->vqs[i].upend_idx = 0;
312 n->vqs[i].ubufs = NULL;
313 n->vqs[i].vhost_hlen = 0;
314 n->vqs[i].sock_hlen = 0;
315 vhost_net_buf_init(&n->vqs[i].rxq);
320 static void vhost_net_tx_packet(struct vhost_net *net)
323 if (net->tx_packets < 1024)
326 net->tx_zcopy_err = 0;
329 static void vhost_net_tx_err(struct vhost_net *net)
334 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
336 /* TX flush waits for outstanding DMAs to be done.
337 * Don't start new DMAs.
339 return !net->tx_flush &&
340 net->tx_packets / 64 >= net->tx_zcopy_err;
343 static bool vhost_sock_zcopy(struct socket *sock)
345 return unlikely(experimental_zcopytx) &&
346 sock_flag(sock->sk, SOCK_ZEROCOPY);
349 static bool vhost_sock_xdp(struct socket *sock)
351 return sock_flag(sock->sk, SOCK_XDP);
354 /* In case of DMA done not in order in lower device driver for some reason.
355 * upend_idx is used to track end of used idx, done_idx is used to track head
356 * of used idx. Once lower device DMA done contiguously, we will signal KVM
359 static void vhost_zerocopy_signal_used(struct vhost_net *net,
360 struct vhost_virtqueue *vq)
362 struct vhost_net_virtqueue *nvq =
363 container_of(vq, struct vhost_net_virtqueue, vq);
367 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
368 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
369 vhost_net_tx_err(net);
370 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
371 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
377 add = min(UIO_MAXIOV - nvq->done_idx, j);
378 vhost_add_used_and_signal_n(vq->dev, vq,
379 &vq->heads[nvq->done_idx], add);
380 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
385 static void vhost_zerocopy_callback(struct sk_buff *skb,
386 struct ubuf_info *ubuf_base, bool success)
388 struct ubuf_info_msgzc *ubuf = uarg_to_msgzc(ubuf_base);
389 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
390 struct vhost_virtqueue *vq = ubufs->vq;
395 /* set len to mark this desc buffers done DMA */
396 vq->heads[ubuf->desc].len = success ?
397 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
398 cnt = vhost_net_ubuf_put(ubufs);
401 * Trigger polling thread if guest stopped submitting new buffers:
402 * in this case, the refcount after decrement will eventually reach 1.
403 * We also trigger polling periodically after each 16 packets
404 * (the value 16 here is more or less arbitrary, it's tuned to trigger
405 * less than 10% of times).
407 if (cnt <= 1 || !(cnt % 16))
408 vhost_poll_queue(&vq->poll);
410 rcu_read_unlock_bh();
413 static inline unsigned long busy_clock(void)
415 return local_clock() >> 10;
418 static bool vhost_can_busy_poll(unsigned long endtime)
420 return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
421 !signal_pending(current));
424 static void vhost_net_disable_vq(struct vhost_net *n,
425 struct vhost_virtqueue *vq)
427 struct vhost_net_virtqueue *nvq =
428 container_of(vq, struct vhost_net_virtqueue, vq);
429 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
430 if (!vhost_vq_get_backend(vq))
432 vhost_poll_stop(poll);
435 static int vhost_net_enable_vq(struct vhost_net *n,
436 struct vhost_virtqueue *vq)
438 struct vhost_net_virtqueue *nvq =
439 container_of(vq, struct vhost_net_virtqueue, vq);
440 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
443 sock = vhost_vq_get_backend(vq);
447 return vhost_poll_start(poll, sock->file);
450 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
452 struct vhost_virtqueue *vq = &nvq->vq;
453 struct vhost_dev *dev = vq->dev;
458 vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
462 static void vhost_tx_batch(struct vhost_net *net,
463 struct vhost_net_virtqueue *nvq,
465 struct msghdr *msghdr)
467 struct tun_msg_ctl ctl = {
469 .num = nvq->batched_xdp,
474 if (nvq->batched_xdp == 0)
477 msghdr->msg_control = &ctl;
478 msghdr->msg_controllen = sizeof(ctl);
479 err = sock->ops->sendmsg(sock, msghdr, 0);
480 if (unlikely(err < 0)) {
481 vq_err(&nvq->vq, "Fail to batch sending packets\n");
483 /* free pages owned by XDP; since this is an unlikely error path,
484 * keep it simple and avoid more complex bulk update for the
487 for (i = 0; i < nvq->batched_xdp; ++i)
488 put_page(virt_to_head_page(nvq->xdp[i].data));
489 nvq->batched_xdp = 0;
495 vhost_net_signal_used(nvq);
496 nvq->batched_xdp = 0;
499 static int sock_has_rx_data(struct socket *sock)
504 if (sock->ops->peek_len)
505 return sock->ops->peek_len(sock);
507 return skb_queue_empty(&sock->sk->sk_receive_queue);
510 static void vhost_net_busy_poll_try_queue(struct vhost_net *net,
511 struct vhost_virtqueue *vq)
513 if (!vhost_vq_avail_empty(&net->dev, vq)) {
514 vhost_poll_queue(&vq->poll);
515 } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
516 vhost_disable_notify(&net->dev, vq);
517 vhost_poll_queue(&vq->poll);
521 static void vhost_net_busy_poll(struct vhost_net *net,
522 struct vhost_virtqueue *rvq,
523 struct vhost_virtqueue *tvq,
527 unsigned long busyloop_timeout;
528 unsigned long endtime;
530 struct vhost_virtqueue *vq = poll_rx ? tvq : rvq;
532 /* Try to hold the vq mutex of the paired virtqueue. We can't
533 * use mutex_lock() here since we could not guarantee a
534 * consistenet lock ordering.
536 if (!mutex_trylock(&vq->mutex))
539 vhost_disable_notify(&net->dev, vq);
540 sock = vhost_vq_get_backend(rvq);
542 busyloop_timeout = poll_rx ? rvq->busyloop_timeout:
543 tvq->busyloop_timeout;
546 endtime = busy_clock() + busyloop_timeout;
548 while (vhost_can_busy_poll(endtime)) {
549 if (vhost_has_work(&net->dev)) {
550 *busyloop_intr = true;
554 if ((sock_has_rx_data(sock) &&
555 !vhost_vq_avail_empty(&net->dev, rvq)) ||
556 !vhost_vq_avail_empty(&net->dev, tvq))
564 if (poll_rx || sock_has_rx_data(sock))
565 vhost_net_busy_poll_try_queue(net, vq);
566 else if (!poll_rx) /* On tx here, sock has no rx data. */
567 vhost_enable_notify(&net->dev, rvq);
569 mutex_unlock(&vq->mutex);
572 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
573 struct vhost_net_virtqueue *tnvq,
574 unsigned int *out_num, unsigned int *in_num,
575 struct msghdr *msghdr, bool *busyloop_intr)
577 struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
578 struct vhost_virtqueue *rvq = &rnvq->vq;
579 struct vhost_virtqueue *tvq = &tnvq->vq;
581 int r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
582 out_num, in_num, NULL, NULL);
584 if (r == tvq->num && tvq->busyloop_timeout) {
585 /* Flush batched packets first */
586 if (!vhost_sock_zcopy(vhost_vq_get_backend(tvq)))
587 vhost_tx_batch(net, tnvq,
588 vhost_vq_get_backend(tvq),
591 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, false);
593 r = vhost_get_vq_desc(tvq, tvq->iov, ARRAY_SIZE(tvq->iov),
594 out_num, in_num, NULL, NULL);
600 static bool vhost_exceeds_maxpend(struct vhost_net *net)
602 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
603 struct vhost_virtqueue *vq = &nvq->vq;
605 return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
606 min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
609 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
610 size_t hdr_size, int out)
612 /* Skip header. TODO: support TSO. */
613 size_t len = iov_length(vq->iov, out);
615 iov_iter_init(iter, ITER_SOURCE, vq->iov, out, len);
616 iov_iter_advance(iter, hdr_size);
618 return iov_iter_count(iter);
621 static int get_tx_bufs(struct vhost_net *net,
622 struct vhost_net_virtqueue *nvq,
624 unsigned int *out, unsigned int *in,
625 size_t *len, bool *busyloop_intr)
627 struct vhost_virtqueue *vq = &nvq->vq;
630 ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, msg, busyloop_intr);
632 if (ret < 0 || ret == vq->num)
636 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
642 *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
644 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
645 *len, nvq->vhost_hlen);
652 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
654 return total_len < VHOST_NET_WEIGHT &&
655 !vhost_vq_avail_empty(vq->dev, vq);
658 static bool vhost_net_page_frag_refill(struct vhost_net *net, unsigned int sz,
659 struct page_frag *pfrag, gfp_t gfp)
662 if (pfrag->offset + sz <= pfrag->size)
664 __page_frag_cache_drain(pfrag->page, net->refcnt_bias);
668 net->refcnt_bias = 0;
669 if (SKB_FRAG_PAGE_ORDER) {
670 /* Avoid direct reclaim but allow kswapd to wake */
671 pfrag->page = alloc_pages((gfp & ~__GFP_DIRECT_RECLAIM) |
672 __GFP_COMP | __GFP_NOWARN |
674 SKB_FRAG_PAGE_ORDER);
675 if (likely(pfrag->page)) {
676 pfrag->size = PAGE_SIZE << SKB_FRAG_PAGE_ORDER;
680 pfrag->page = alloc_page(gfp);
681 if (likely(pfrag->page)) {
682 pfrag->size = PAGE_SIZE;
688 net->refcnt_bias = USHRT_MAX;
689 page_ref_add(pfrag->page, USHRT_MAX - 1);
693 #define VHOST_NET_RX_PAD (NET_IP_ALIGN + NET_SKB_PAD)
695 static int vhost_net_build_xdp(struct vhost_net_virtqueue *nvq,
696 struct iov_iter *from)
698 struct vhost_virtqueue *vq = &nvq->vq;
699 struct vhost_net *net = container_of(vq->dev, struct vhost_net,
701 struct socket *sock = vhost_vq_get_backend(vq);
702 struct page_frag *alloc_frag = &net->page_frag;
703 struct virtio_net_hdr *gso;
704 struct xdp_buff *xdp = &nvq->xdp[nvq->batched_xdp];
705 struct tun_xdp_hdr *hdr;
706 size_t len = iov_iter_count(from);
707 int headroom = vhost_sock_xdp(sock) ? XDP_PACKET_HEADROOM : 0;
708 int buflen = SKB_DATA_ALIGN(sizeof(struct skb_shared_info));
709 int pad = SKB_DATA_ALIGN(VHOST_NET_RX_PAD + headroom + nvq->sock_hlen);
710 int sock_hlen = nvq->sock_hlen;
714 if (unlikely(len < nvq->sock_hlen))
717 if (SKB_DATA_ALIGN(len + pad) +
718 SKB_DATA_ALIGN(sizeof(struct skb_shared_info)) > PAGE_SIZE)
721 buflen += SKB_DATA_ALIGN(len + pad);
722 alloc_frag->offset = ALIGN((u64)alloc_frag->offset, SMP_CACHE_BYTES);
723 if (unlikely(!vhost_net_page_frag_refill(net, buflen,
724 alloc_frag, GFP_KERNEL)))
727 buf = (char *)page_address(alloc_frag->page) + alloc_frag->offset;
728 copied = copy_page_from_iter(alloc_frag->page,
730 offsetof(struct tun_xdp_hdr, gso),
732 if (copied != sock_hlen)
738 if ((gso->flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
739 vhost16_to_cpu(vq, gso->csum_start) +
740 vhost16_to_cpu(vq, gso->csum_offset) + 2 >
741 vhost16_to_cpu(vq, gso->hdr_len)) {
742 gso->hdr_len = cpu_to_vhost16(vq,
743 vhost16_to_cpu(vq, gso->csum_start) +
744 vhost16_to_cpu(vq, gso->csum_offset) + 2);
746 if (vhost16_to_cpu(vq, gso->hdr_len) > len)
751 copied = copy_page_from_iter(alloc_frag->page,
752 alloc_frag->offset + pad,
757 xdp_init_buff(xdp, buflen, NULL);
758 xdp_prepare_buff(xdp, buf, pad, len, true);
759 hdr->buflen = buflen;
762 alloc_frag->offset += buflen;
769 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
771 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
772 struct vhost_virtqueue *vq = &nvq->vq;
775 struct msghdr msg = {
780 .msg_flags = MSG_DONTWAIT,
782 size_t len, total_len = 0;
785 bool sock_can_batch = (sock->sk->sk_sndbuf == INT_MAX);
788 bool busyloop_intr = false;
790 if (nvq->done_idx == VHOST_NET_BATCH)
791 vhost_tx_batch(net, nvq, sock, &msg);
793 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
795 /* On error, stop handling until the next kick. */
796 if (unlikely(head < 0))
798 /* Nothing new? Wait for eventfd to tell us they refilled. */
799 if (head == vq->num) {
800 if (unlikely(busyloop_intr)) {
801 vhost_poll_queue(&vq->poll);
802 } else if (unlikely(vhost_enable_notify(&net->dev,
804 vhost_disable_notify(&net->dev, vq);
812 /* For simplicity, TX batching is only enabled if
813 * sndbuf is unlimited.
815 if (sock_can_batch) {
816 err = vhost_net_build_xdp(nvq, &msg.msg_iter);
819 } else if (unlikely(err != -ENOSPC)) {
820 vhost_tx_batch(net, nvq, sock, &msg);
821 vhost_discard_vq_desc(vq, 1);
822 vhost_net_enable_vq(net, vq);
826 /* We can't build XDP buff, go for single
827 * packet path but let's flush batched
830 vhost_tx_batch(net, nvq, sock, &msg);
831 msg.msg_control = NULL;
833 if (tx_can_batch(vq, total_len))
834 msg.msg_flags |= MSG_MORE;
836 msg.msg_flags &= ~MSG_MORE;
839 err = sock->ops->sendmsg(sock, &msg, len);
840 if (unlikely(err < 0)) {
841 if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
842 vhost_discard_vq_desc(vq, 1);
843 vhost_net_enable_vq(net, vq);
846 pr_debug("Fail to send packet: err %d", err);
847 } else if (unlikely(err != len))
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_msgzc *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->ctx = nvq->ubufs;
913 ubuf->desc = nvq->upend_idx;
914 ubuf->ubuf.callback = vhost_zerocopy_callback;
915 ubuf->ubuf.flags = SKBFL_ZEROCOPY_FRAG;
916 refcount_set(&ubuf->ubuf.refcnt, 1);
917 msg.msg_control = &ctl;
918 ctl.type = TUN_MSG_UBUF;
919 ctl.ptr = &ubuf->ubuf;
920 msg.msg_controllen = sizeof(ctl);
922 atomic_inc(&ubufs->refcount);
923 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
925 msg.msg_control = NULL;
929 if (tx_can_batch(vq, total_len) &&
930 likely(!vhost_exceeds_maxpend(net))) {
931 msg.msg_flags |= MSG_MORE;
933 msg.msg_flags &= ~MSG_MORE;
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 if (err == -EAGAIN || err == -ENOMEM || err == -ENOBUFS) {
945 vhost_discard_vq_desc(vq, 1);
946 vhost_net_enable_vq(net, vq);
949 pr_debug("Fail to send packet: err %d", err);
950 } else if (unlikely(err != len))
951 pr_debug("Truncated TX packet: "
952 " len %d != %zd\n", err, len);
954 vhost_add_used_and_signal(&net->dev, vq, head, 0);
956 vhost_zerocopy_signal_used(net, vq);
957 vhost_net_tx_packet(net);
958 } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
961 /* Expects to be always run from workqueue - which acts as
962 * read-size critical section for our kind of RCU. */
963 static void handle_tx(struct vhost_net *net)
965 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
966 struct vhost_virtqueue *vq = &nvq->vq;
969 mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_TX);
970 sock = vhost_vq_get_backend(vq);
974 if (!vq_meta_prefetch(vq))
977 vhost_disable_notify(&net->dev, vq);
978 vhost_net_disable_vq(net, vq);
980 if (vhost_sock_zcopy(sock))
981 handle_tx_zerocopy(net, sock);
983 handle_tx_copy(net, sock);
986 mutex_unlock(&vq->mutex);
989 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
991 struct sk_buff *head;
996 return vhost_net_buf_peek(rvq);
998 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
999 head = skb_peek(&sk->sk_receive_queue);
1002 if (skb_vlan_tag_present(head))
1006 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
1010 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
1011 bool *busyloop_intr)
1013 struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
1014 struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
1015 struct vhost_virtqueue *rvq = &rnvq->vq;
1016 struct vhost_virtqueue *tvq = &tnvq->vq;
1017 int len = peek_head_len(rnvq, sk);
1019 if (!len && rvq->busyloop_timeout) {
1020 /* Flush batched heads first */
1021 vhost_net_signal_used(rnvq);
1022 /* Both tx vq and rx socket were polled here */
1023 vhost_net_busy_poll(net, rvq, tvq, busyloop_intr, true);
1025 len = peek_head_len(rnvq, sk);
1031 /* This is a multi-buffer version of vhost_get_desc, that works if
1032 * vq has read descriptors only.
1033 * @vq - the relevant virtqueue
1034 * @datalen - data length we'll be reading
1035 * @iovcount - returned count of io vectors we fill
1037 * @log_num - log offset
1038 * @quota - headcount quota, 1 for big buffer
1039 * returns number of buffer heads allocated, negative on error
1041 static int get_rx_bufs(struct vhost_virtqueue *vq,
1042 struct vring_used_elem *heads,
1045 struct vhost_log *log,
1049 unsigned int out, in;
1054 /* len is always initialized before use since we are always called with
1059 while (datalen > 0 && headcount < quota) {
1060 if (unlikely(seg >= UIO_MAXIOV)) {
1064 r = vhost_get_vq_desc(vq, vq->iov + seg,
1065 ARRAY_SIZE(vq->iov) - seg, &out,
1067 if (unlikely(r < 0))
1075 if (unlikely(out || in <= 0)) {
1076 vq_err(vq, "unexpected descriptor format for RX: "
1077 "out %d, in %d\n", out, in);
1081 if (unlikely(log)) {
1085 heads[headcount].id = cpu_to_vhost32(vq, d);
1086 len = iov_length(vq->iov + seg, in);
1087 heads[headcount].len = cpu_to_vhost32(vq, len);
1092 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
1097 /* Detect overrun */
1098 if (unlikely(datalen > 0)) {
1104 vhost_discard_vq_desc(vq, headcount);
1108 /* Expects to be always run from workqueue - which acts as
1109 * read-size critical section for our kind of RCU. */
1110 static void handle_rx(struct vhost_net *net)
1112 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
1113 struct vhost_virtqueue *vq = &nvq->vq;
1115 struct vhost_log *vq_log;
1116 struct msghdr msg = {
1119 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
1120 .msg_controllen = 0,
1121 .msg_flags = MSG_DONTWAIT,
1123 struct virtio_net_hdr hdr = {
1125 .gso_type = VIRTIO_NET_HDR_GSO_NONE
1127 size_t total_len = 0;
1130 size_t vhost_hlen, sock_hlen;
1131 size_t vhost_len, sock_len;
1132 bool busyloop_intr = false;
1133 struct socket *sock;
1134 struct iov_iter fixup;
1135 __virtio16 num_buffers;
1138 mutex_lock_nested(&vq->mutex, VHOST_NET_VQ_RX);
1139 sock = vhost_vq_get_backend(vq);
1143 if (!vq_meta_prefetch(vq))
1146 vhost_disable_notify(&net->dev, vq);
1147 vhost_net_disable_vq(net, vq);
1149 vhost_hlen = nvq->vhost_hlen;
1150 sock_hlen = nvq->sock_hlen;
1152 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
1154 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
1157 sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
1161 sock_len += sock_hlen;
1162 vhost_len = sock_len + vhost_hlen;
1163 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
1164 vhost_len, &in, vq_log, &log,
1165 likely(mergeable) ? UIO_MAXIOV : 1);
1166 /* On error, stop handling until the next kick. */
1167 if (unlikely(headcount < 0))
1169 /* OK, now we need to know about added descriptors. */
1171 if (unlikely(busyloop_intr)) {
1172 vhost_poll_queue(&vq->poll);
1173 } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
1174 /* They have slipped one in as we were
1175 * doing that: check again. */
1176 vhost_disable_notify(&net->dev, vq);
1179 /* Nothing new? Wait for eventfd to tell us
1183 busyloop_intr = false;
1185 msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
1186 /* On overrun, truncate and discard */
1187 if (unlikely(headcount > UIO_MAXIOV)) {
1188 iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, 1, 1);
1189 err = sock->ops->recvmsg(sock, &msg,
1190 1, MSG_DONTWAIT | MSG_TRUNC);
1191 pr_debug("Discarded rx packet: len %zd\n", sock_len);
1194 /* We don't need to be notified again. */
1195 iov_iter_init(&msg.msg_iter, ITER_DEST, vq->iov, in, vhost_len);
1196 fixup = msg.msg_iter;
1197 if (unlikely((vhost_hlen))) {
1198 /* We will supply the header ourselves
1199 * TODO: support TSO.
1201 iov_iter_advance(&msg.msg_iter, vhost_hlen);
1203 err = sock->ops->recvmsg(sock, &msg,
1204 sock_len, MSG_DONTWAIT | MSG_TRUNC);
1205 /* Userspace might have consumed the packet meanwhile:
1206 * it's not supposed to do this usually, but might be hard
1207 * to prevent. Discard data we got (if any) and keep going. */
1208 if (unlikely(err != sock_len)) {
1209 pr_debug("Discarded rx packet: "
1210 " len %d, expected %zd\n", err, sock_len);
1211 vhost_discard_vq_desc(vq, headcount);
1214 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
1215 if (unlikely(vhost_hlen)) {
1216 if (copy_to_iter(&hdr, sizeof(hdr),
1217 &fixup) != sizeof(hdr)) {
1218 vq_err(vq, "Unable to write vnet_hdr "
1219 "at addr %p\n", vq->iov->iov_base);
1223 /* Header came from socket; we'll need to patch
1224 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
1226 iov_iter_advance(&fixup, sizeof(hdr));
1228 /* TODO: Should check and handle checksum. */
1230 num_buffers = cpu_to_vhost16(vq, headcount);
1231 if (likely(mergeable) &&
1232 copy_to_iter(&num_buffers, sizeof num_buffers,
1233 &fixup) != sizeof num_buffers) {
1234 vq_err(vq, "Failed num_buffers write");
1235 vhost_discard_vq_desc(vq, headcount);
1238 nvq->done_idx += headcount;
1239 if (nvq->done_idx > VHOST_NET_BATCH)
1240 vhost_net_signal_used(nvq);
1241 if (unlikely(vq_log))
1242 vhost_log_write(vq, vq_log, log, vhost_len,
1244 total_len += vhost_len;
1245 } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1247 if (unlikely(busyloop_intr))
1248 vhost_poll_queue(&vq->poll);
1250 vhost_net_enable_vq(net, vq);
1252 vhost_net_signal_used(nvq);
1253 mutex_unlock(&vq->mutex);
1256 static void handle_tx_kick(struct vhost_work *work)
1258 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1260 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1265 static void handle_rx_kick(struct vhost_work *work)
1267 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1269 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1274 static void handle_tx_net(struct vhost_work *work)
1276 struct vhost_net *net = container_of(work, struct vhost_net,
1277 poll[VHOST_NET_VQ_TX].work);
1281 static void handle_rx_net(struct vhost_work *work)
1283 struct vhost_net *net = container_of(work, struct vhost_net,
1284 poll[VHOST_NET_VQ_RX].work);
1288 static int vhost_net_open(struct inode *inode, struct file *f)
1290 struct vhost_net *n;
1291 struct vhost_dev *dev;
1292 struct vhost_virtqueue **vqs;
1294 struct xdp_buff *xdp;
1297 n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1300 vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1306 queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1313 n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1315 xdp = kmalloc_array(VHOST_NET_BATCH, sizeof(*xdp), GFP_KERNEL);
1322 n->vqs[VHOST_NET_VQ_TX].xdp = xdp;
1325 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1326 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1327 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1328 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1329 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1330 n->vqs[i].ubufs = NULL;
1331 n->vqs[i].ubuf_info = NULL;
1332 n->vqs[i].upend_idx = 0;
1333 n->vqs[i].done_idx = 0;
1334 n->vqs[i].batched_xdp = 0;
1335 n->vqs[i].vhost_hlen = 0;
1336 n->vqs[i].sock_hlen = 0;
1337 n->vqs[i].rx_ring = NULL;
1338 vhost_net_buf_init(&n->vqs[i].rxq);
1340 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1341 UIO_MAXIOV + VHOST_NET_BATCH,
1342 VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT, true,
1345 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1346 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1348 f->private_data = n;
1349 n->page_frag.page = NULL;
1355 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1356 struct vhost_virtqueue *vq)
1358 struct socket *sock;
1359 struct vhost_net_virtqueue *nvq =
1360 container_of(vq, struct vhost_net_virtqueue, vq);
1362 mutex_lock(&vq->mutex);
1363 sock = vhost_vq_get_backend(vq);
1364 vhost_net_disable_vq(n, vq);
1365 vhost_vq_set_backend(vq, NULL);
1366 vhost_net_buf_unproduce(nvq);
1367 nvq->rx_ring = NULL;
1368 mutex_unlock(&vq->mutex);
1372 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1373 struct socket **rx_sock)
1375 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1376 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1379 static void vhost_net_flush(struct vhost_net *n)
1381 vhost_dev_flush(&n->dev);
1382 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1383 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1385 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1386 /* Wait for all lower device DMAs done. */
1387 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1388 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1389 n->tx_flush = false;
1390 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1391 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1395 static int vhost_net_release(struct inode *inode, struct file *f)
1397 struct vhost_net *n = f->private_data;
1398 struct socket *tx_sock;
1399 struct socket *rx_sock;
1401 vhost_net_stop(n, &tx_sock, &rx_sock);
1403 vhost_dev_stop(&n->dev);
1404 vhost_dev_cleanup(&n->dev);
1405 vhost_net_vq_reset(n);
1407 sockfd_put(tx_sock);
1409 sockfd_put(rx_sock);
1410 /* Make sure no callbacks are outstanding */
1412 /* We do an extra flush before freeing memory,
1413 * since jobs can re-queue themselves. */
1415 kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1416 kfree(n->vqs[VHOST_NET_VQ_TX].xdp);
1418 if (n->page_frag.page)
1419 __page_frag_cache_drain(n->page_frag.page, n->refcnt_bias);
1424 static struct socket *get_raw_socket(int fd)
1427 struct socket *sock = sockfd_lookup(fd, &r);
1430 return ERR_PTR(-ENOTSOCK);
1432 /* Parameter checking */
1433 if (sock->sk->sk_type != SOCK_RAW) {
1434 r = -ESOCKTNOSUPPORT;
1438 if (sock->sk->sk_family != AF_PACKET) {
1448 static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1450 struct ptr_ring *ring;
1451 ring = tun_get_tx_ring(file);
1454 ring = tap_get_ptr_ring(file);
1462 static struct socket *get_tap_socket(int fd)
1464 struct file *file = fget(fd);
1465 struct socket *sock;
1468 return ERR_PTR(-EBADF);
1469 sock = tun_get_socket(file);
1472 sock = tap_get_socket(file);
1478 static struct socket *get_socket(int fd)
1480 struct socket *sock;
1482 /* special case to disable backend */
1485 sock = get_raw_socket(fd);
1488 sock = get_tap_socket(fd);
1491 return ERR_PTR(-ENOTSOCK);
1494 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1496 struct socket *sock, *oldsock;
1497 struct vhost_virtqueue *vq;
1498 struct vhost_net_virtqueue *nvq;
1499 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1502 mutex_lock(&n->dev.mutex);
1503 r = vhost_dev_check_owner(&n->dev);
1507 if (index >= VHOST_NET_VQ_MAX) {
1511 vq = &n->vqs[index].vq;
1512 nvq = &n->vqs[index];
1513 mutex_lock(&vq->mutex);
1516 vhost_clear_msg(&n->dev);
1518 /* Verify that ring has been setup correctly. */
1519 if (!vhost_vq_access_ok(vq)) {
1523 sock = get_socket(fd);
1529 /* start polling new socket */
1530 oldsock = vhost_vq_get_backend(vq);
1531 if (sock != oldsock) {
1532 ubufs = vhost_net_ubuf_alloc(vq,
1533 sock && vhost_sock_zcopy(sock));
1534 if (IS_ERR(ubufs)) {
1539 vhost_net_disable_vq(n, vq);
1540 vhost_vq_set_backend(vq, sock);
1541 vhost_net_buf_unproduce(nvq);
1542 r = vhost_vq_init_access(vq);
1545 r = vhost_net_enable_vq(n, vq);
1548 if (index == VHOST_NET_VQ_RX) {
1550 nvq->rx_ring = get_tap_ptr_ring(sock->file);
1552 nvq->rx_ring = NULL;
1555 oldubufs = nvq->ubufs;
1559 n->tx_zcopy_err = 0;
1560 n->tx_flush = false;
1563 mutex_unlock(&vq->mutex);
1566 vhost_net_ubuf_put_wait_and_free(oldubufs);
1567 mutex_lock(&vq->mutex);
1568 vhost_zerocopy_signal_used(n, vq);
1569 mutex_unlock(&vq->mutex);
1573 vhost_dev_flush(&n->dev);
1574 sockfd_put(oldsock);
1577 mutex_unlock(&n->dev.mutex);
1581 vhost_vq_set_backend(vq, oldsock);
1582 vhost_net_enable_vq(n, vq);
1584 vhost_net_ubuf_put_wait_and_free(ubufs);
1589 mutex_unlock(&vq->mutex);
1591 mutex_unlock(&n->dev.mutex);
1595 static long vhost_net_reset_owner(struct vhost_net *n)
1597 struct socket *tx_sock = NULL;
1598 struct socket *rx_sock = NULL;
1600 struct vhost_iotlb *umem;
1602 mutex_lock(&n->dev.mutex);
1603 err = vhost_dev_check_owner(&n->dev);
1606 umem = vhost_dev_reset_owner_prepare();
1611 vhost_net_stop(n, &tx_sock, &rx_sock);
1613 vhost_dev_stop(&n->dev);
1614 vhost_dev_reset_owner(&n->dev, umem);
1615 vhost_net_vq_reset(n);
1617 mutex_unlock(&n->dev.mutex);
1619 sockfd_put(tx_sock);
1621 sockfd_put(rx_sock);
1625 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1627 size_t vhost_hlen, sock_hlen, hdr_len;
1630 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1631 (1ULL << VIRTIO_F_VERSION_1))) ?
1632 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1633 sizeof(struct virtio_net_hdr);
1634 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1635 /* vhost provides vnet_hdr */
1636 vhost_hlen = hdr_len;
1639 /* socket provides vnet_hdr */
1641 sock_hlen = hdr_len;
1643 mutex_lock(&n->dev.mutex);
1644 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1645 !vhost_log_access_ok(&n->dev))
1648 if ((features & (1ULL << VIRTIO_F_ACCESS_PLATFORM))) {
1649 if (vhost_init_device_iotlb(&n->dev))
1653 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1654 mutex_lock(&n->vqs[i].vq.mutex);
1655 n->vqs[i].vq.acked_features = features;
1656 n->vqs[i].vhost_hlen = vhost_hlen;
1657 n->vqs[i].sock_hlen = sock_hlen;
1658 mutex_unlock(&n->vqs[i].vq.mutex);
1660 mutex_unlock(&n->dev.mutex);
1664 mutex_unlock(&n->dev.mutex);
1668 static long vhost_net_set_owner(struct vhost_net *n)
1672 mutex_lock(&n->dev.mutex);
1673 if (vhost_dev_has_owner(&n->dev)) {
1677 r = vhost_net_set_ubuf_info(n);
1680 r = vhost_dev_set_owner(&n->dev);
1682 vhost_net_clear_ubuf_info(n);
1685 mutex_unlock(&n->dev.mutex);
1689 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1692 struct vhost_net *n = f->private_data;
1693 void __user *argp = (void __user *)arg;
1694 u64 __user *featurep = argp;
1695 struct vhost_vring_file backend;
1700 case VHOST_NET_SET_BACKEND:
1701 if (copy_from_user(&backend, argp, sizeof backend))
1703 return vhost_net_set_backend(n, backend.index, backend.fd);
1704 case VHOST_GET_FEATURES:
1705 features = VHOST_NET_FEATURES;
1706 if (copy_to_user(featurep, &features, sizeof features))
1709 case VHOST_SET_FEATURES:
1710 if (copy_from_user(&features, featurep, sizeof features))
1712 if (features & ~VHOST_NET_FEATURES)
1714 return vhost_net_set_features(n, features);
1715 case VHOST_GET_BACKEND_FEATURES:
1716 features = VHOST_NET_BACKEND_FEATURES;
1717 if (copy_to_user(featurep, &features, sizeof(features)))
1720 case VHOST_SET_BACKEND_FEATURES:
1721 if (copy_from_user(&features, featurep, sizeof(features)))
1723 if (features & ~VHOST_NET_BACKEND_FEATURES)
1725 vhost_set_backend_features(&n->dev, features);
1727 case VHOST_RESET_OWNER:
1728 return vhost_net_reset_owner(n);
1729 case VHOST_SET_OWNER:
1730 return vhost_net_set_owner(n);
1732 mutex_lock(&n->dev.mutex);
1733 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1734 if (r == -ENOIOCTLCMD)
1735 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1738 mutex_unlock(&n->dev.mutex);
1743 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1745 struct file *file = iocb->ki_filp;
1746 struct vhost_net *n = file->private_data;
1747 struct vhost_dev *dev = &n->dev;
1748 int noblock = file->f_flags & O_NONBLOCK;
1750 return vhost_chr_read_iter(dev, to, noblock);
1753 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1754 struct iov_iter *from)
1756 struct file *file = iocb->ki_filp;
1757 struct vhost_net *n = file->private_data;
1758 struct vhost_dev *dev = &n->dev;
1760 return vhost_chr_write_iter(dev, from);
1763 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1765 struct vhost_net *n = file->private_data;
1766 struct vhost_dev *dev = &n->dev;
1768 return vhost_chr_poll(file, dev, wait);
1771 static const struct file_operations vhost_net_fops = {
1772 .owner = THIS_MODULE,
1773 .release = vhost_net_release,
1774 .read_iter = vhost_net_chr_read_iter,
1775 .write_iter = vhost_net_chr_write_iter,
1776 .poll = vhost_net_chr_poll,
1777 .unlocked_ioctl = vhost_net_ioctl,
1778 .compat_ioctl = compat_ptr_ioctl,
1779 .open = vhost_net_open,
1780 .llseek = noop_llseek,
1783 static struct miscdevice vhost_net_misc = {
1784 .minor = VHOST_NET_MINOR,
1785 .name = "vhost-net",
1786 .fops = &vhost_net_fops,
1789 static int __init vhost_net_init(void)
1791 if (experimental_zcopytx)
1792 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1793 return misc_register(&vhost_net_misc);
1795 module_init(vhost_net_init);
1797 static void __exit vhost_net_exit(void)
1799 misc_deregister(&vhost_net_misc);
1801 module_exit(vhost_net_exit);
1803 MODULE_VERSION("0.0.1");
1804 MODULE_LICENSE("GPL v2");
1805 MODULE_AUTHOR("Michael S. Tsirkin");
1806 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1807 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1808 MODULE_ALIAS("devname:vhost-net");