Loading...
1/* Copyright (C) 2009 Red Hat, Inc.
2 * Author: Michael S. Tsirkin <mst@redhat.com>
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2.
5 *
6 * virtio-net server in host kernel.
7 */
8
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/rcupdate.h>
19#include <linux/file.h>
20#include <linux/slab.h>
21
22#include <linux/net.h>
23#include <linux/if_packet.h>
24#include <linux/if_arp.h>
25#include <linux/if_tun.h>
26#include <linux/if_macvlan.h>
27
28#include <net/sock.h>
29
30#include "vhost.h"
31
32static int experimental_zcopytx;
33module_param(experimental_zcopytx, int, 0444);
34MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
35
36/* Max number of bytes transferred before requeueing the job.
37 * Using this limit prevents one virtqueue from starving others. */
38#define VHOST_NET_WEIGHT 0x80000
39
40/* MAX number of TX used buffers for outstanding zerocopy */
41#define VHOST_MAX_PEND 128
42#define VHOST_GOODCOPY_LEN 256
43
44enum {
45 VHOST_NET_VQ_RX = 0,
46 VHOST_NET_VQ_TX = 1,
47 VHOST_NET_VQ_MAX = 2,
48};
49
50enum vhost_net_poll_state {
51 VHOST_NET_POLL_DISABLED = 0,
52 VHOST_NET_POLL_STARTED = 1,
53 VHOST_NET_POLL_STOPPED = 2,
54};
55
56struct vhost_net {
57 struct vhost_dev dev;
58 struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
59 struct vhost_poll poll[VHOST_NET_VQ_MAX];
60 /* Tells us whether we are polling a socket for TX.
61 * We only do this when socket buffer fills up.
62 * Protected by tx vq lock. */
63 enum vhost_net_poll_state tx_poll_state;
64};
65
66static bool vhost_sock_zcopy(struct socket *sock)
67{
68 return unlikely(experimental_zcopytx) &&
69 sock_flag(sock->sk, SOCK_ZEROCOPY);
70}
71
72/* Pop first len bytes from iovec. Return number of segments used. */
73static int move_iovec_hdr(struct iovec *from, struct iovec *to,
74 size_t len, int iov_count)
75{
76 int seg = 0;
77 size_t size;
78
79 while (len && seg < iov_count) {
80 size = min(from->iov_len, len);
81 to->iov_base = from->iov_base;
82 to->iov_len = size;
83 from->iov_len -= size;
84 from->iov_base += size;
85 len -= size;
86 ++from;
87 ++to;
88 ++seg;
89 }
90 return seg;
91}
92/* Copy iovec entries for len bytes from iovec. */
93static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
94 size_t len, int iovcount)
95{
96 int seg = 0;
97 size_t size;
98
99 while (len && seg < iovcount) {
100 size = min(from->iov_len, len);
101 to->iov_base = from->iov_base;
102 to->iov_len = size;
103 len -= size;
104 ++from;
105 ++to;
106 ++seg;
107 }
108}
109
110/* Caller must have TX VQ lock */
111static void tx_poll_stop(struct vhost_net *net)
112{
113 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
114 return;
115 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
116 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
117}
118
119/* Caller must have TX VQ lock */
120static void tx_poll_start(struct vhost_net *net, struct socket *sock)
121{
122 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
123 return;
124 vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
125 net->tx_poll_state = VHOST_NET_POLL_STARTED;
126}
127
128/* Expects to be always run from workqueue - which acts as
129 * read-size critical section for our kind of RCU. */
130static void handle_tx(struct vhost_net *net)
131{
132 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
133 unsigned out, in, s;
134 int head;
135 struct msghdr msg = {
136 .msg_name = NULL,
137 .msg_namelen = 0,
138 .msg_control = NULL,
139 .msg_controllen = 0,
140 .msg_iov = vq->iov,
141 .msg_flags = MSG_DONTWAIT,
142 };
143 size_t len, total_len = 0;
144 int err, wmem;
145 size_t hdr_size;
146 struct socket *sock;
147 struct vhost_ubuf_ref *uninitialized_var(ubufs);
148 bool zcopy;
149
150 /* TODO: check that we are running from vhost_worker? */
151 sock = rcu_dereference_check(vq->private_data, 1);
152 if (!sock)
153 return;
154
155 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
156 if (wmem >= sock->sk->sk_sndbuf) {
157 mutex_lock(&vq->mutex);
158 tx_poll_start(net, sock);
159 mutex_unlock(&vq->mutex);
160 return;
161 }
162
163 mutex_lock(&vq->mutex);
164 vhost_disable_notify(&net->dev, vq);
165
166 if (wmem < sock->sk->sk_sndbuf / 2)
167 tx_poll_stop(net);
168 hdr_size = vq->vhost_hlen;
169 zcopy = vhost_sock_zcopy(sock);
170
171 for (;;) {
172 /* Release DMAs done buffers first */
173 if (zcopy)
174 vhost_zerocopy_signal_used(vq);
175
176 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
177 ARRAY_SIZE(vq->iov),
178 &out, &in,
179 NULL, NULL);
180 /* On error, stop handling until the next kick. */
181 if (unlikely(head < 0))
182 break;
183 /* Nothing new? Wait for eventfd to tell us they refilled. */
184 if (head == vq->num) {
185 int num_pends;
186
187 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
188 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
189 tx_poll_start(net, sock);
190 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
191 break;
192 }
193 /* If more outstanding DMAs, queue the work.
194 * Handle upend_idx wrap around
195 */
196 num_pends = likely(vq->upend_idx >= vq->done_idx) ?
197 (vq->upend_idx - vq->done_idx) :
198 (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
199 if (unlikely(num_pends > VHOST_MAX_PEND)) {
200 tx_poll_start(net, sock);
201 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
202 break;
203 }
204 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
205 vhost_disable_notify(&net->dev, vq);
206 continue;
207 }
208 break;
209 }
210 if (in) {
211 vq_err(vq, "Unexpected descriptor format for TX: "
212 "out %d, int %d\n", out, in);
213 break;
214 }
215 /* Skip header. TODO: support TSO. */
216 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
217 msg.msg_iovlen = out;
218 len = iov_length(vq->iov, out);
219 /* Sanity check */
220 if (!len) {
221 vq_err(vq, "Unexpected header len for TX: "
222 "%zd expected %zd\n",
223 iov_length(vq->hdr, s), hdr_size);
224 break;
225 }
226 /* use msg_control to pass vhost zerocopy ubuf info to skb */
227 if (zcopy) {
228 vq->heads[vq->upend_idx].id = head;
229 if (len < VHOST_GOODCOPY_LEN) {
230 /* copy don't need to wait for DMA done */
231 vq->heads[vq->upend_idx].len =
232 VHOST_DMA_DONE_LEN;
233 msg.msg_control = NULL;
234 msg.msg_controllen = 0;
235 ubufs = NULL;
236 } else {
237 struct ubuf_info *ubuf = &vq->ubuf_info[head];
238
239 vq->heads[vq->upend_idx].len = len;
240 ubuf->callback = vhost_zerocopy_callback;
241 ubuf->arg = vq->ubufs;
242 ubuf->desc = vq->upend_idx;
243 msg.msg_control = ubuf;
244 msg.msg_controllen = sizeof(ubuf);
245 ubufs = vq->ubufs;
246 kref_get(&ubufs->kref);
247 }
248 vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
249 }
250 /* TODO: Check specific error and bomb out unless ENOBUFS? */
251 err = sock->ops->sendmsg(NULL, sock, &msg, len);
252 if (unlikely(err < 0)) {
253 if (zcopy) {
254 if (ubufs)
255 vhost_ubuf_put(ubufs);
256 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
257 UIO_MAXIOV;
258 }
259 vhost_discard_vq_desc(vq, 1);
260 tx_poll_start(net, sock);
261 break;
262 }
263 if (err != len)
264 pr_debug("Truncated TX packet: "
265 " len %d != %zd\n", err, len);
266 if (!zcopy)
267 vhost_add_used_and_signal(&net->dev, vq, head, 0);
268 total_len += len;
269 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
270 vhost_poll_queue(&vq->poll);
271 break;
272 }
273 }
274
275 mutex_unlock(&vq->mutex);
276}
277
278static int peek_head_len(struct sock *sk)
279{
280 struct sk_buff *head;
281 int len = 0;
282 unsigned long flags;
283
284 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
285 head = skb_peek(&sk->sk_receive_queue);
286 if (likely(head))
287 len = head->len;
288 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
289 return len;
290}
291
292/* This is a multi-buffer version of vhost_get_desc, that works if
293 * vq has read descriptors only.
294 * @vq - the relevant virtqueue
295 * @datalen - data length we'll be reading
296 * @iovcount - returned count of io vectors we fill
297 * @log - vhost log
298 * @log_num - log offset
299 * @quota - headcount quota, 1 for big buffer
300 * returns number of buffer heads allocated, negative on error
301 */
302static int get_rx_bufs(struct vhost_virtqueue *vq,
303 struct vring_used_elem *heads,
304 int datalen,
305 unsigned *iovcount,
306 struct vhost_log *log,
307 unsigned *log_num,
308 unsigned int quota)
309{
310 unsigned int out, in;
311 int seg = 0;
312 int headcount = 0;
313 unsigned d;
314 int r, nlogs = 0;
315
316 while (datalen > 0 && headcount < quota) {
317 if (unlikely(seg >= UIO_MAXIOV)) {
318 r = -ENOBUFS;
319 goto err;
320 }
321 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
322 ARRAY_SIZE(vq->iov) - seg, &out,
323 &in, log, log_num);
324 if (d == vq->num) {
325 r = 0;
326 goto err;
327 }
328 if (unlikely(out || in <= 0)) {
329 vq_err(vq, "unexpected descriptor format for RX: "
330 "out %d, in %d\n", out, in);
331 r = -EINVAL;
332 goto err;
333 }
334 if (unlikely(log)) {
335 nlogs += *log_num;
336 log += *log_num;
337 }
338 heads[headcount].id = d;
339 heads[headcount].len = iov_length(vq->iov + seg, in);
340 datalen -= heads[headcount].len;
341 ++headcount;
342 seg += in;
343 }
344 heads[headcount - 1].len += datalen;
345 *iovcount = seg;
346 if (unlikely(log))
347 *log_num = nlogs;
348 return headcount;
349err:
350 vhost_discard_vq_desc(vq, headcount);
351 return r;
352}
353
354/* Expects to be always run from workqueue - which acts as
355 * read-size critical section for our kind of RCU. */
356static void handle_rx(struct vhost_net *net)
357{
358 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
359 unsigned uninitialized_var(in), log;
360 struct vhost_log *vq_log;
361 struct msghdr msg = {
362 .msg_name = NULL,
363 .msg_namelen = 0,
364 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
365 .msg_controllen = 0,
366 .msg_iov = vq->iov,
367 .msg_flags = MSG_DONTWAIT,
368 };
369 struct virtio_net_hdr_mrg_rxbuf hdr = {
370 .hdr.flags = 0,
371 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
372 };
373 size_t total_len = 0;
374 int err, headcount, mergeable;
375 size_t vhost_hlen, sock_hlen;
376 size_t vhost_len, sock_len;
377 /* TODO: check that we are running from vhost_worker? */
378 struct socket *sock = rcu_dereference_check(vq->private_data, 1);
379
380 if (!sock)
381 return;
382
383 mutex_lock(&vq->mutex);
384 vhost_disable_notify(&net->dev, vq);
385 vhost_hlen = vq->vhost_hlen;
386 sock_hlen = vq->sock_hlen;
387
388 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
389 vq->log : NULL;
390 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
391
392 while ((sock_len = peek_head_len(sock->sk))) {
393 sock_len += sock_hlen;
394 vhost_len = sock_len + vhost_hlen;
395 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
396 &in, vq_log, &log,
397 likely(mergeable) ? UIO_MAXIOV : 1);
398 /* On error, stop handling until the next kick. */
399 if (unlikely(headcount < 0))
400 break;
401 /* OK, now we need to know about added descriptors. */
402 if (!headcount) {
403 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
404 /* They have slipped one in as we were
405 * doing that: check again. */
406 vhost_disable_notify(&net->dev, vq);
407 continue;
408 }
409 /* Nothing new? Wait for eventfd to tell us
410 * they refilled. */
411 break;
412 }
413 /* We don't need to be notified again. */
414 if (unlikely((vhost_hlen)))
415 /* Skip header. TODO: support TSO. */
416 move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
417 else
418 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
419 * needed because recvmsg can modify msg_iov. */
420 copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
421 msg.msg_iovlen = in;
422 err = sock->ops->recvmsg(NULL, sock, &msg,
423 sock_len, MSG_DONTWAIT | MSG_TRUNC);
424 /* Userspace might have consumed the packet meanwhile:
425 * it's not supposed to do this usually, but might be hard
426 * to prevent. Discard data we got (if any) and keep going. */
427 if (unlikely(err != sock_len)) {
428 pr_debug("Discarded rx packet: "
429 " len %d, expected %zd\n", err, sock_len);
430 vhost_discard_vq_desc(vq, headcount);
431 continue;
432 }
433 if (unlikely(vhost_hlen) &&
434 memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
435 vhost_hlen)) {
436 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
437 vq->iov->iov_base);
438 break;
439 }
440 /* TODO: Should check and handle checksum. */
441 if (likely(mergeable) &&
442 memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
443 offsetof(typeof(hdr), num_buffers),
444 sizeof hdr.num_buffers)) {
445 vq_err(vq, "Failed num_buffers write");
446 vhost_discard_vq_desc(vq, headcount);
447 break;
448 }
449 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
450 headcount);
451 if (unlikely(vq_log))
452 vhost_log_write(vq, vq_log, log, vhost_len);
453 total_len += vhost_len;
454 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
455 vhost_poll_queue(&vq->poll);
456 break;
457 }
458 }
459
460 mutex_unlock(&vq->mutex);
461}
462
463static void handle_tx_kick(struct vhost_work *work)
464{
465 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
466 poll.work);
467 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
468
469 handle_tx(net);
470}
471
472static void handle_rx_kick(struct vhost_work *work)
473{
474 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
475 poll.work);
476 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
477
478 handle_rx(net);
479}
480
481static void handle_tx_net(struct vhost_work *work)
482{
483 struct vhost_net *net = container_of(work, struct vhost_net,
484 poll[VHOST_NET_VQ_TX].work);
485 handle_tx(net);
486}
487
488static void handle_rx_net(struct vhost_work *work)
489{
490 struct vhost_net *net = container_of(work, struct vhost_net,
491 poll[VHOST_NET_VQ_RX].work);
492 handle_rx(net);
493}
494
495static int vhost_net_open(struct inode *inode, struct file *f)
496{
497 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
498 struct vhost_dev *dev;
499 int r;
500
501 if (!n)
502 return -ENOMEM;
503
504 dev = &n->dev;
505 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
506 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
507 r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
508 if (r < 0) {
509 kfree(n);
510 return r;
511 }
512
513 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
514 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
515 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
516
517 f->private_data = n;
518
519 return 0;
520}
521
522static void vhost_net_disable_vq(struct vhost_net *n,
523 struct vhost_virtqueue *vq)
524{
525 if (!vq->private_data)
526 return;
527 if (vq == n->vqs + VHOST_NET_VQ_TX) {
528 tx_poll_stop(n);
529 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
530 } else
531 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
532}
533
534static void vhost_net_enable_vq(struct vhost_net *n,
535 struct vhost_virtqueue *vq)
536{
537 struct socket *sock;
538
539 sock = rcu_dereference_protected(vq->private_data,
540 lockdep_is_held(&vq->mutex));
541 if (!sock)
542 return;
543 if (vq == n->vqs + VHOST_NET_VQ_TX) {
544 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
545 tx_poll_start(n, sock);
546 } else
547 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
548}
549
550static struct socket *vhost_net_stop_vq(struct vhost_net *n,
551 struct vhost_virtqueue *vq)
552{
553 struct socket *sock;
554
555 mutex_lock(&vq->mutex);
556 sock = rcu_dereference_protected(vq->private_data,
557 lockdep_is_held(&vq->mutex));
558 vhost_net_disable_vq(n, vq);
559 rcu_assign_pointer(vq->private_data, NULL);
560 mutex_unlock(&vq->mutex);
561 return sock;
562}
563
564static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
565 struct socket **rx_sock)
566{
567 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
568 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
569}
570
571static void vhost_net_flush_vq(struct vhost_net *n, int index)
572{
573 vhost_poll_flush(n->poll + index);
574 vhost_poll_flush(&n->dev.vqs[index].poll);
575}
576
577static void vhost_net_flush(struct vhost_net *n)
578{
579 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
580 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
581}
582
583static int vhost_net_release(struct inode *inode, struct file *f)
584{
585 struct vhost_net *n = f->private_data;
586 struct socket *tx_sock;
587 struct socket *rx_sock;
588
589 vhost_net_stop(n, &tx_sock, &rx_sock);
590 vhost_net_flush(n);
591 vhost_dev_cleanup(&n->dev);
592 if (tx_sock)
593 fput(tx_sock->file);
594 if (rx_sock)
595 fput(rx_sock->file);
596 /* We do an extra flush before freeing memory,
597 * since jobs can re-queue themselves. */
598 vhost_net_flush(n);
599 kfree(n);
600 return 0;
601}
602
603static struct socket *get_raw_socket(int fd)
604{
605 struct {
606 struct sockaddr_ll sa;
607 char buf[MAX_ADDR_LEN];
608 } uaddr;
609 int uaddr_len = sizeof uaddr, r;
610 struct socket *sock = sockfd_lookup(fd, &r);
611
612 if (!sock)
613 return ERR_PTR(-ENOTSOCK);
614
615 /* Parameter checking */
616 if (sock->sk->sk_type != SOCK_RAW) {
617 r = -ESOCKTNOSUPPORT;
618 goto err;
619 }
620
621 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
622 &uaddr_len, 0);
623 if (r)
624 goto err;
625
626 if (uaddr.sa.sll_family != AF_PACKET) {
627 r = -EPFNOSUPPORT;
628 goto err;
629 }
630 return sock;
631err:
632 fput(sock->file);
633 return ERR_PTR(r);
634}
635
636static struct socket *get_tap_socket(int fd)
637{
638 struct file *file = fget(fd);
639 struct socket *sock;
640
641 if (!file)
642 return ERR_PTR(-EBADF);
643 sock = tun_get_socket(file);
644 if (!IS_ERR(sock))
645 return sock;
646 sock = macvtap_get_socket(file);
647 if (IS_ERR(sock))
648 fput(file);
649 return sock;
650}
651
652static struct socket *get_socket(int fd)
653{
654 struct socket *sock;
655
656 /* special case to disable backend */
657 if (fd == -1)
658 return NULL;
659 sock = get_raw_socket(fd);
660 if (!IS_ERR(sock))
661 return sock;
662 sock = get_tap_socket(fd);
663 if (!IS_ERR(sock))
664 return sock;
665 return ERR_PTR(-ENOTSOCK);
666}
667
668static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
669{
670 struct socket *sock, *oldsock;
671 struct vhost_virtqueue *vq;
672 struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
673 int r;
674
675 mutex_lock(&n->dev.mutex);
676 r = vhost_dev_check_owner(&n->dev);
677 if (r)
678 goto err;
679
680 if (index >= VHOST_NET_VQ_MAX) {
681 r = -ENOBUFS;
682 goto err;
683 }
684 vq = n->vqs + index;
685 mutex_lock(&vq->mutex);
686
687 /* Verify that ring has been setup correctly. */
688 if (!vhost_vq_access_ok(vq)) {
689 r = -EFAULT;
690 goto err_vq;
691 }
692 sock = get_socket(fd);
693 if (IS_ERR(sock)) {
694 r = PTR_ERR(sock);
695 goto err_vq;
696 }
697
698 /* start polling new socket */
699 oldsock = rcu_dereference_protected(vq->private_data,
700 lockdep_is_held(&vq->mutex));
701 if (sock != oldsock) {
702 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
703 if (IS_ERR(ubufs)) {
704 r = PTR_ERR(ubufs);
705 goto err_ubufs;
706 }
707 oldubufs = vq->ubufs;
708 vq->ubufs = ubufs;
709 vhost_net_disable_vq(n, vq);
710 rcu_assign_pointer(vq->private_data, sock);
711 vhost_net_enable_vq(n, vq);
712
713 r = vhost_init_used(vq);
714 if (r)
715 goto err_vq;
716 }
717
718 mutex_unlock(&vq->mutex);
719
720 if (oldubufs) {
721 vhost_ubuf_put_and_wait(oldubufs);
722 mutex_lock(&vq->mutex);
723 vhost_zerocopy_signal_used(vq);
724 mutex_unlock(&vq->mutex);
725 }
726
727 if (oldsock) {
728 vhost_net_flush_vq(n, index);
729 fput(oldsock->file);
730 }
731
732 mutex_unlock(&n->dev.mutex);
733 return 0;
734
735err_ubufs:
736 fput(sock->file);
737err_vq:
738 mutex_unlock(&vq->mutex);
739err:
740 mutex_unlock(&n->dev.mutex);
741 return r;
742}
743
744static long vhost_net_reset_owner(struct vhost_net *n)
745{
746 struct socket *tx_sock = NULL;
747 struct socket *rx_sock = NULL;
748 long err;
749
750 mutex_lock(&n->dev.mutex);
751 err = vhost_dev_check_owner(&n->dev);
752 if (err)
753 goto done;
754 vhost_net_stop(n, &tx_sock, &rx_sock);
755 vhost_net_flush(n);
756 err = vhost_dev_reset_owner(&n->dev);
757done:
758 mutex_unlock(&n->dev.mutex);
759 if (tx_sock)
760 fput(tx_sock->file);
761 if (rx_sock)
762 fput(rx_sock->file);
763 return err;
764}
765
766static int vhost_net_set_features(struct vhost_net *n, u64 features)
767{
768 size_t vhost_hlen, sock_hlen, hdr_len;
769 int i;
770
771 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
772 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
773 sizeof(struct virtio_net_hdr);
774 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
775 /* vhost provides vnet_hdr */
776 vhost_hlen = hdr_len;
777 sock_hlen = 0;
778 } else {
779 /* socket provides vnet_hdr */
780 vhost_hlen = 0;
781 sock_hlen = hdr_len;
782 }
783 mutex_lock(&n->dev.mutex);
784 if ((features & (1 << VHOST_F_LOG_ALL)) &&
785 !vhost_log_access_ok(&n->dev)) {
786 mutex_unlock(&n->dev.mutex);
787 return -EFAULT;
788 }
789 n->dev.acked_features = features;
790 smp_wmb();
791 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
792 mutex_lock(&n->vqs[i].mutex);
793 n->vqs[i].vhost_hlen = vhost_hlen;
794 n->vqs[i].sock_hlen = sock_hlen;
795 mutex_unlock(&n->vqs[i].mutex);
796 }
797 vhost_net_flush(n);
798 mutex_unlock(&n->dev.mutex);
799 return 0;
800}
801
802static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
803 unsigned long arg)
804{
805 struct vhost_net *n = f->private_data;
806 void __user *argp = (void __user *)arg;
807 u64 __user *featurep = argp;
808 struct vhost_vring_file backend;
809 u64 features;
810 int r;
811
812 switch (ioctl) {
813 case VHOST_NET_SET_BACKEND:
814 if (copy_from_user(&backend, argp, sizeof backend))
815 return -EFAULT;
816 return vhost_net_set_backend(n, backend.index, backend.fd);
817 case VHOST_GET_FEATURES:
818 features = VHOST_FEATURES;
819 if (copy_to_user(featurep, &features, sizeof features))
820 return -EFAULT;
821 return 0;
822 case VHOST_SET_FEATURES:
823 if (copy_from_user(&features, featurep, sizeof features))
824 return -EFAULT;
825 if (features & ~VHOST_FEATURES)
826 return -EOPNOTSUPP;
827 return vhost_net_set_features(n, features);
828 case VHOST_RESET_OWNER:
829 return vhost_net_reset_owner(n);
830 default:
831 mutex_lock(&n->dev.mutex);
832 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
833 vhost_net_flush(n);
834 mutex_unlock(&n->dev.mutex);
835 return r;
836 }
837}
838
839#ifdef CONFIG_COMPAT
840static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
841 unsigned long arg)
842{
843 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
844}
845#endif
846
847static const struct file_operations vhost_net_fops = {
848 .owner = THIS_MODULE,
849 .release = vhost_net_release,
850 .unlocked_ioctl = vhost_net_ioctl,
851#ifdef CONFIG_COMPAT
852 .compat_ioctl = vhost_net_compat_ioctl,
853#endif
854 .open = vhost_net_open,
855 .llseek = noop_llseek,
856};
857
858static struct miscdevice vhost_net_misc = {
859 MISC_DYNAMIC_MINOR,
860 "vhost-net",
861 &vhost_net_fops,
862};
863
864static int vhost_net_init(void)
865{
866 if (experimental_zcopytx)
867 vhost_enable_zcopy(VHOST_NET_VQ_TX);
868 return misc_register(&vhost_net_misc);
869}
870module_init(vhost_net_init);
871
872static void vhost_net_exit(void)
873{
874 misc_deregister(&vhost_net_misc);
875}
876module_exit(vhost_net_exit);
877
878MODULE_VERSION("0.0.1");
879MODULE_LICENSE("GPL v2");
880MODULE_AUTHOR("Michael S. Tsirkin");
881MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1/* Copyright (C) 2009 Red Hat, Inc.
2 * Author: Michael S. Tsirkin <mst@redhat.com>
3 *
4 * This work is licensed under the terms of the GNU GPL, version 2.
5 *
6 * virtio-net server in host kernel.
7 */
8
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>
23
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>
33
34#include <net/sock.h>
35
36#include "vhost.h"
37
38static int experimental_zcopytx = 1;
39module_param(experimental_zcopytx, int, 0444);
40MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
41 " 1 -Enable; 0 - Disable");
42
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
46
47/* Max number of packets transferred before requeueing the job.
48 * Using this limit prevents one virtqueue from starving rx. */
49#define VHOST_NET_PKT_WEIGHT(vq) ((vq)->num * 2)
50
51/* MAX number of TX used buffers for outstanding zerocopy */
52#define VHOST_MAX_PEND 128
53#define VHOST_GOODCOPY_LEN 256
54
55/*
56 * For transmit, used buffer len is unused; we override it to track buffer
57 * status internally; used for zerocopy tx only.
58 */
59/* Lower device DMA failed */
60#define VHOST_DMA_FAILED_LEN ((__force __virtio32)3)
61/* Lower device DMA done */
62#define VHOST_DMA_DONE_LEN ((__force __virtio32)2)
63/* Lower device DMA in progress */
64#define VHOST_DMA_IN_PROGRESS ((__force __virtio32)1)
65/* Buffer unused */
66#define VHOST_DMA_CLEAR_LEN ((__force __virtio32)0)
67
68#define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
69
70enum {
71 VHOST_NET_FEATURES = VHOST_FEATURES |
72 (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
73 (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
74 (1ULL << VIRTIO_F_IOMMU_PLATFORM)
75};
76
77enum {
78 VHOST_NET_VQ_RX = 0,
79 VHOST_NET_VQ_TX = 1,
80 VHOST_NET_VQ_MAX = 2,
81};
82
83struct vhost_net_ubuf_ref {
84 /* refcount follows semantics similar to kref:
85 * 0: object is released
86 * 1: no outstanding ubufs
87 * >1: outstanding ubufs
88 */
89 atomic_t refcount;
90 wait_queue_head_t wait;
91 struct vhost_virtqueue *vq;
92};
93
94#define VHOST_RX_BATCH 64
95struct vhost_net_buf {
96 void **queue;
97 int tail;
98 int head;
99};
100
101struct vhost_net_virtqueue {
102 struct vhost_virtqueue vq;
103 size_t vhost_hlen;
104 size_t sock_hlen;
105 /* vhost zerocopy support fields below: */
106 /* last used idx for outstanding DMA zerocopy buffers */
107 int upend_idx;
108 /* For TX, first used idx for DMA done zerocopy buffers
109 * For RX, number of batched heads
110 */
111 int done_idx;
112 /* an array of userspace buffers info */
113 struct ubuf_info *ubuf_info;
114 /* Reference counting for outstanding ubufs.
115 * Protected by vq mutex. Writers must also take device mutex. */
116 struct vhost_net_ubuf_ref *ubufs;
117 struct ptr_ring *rx_ring;
118 struct vhost_net_buf rxq;
119};
120
121struct vhost_net {
122 struct vhost_dev dev;
123 struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
124 struct vhost_poll poll[VHOST_NET_VQ_MAX];
125 /* Number of TX recently submitted.
126 * Protected by tx vq lock. */
127 unsigned tx_packets;
128 /* Number of times zerocopy TX recently failed.
129 * Protected by tx vq lock. */
130 unsigned tx_zcopy_err;
131 /* Flush in progress. Protected by tx vq lock. */
132 bool tx_flush;
133};
134
135static unsigned vhost_net_zcopy_mask __read_mostly;
136
137static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
138{
139 if (rxq->tail != rxq->head)
140 return rxq->queue[rxq->head];
141 else
142 return NULL;
143}
144
145static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
146{
147 return rxq->tail - rxq->head;
148}
149
150static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
151{
152 return rxq->tail == rxq->head;
153}
154
155static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
156{
157 void *ret = vhost_net_buf_get_ptr(rxq);
158 ++rxq->head;
159 return ret;
160}
161
162static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
163{
164 struct vhost_net_buf *rxq = &nvq->rxq;
165
166 rxq->head = 0;
167 rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
168 VHOST_RX_BATCH);
169 return rxq->tail;
170}
171
172static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
173{
174 struct vhost_net_buf *rxq = &nvq->rxq;
175
176 if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
177 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
178 vhost_net_buf_get_size(rxq),
179 tun_ptr_free);
180 rxq->head = rxq->tail = 0;
181 }
182}
183
184static int vhost_net_buf_peek_len(void *ptr)
185{
186 if (tun_is_xdp_buff(ptr)) {
187 struct xdp_buff *xdp = tun_ptr_to_xdp(ptr);
188
189 return xdp->data_end - xdp->data;
190 }
191
192 return __skb_array_len_with_tag(ptr);
193}
194
195static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
196{
197 struct vhost_net_buf *rxq = &nvq->rxq;
198
199 if (!vhost_net_buf_is_empty(rxq))
200 goto out;
201
202 if (!vhost_net_buf_produce(nvq))
203 return 0;
204
205out:
206 return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
207}
208
209static void vhost_net_buf_init(struct vhost_net_buf *rxq)
210{
211 rxq->head = rxq->tail = 0;
212}
213
214static void vhost_net_enable_zcopy(int vq)
215{
216 vhost_net_zcopy_mask |= 0x1 << vq;
217}
218
219static struct vhost_net_ubuf_ref *
220vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
221{
222 struct vhost_net_ubuf_ref *ubufs;
223 /* No zero copy backend? Nothing to count. */
224 if (!zcopy)
225 return NULL;
226 ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
227 if (!ubufs)
228 return ERR_PTR(-ENOMEM);
229 atomic_set(&ubufs->refcount, 1);
230 init_waitqueue_head(&ubufs->wait);
231 ubufs->vq = vq;
232 return ubufs;
233}
234
235static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
236{
237 int r = atomic_sub_return(1, &ubufs->refcount);
238 if (unlikely(!r))
239 wake_up(&ubufs->wait);
240 return r;
241}
242
243static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
244{
245 vhost_net_ubuf_put(ubufs);
246 wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
247}
248
249static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
250{
251 vhost_net_ubuf_put_and_wait(ubufs);
252 kfree(ubufs);
253}
254
255static void vhost_net_clear_ubuf_info(struct vhost_net *n)
256{
257 int i;
258
259 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
260 kfree(n->vqs[i].ubuf_info);
261 n->vqs[i].ubuf_info = NULL;
262 }
263}
264
265static int vhost_net_set_ubuf_info(struct vhost_net *n)
266{
267 bool zcopy;
268 int i;
269
270 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
271 zcopy = vhost_net_zcopy_mask & (0x1 << i);
272 if (!zcopy)
273 continue;
274 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
275 UIO_MAXIOV, GFP_KERNEL);
276 if (!n->vqs[i].ubuf_info)
277 goto err;
278 }
279 return 0;
280
281err:
282 vhost_net_clear_ubuf_info(n);
283 return -ENOMEM;
284}
285
286static void vhost_net_vq_reset(struct vhost_net *n)
287{
288 int i;
289
290 vhost_net_clear_ubuf_info(n);
291
292 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
293 n->vqs[i].done_idx = 0;
294 n->vqs[i].upend_idx = 0;
295 n->vqs[i].ubufs = NULL;
296 n->vqs[i].vhost_hlen = 0;
297 n->vqs[i].sock_hlen = 0;
298 vhost_net_buf_init(&n->vqs[i].rxq);
299 }
300
301}
302
303static void vhost_net_tx_packet(struct vhost_net *net)
304{
305 ++net->tx_packets;
306 if (net->tx_packets < 1024)
307 return;
308 net->tx_packets = 0;
309 net->tx_zcopy_err = 0;
310}
311
312static void vhost_net_tx_err(struct vhost_net *net)
313{
314 ++net->tx_zcopy_err;
315}
316
317static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
318{
319 /* TX flush waits for outstanding DMAs to be done.
320 * Don't start new DMAs.
321 */
322 return !net->tx_flush &&
323 net->tx_packets / 64 >= net->tx_zcopy_err;
324}
325
326static bool vhost_sock_zcopy(struct socket *sock)
327{
328 return unlikely(experimental_zcopytx) &&
329 sock_flag(sock->sk, SOCK_ZEROCOPY);
330}
331
332/* In case of DMA done not in order in lower device driver for some reason.
333 * upend_idx is used to track end of used idx, done_idx is used to track head
334 * of used idx. Once lower device DMA done contiguously, we will signal KVM
335 * guest used idx.
336 */
337static void vhost_zerocopy_signal_used(struct vhost_net *net,
338 struct vhost_virtqueue *vq)
339{
340 struct vhost_net_virtqueue *nvq =
341 container_of(vq, struct vhost_net_virtqueue, vq);
342 int i, add;
343 int j = 0;
344
345 for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
346 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
347 vhost_net_tx_err(net);
348 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
349 vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
350 ++j;
351 } else
352 break;
353 }
354 while (j) {
355 add = min(UIO_MAXIOV - nvq->done_idx, j);
356 vhost_add_used_and_signal_n(vq->dev, vq,
357 &vq->heads[nvq->done_idx], add);
358 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
359 j -= add;
360 }
361}
362
363static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
364{
365 struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
366 struct vhost_virtqueue *vq = ubufs->vq;
367 int cnt;
368
369 rcu_read_lock_bh();
370
371 /* set len to mark this desc buffers done DMA */
372 vq->heads[ubuf->desc].len = success ?
373 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
374 cnt = vhost_net_ubuf_put(ubufs);
375
376 /*
377 * Trigger polling thread if guest stopped submitting new buffers:
378 * in this case, the refcount after decrement will eventually reach 1.
379 * We also trigger polling periodically after each 16 packets
380 * (the value 16 here is more or less arbitrary, it's tuned to trigger
381 * less than 10% of times).
382 */
383 if (cnt <= 1 || !(cnt % 16))
384 vhost_poll_queue(&vq->poll);
385
386 rcu_read_unlock_bh();
387}
388
389static inline unsigned long busy_clock(void)
390{
391 return local_clock() >> 10;
392}
393
394static bool vhost_can_busy_poll(struct vhost_dev *dev,
395 unsigned long endtime)
396{
397 return likely(!need_resched()) &&
398 likely(!time_after(busy_clock(), endtime)) &&
399 likely(!signal_pending(current)) &&
400 !vhost_has_work(dev);
401}
402
403static void vhost_net_disable_vq(struct vhost_net *n,
404 struct vhost_virtqueue *vq)
405{
406 struct vhost_net_virtqueue *nvq =
407 container_of(vq, struct vhost_net_virtqueue, vq);
408 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
409 if (!vq->private_data)
410 return;
411 vhost_poll_stop(poll);
412}
413
414static int vhost_net_enable_vq(struct vhost_net *n,
415 struct vhost_virtqueue *vq)
416{
417 struct vhost_net_virtqueue *nvq =
418 container_of(vq, struct vhost_net_virtqueue, vq);
419 struct vhost_poll *poll = n->poll + (nvq - n->vqs);
420 struct socket *sock;
421
422 sock = vq->private_data;
423 if (!sock)
424 return 0;
425
426 return vhost_poll_start(poll, sock->file);
427}
428
429static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
430 struct vhost_virtqueue *vq,
431 struct iovec iov[], unsigned int iov_size,
432 unsigned int *out_num, unsigned int *in_num)
433{
434 unsigned long uninitialized_var(endtime);
435 int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
436 out_num, in_num, NULL, NULL);
437
438 if (r == vq->num && vq->busyloop_timeout) {
439 preempt_disable();
440 endtime = busy_clock() + vq->busyloop_timeout;
441 while (vhost_can_busy_poll(vq->dev, endtime) &&
442 vhost_vq_avail_empty(vq->dev, vq))
443 cpu_relax();
444 preempt_enable();
445 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
446 out_num, in_num, NULL, NULL);
447 }
448
449 return r;
450}
451
452static bool vhost_exceeds_maxpend(struct vhost_net *net)
453{
454 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
455 struct vhost_virtqueue *vq = &nvq->vq;
456
457 return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
458 min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
459}
460
461/* Expects to be always run from workqueue - which acts as
462 * read-size critical section for our kind of RCU. */
463static void handle_tx(struct vhost_net *net)
464{
465 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
466 struct vhost_virtqueue *vq = &nvq->vq;
467 unsigned out, in;
468 int head;
469 struct msghdr msg = {
470 .msg_name = NULL,
471 .msg_namelen = 0,
472 .msg_control = NULL,
473 .msg_controllen = 0,
474 .msg_flags = MSG_DONTWAIT,
475 };
476 size_t len, total_len = 0;
477 int err;
478 size_t hdr_size;
479 struct socket *sock;
480 struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
481 bool zcopy, zcopy_used;
482 int sent_pkts = 0;
483
484 mutex_lock(&vq->mutex);
485 sock = vq->private_data;
486 if (!sock)
487 goto out;
488
489 if (!vq_iotlb_prefetch(vq))
490 goto out;
491
492 vhost_disable_notify(&net->dev, vq);
493 vhost_net_disable_vq(net, vq);
494
495 hdr_size = nvq->vhost_hlen;
496 zcopy = nvq->ubufs;
497
498 for (;;) {
499 /* Release DMAs done buffers first */
500 if (zcopy)
501 vhost_zerocopy_signal_used(net, vq);
502
503
504 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
505 ARRAY_SIZE(vq->iov),
506 &out, &in);
507 /* On error, stop handling until the next kick. */
508 if (unlikely(head < 0))
509 break;
510 /* Nothing new? Wait for eventfd to tell us they refilled. */
511 if (head == vq->num) {
512 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
513 vhost_disable_notify(&net->dev, vq);
514 continue;
515 }
516 break;
517 }
518 if (in) {
519 vq_err(vq, "Unexpected descriptor format for TX: "
520 "out %d, int %d\n", out, in);
521 break;
522 }
523 /* Skip header. TODO: support TSO. */
524 len = iov_length(vq->iov, out);
525 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
526 iov_iter_advance(&msg.msg_iter, hdr_size);
527 /* Sanity check */
528 if (!msg_data_left(&msg)) {
529 vq_err(vq, "Unexpected header len for TX: "
530 "%zd expected %zd\n",
531 len, hdr_size);
532 break;
533 }
534 len = msg_data_left(&msg);
535
536 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
537 && !vhost_exceeds_maxpend(net)
538 && vhost_net_tx_select_zcopy(net);
539
540 /* use msg_control to pass vhost zerocopy ubuf info to skb */
541 if (zcopy_used) {
542 struct ubuf_info *ubuf;
543 ubuf = nvq->ubuf_info + nvq->upend_idx;
544
545 vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
546 vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
547 ubuf->callback = vhost_zerocopy_callback;
548 ubuf->ctx = nvq->ubufs;
549 ubuf->desc = nvq->upend_idx;
550 refcount_set(&ubuf->refcnt, 1);
551 msg.msg_control = ubuf;
552 msg.msg_controllen = sizeof(ubuf);
553 ubufs = nvq->ubufs;
554 atomic_inc(&ubufs->refcount);
555 nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
556 } else {
557 msg.msg_control = NULL;
558 ubufs = NULL;
559 }
560
561 total_len += len;
562 if (total_len < VHOST_NET_WEIGHT &&
563 !vhost_vq_avail_empty(&net->dev, vq) &&
564 likely(!vhost_exceeds_maxpend(net))) {
565 msg.msg_flags |= MSG_MORE;
566 } else {
567 msg.msg_flags &= ~MSG_MORE;
568 }
569
570 /* TODO: Check specific error and bomb out unless ENOBUFS? */
571 err = sock->ops->sendmsg(sock, &msg, len);
572 if (unlikely(err < 0)) {
573 if (zcopy_used) {
574 vhost_net_ubuf_put(ubufs);
575 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
576 % UIO_MAXIOV;
577 }
578 vhost_discard_vq_desc(vq, 1);
579 vhost_net_enable_vq(net, vq);
580 break;
581 }
582 if (err != len)
583 pr_debug("Truncated TX packet: "
584 " len %d != %zd\n", err, len);
585 if (!zcopy_used)
586 vhost_add_used_and_signal(&net->dev, vq, head, 0);
587 else
588 vhost_zerocopy_signal_used(net, vq);
589 vhost_net_tx_packet(net);
590 if (unlikely(total_len >= VHOST_NET_WEIGHT) ||
591 unlikely(++sent_pkts >= VHOST_NET_PKT_WEIGHT(vq))) {
592 vhost_poll_queue(&vq->poll);
593 break;
594 }
595 }
596out:
597 mutex_unlock(&vq->mutex);
598}
599
600static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
601{
602 struct sk_buff *head;
603 int len = 0;
604 unsigned long flags;
605
606 if (rvq->rx_ring)
607 return vhost_net_buf_peek(rvq);
608
609 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
610 head = skb_peek(&sk->sk_receive_queue);
611 if (likely(head)) {
612 len = head->len;
613 if (skb_vlan_tag_present(head))
614 len += VLAN_HLEN;
615 }
616
617 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
618 return len;
619}
620
621static int sk_has_rx_data(struct sock *sk)
622{
623 struct socket *sock = sk->sk_socket;
624
625 if (sock->ops->peek_len)
626 return sock->ops->peek_len(sock);
627
628 return skb_queue_empty(&sk->sk_receive_queue);
629}
630
631static void vhost_rx_signal_used(struct vhost_net_virtqueue *nvq)
632{
633 struct vhost_virtqueue *vq = &nvq->vq;
634 struct vhost_dev *dev = vq->dev;
635
636 if (!nvq->done_idx)
637 return;
638
639 vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
640 nvq->done_idx = 0;
641}
642
643static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
644{
645 struct vhost_net_virtqueue *rvq = &net->vqs[VHOST_NET_VQ_RX];
646 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
647 struct vhost_virtqueue *vq = &nvq->vq;
648 unsigned long uninitialized_var(endtime);
649 int len = peek_head_len(rvq, sk);
650
651 if (!len && vq->busyloop_timeout) {
652 /* Flush batched heads first */
653 vhost_rx_signal_used(rvq);
654 /* Both tx vq and rx socket were polled here */
655 mutex_lock_nested(&vq->mutex, 1);
656 vhost_disable_notify(&net->dev, vq);
657
658 preempt_disable();
659 endtime = busy_clock() + vq->busyloop_timeout;
660
661 while (vhost_can_busy_poll(&net->dev, endtime) &&
662 !sk_has_rx_data(sk) &&
663 vhost_vq_avail_empty(&net->dev, vq))
664 cpu_relax();
665
666 preempt_enable();
667
668 if (!vhost_vq_avail_empty(&net->dev, vq))
669 vhost_poll_queue(&vq->poll);
670 else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
671 vhost_disable_notify(&net->dev, vq);
672 vhost_poll_queue(&vq->poll);
673 }
674
675 mutex_unlock(&vq->mutex);
676
677 len = peek_head_len(rvq, sk);
678 }
679
680 return len;
681}
682
683/* This is a multi-buffer version of vhost_get_desc, that works if
684 * vq has read descriptors only.
685 * @vq - the relevant virtqueue
686 * @datalen - data length we'll be reading
687 * @iovcount - returned count of io vectors we fill
688 * @log - vhost log
689 * @log_num - log offset
690 * @quota - headcount quota, 1 for big buffer
691 * returns number of buffer heads allocated, negative on error
692 */
693static int get_rx_bufs(struct vhost_virtqueue *vq,
694 struct vring_used_elem *heads,
695 int datalen,
696 unsigned *iovcount,
697 struct vhost_log *log,
698 unsigned *log_num,
699 unsigned int quota)
700{
701 unsigned int out, in;
702 int seg = 0;
703 int headcount = 0;
704 unsigned d;
705 int r, nlogs = 0;
706 /* len is always initialized before use since we are always called with
707 * datalen > 0.
708 */
709 u32 uninitialized_var(len);
710
711 while (datalen > 0 && headcount < quota) {
712 if (unlikely(seg >= UIO_MAXIOV)) {
713 r = -ENOBUFS;
714 goto err;
715 }
716 r = vhost_get_vq_desc(vq, vq->iov + seg,
717 ARRAY_SIZE(vq->iov) - seg, &out,
718 &in, log, log_num);
719 if (unlikely(r < 0))
720 goto err;
721
722 d = r;
723 if (d == vq->num) {
724 r = 0;
725 goto err;
726 }
727 if (unlikely(out || in <= 0)) {
728 vq_err(vq, "unexpected descriptor format for RX: "
729 "out %d, in %d\n", out, in);
730 r = -EINVAL;
731 goto err;
732 }
733 if (unlikely(log)) {
734 nlogs += *log_num;
735 log += *log_num;
736 }
737 heads[headcount].id = cpu_to_vhost32(vq, d);
738 len = iov_length(vq->iov + seg, in);
739 heads[headcount].len = cpu_to_vhost32(vq, len);
740 datalen -= len;
741 ++headcount;
742 seg += in;
743 }
744 heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
745 *iovcount = seg;
746 if (unlikely(log))
747 *log_num = nlogs;
748
749 /* Detect overrun */
750 if (unlikely(datalen > 0)) {
751 r = UIO_MAXIOV + 1;
752 goto err;
753 }
754 return headcount;
755err:
756 vhost_discard_vq_desc(vq, headcount);
757 return r;
758}
759
760/* Expects to be always run from workqueue - which acts as
761 * read-size critical section for our kind of RCU. */
762static void handle_rx(struct vhost_net *net)
763{
764 struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
765 struct vhost_virtqueue *vq = &nvq->vq;
766 unsigned uninitialized_var(in), log;
767 struct vhost_log *vq_log;
768 struct msghdr msg = {
769 .msg_name = NULL,
770 .msg_namelen = 0,
771 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
772 .msg_controllen = 0,
773 .msg_flags = MSG_DONTWAIT,
774 };
775 struct virtio_net_hdr hdr = {
776 .flags = 0,
777 .gso_type = VIRTIO_NET_HDR_GSO_NONE
778 };
779 size_t total_len = 0;
780 int err, mergeable;
781 s16 headcount;
782 size_t vhost_hlen, sock_hlen;
783 size_t vhost_len, sock_len;
784 struct socket *sock;
785 struct iov_iter fixup;
786 __virtio16 num_buffers;
787
788 mutex_lock_nested(&vq->mutex, 0);
789 sock = vq->private_data;
790 if (!sock)
791 goto out;
792
793 if (!vq_iotlb_prefetch(vq))
794 goto out;
795
796 vhost_disable_notify(&net->dev, vq);
797 vhost_net_disable_vq(net, vq);
798
799 vhost_hlen = nvq->vhost_hlen;
800 sock_hlen = nvq->sock_hlen;
801
802 vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
803 vq->log : NULL;
804 mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
805
806 while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
807 sock_len += sock_hlen;
808 vhost_len = sock_len + vhost_hlen;
809 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
810 vhost_len, &in, vq_log, &log,
811 likely(mergeable) ? UIO_MAXIOV : 1);
812 /* On error, stop handling until the next kick. */
813 if (unlikely(headcount < 0))
814 goto out;
815 /* OK, now we need to know about added descriptors. */
816 if (!headcount) {
817 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
818 /* They have slipped one in as we were
819 * doing that: check again. */
820 vhost_disable_notify(&net->dev, vq);
821 continue;
822 }
823 /* Nothing new? Wait for eventfd to tell us
824 * they refilled. */
825 goto out;
826 }
827 if (nvq->rx_ring)
828 msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
829 /* On overrun, truncate and discard */
830 if (unlikely(headcount > UIO_MAXIOV)) {
831 iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
832 err = sock->ops->recvmsg(sock, &msg,
833 1, MSG_DONTWAIT | MSG_TRUNC);
834 pr_debug("Discarded rx packet: len %zd\n", sock_len);
835 continue;
836 }
837 /* We don't need to be notified again. */
838 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
839 fixup = msg.msg_iter;
840 if (unlikely((vhost_hlen))) {
841 /* We will supply the header ourselves
842 * TODO: support TSO.
843 */
844 iov_iter_advance(&msg.msg_iter, vhost_hlen);
845 }
846 err = sock->ops->recvmsg(sock, &msg,
847 sock_len, MSG_DONTWAIT | MSG_TRUNC);
848 /* Userspace might have consumed the packet meanwhile:
849 * it's not supposed to do this usually, but might be hard
850 * to prevent. Discard data we got (if any) and keep going. */
851 if (unlikely(err != sock_len)) {
852 pr_debug("Discarded rx packet: "
853 " len %d, expected %zd\n", err, sock_len);
854 vhost_discard_vq_desc(vq, headcount);
855 continue;
856 }
857 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
858 if (unlikely(vhost_hlen)) {
859 if (copy_to_iter(&hdr, sizeof(hdr),
860 &fixup) != sizeof(hdr)) {
861 vq_err(vq, "Unable to write vnet_hdr "
862 "at addr %p\n", vq->iov->iov_base);
863 goto out;
864 }
865 } else {
866 /* Header came from socket; we'll need to patch
867 * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
868 */
869 iov_iter_advance(&fixup, sizeof(hdr));
870 }
871 /* TODO: Should check and handle checksum. */
872
873 num_buffers = cpu_to_vhost16(vq, headcount);
874 if (likely(mergeable) &&
875 copy_to_iter(&num_buffers, sizeof num_buffers,
876 &fixup) != sizeof num_buffers) {
877 vq_err(vq, "Failed num_buffers write");
878 vhost_discard_vq_desc(vq, headcount);
879 goto out;
880 }
881 nvq->done_idx += headcount;
882 if (nvq->done_idx > VHOST_RX_BATCH)
883 vhost_rx_signal_used(nvq);
884 if (unlikely(vq_log))
885 vhost_log_write(vq, vq_log, log, vhost_len);
886 total_len += vhost_len;
887 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
888 vhost_poll_queue(&vq->poll);
889 goto out;
890 }
891 }
892 vhost_net_enable_vq(net, vq);
893out:
894 vhost_rx_signal_used(nvq);
895 mutex_unlock(&vq->mutex);
896}
897
898static void handle_tx_kick(struct vhost_work *work)
899{
900 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
901 poll.work);
902 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
903
904 handle_tx(net);
905}
906
907static void handle_rx_kick(struct vhost_work *work)
908{
909 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
910 poll.work);
911 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
912
913 handle_rx(net);
914}
915
916static void handle_tx_net(struct vhost_work *work)
917{
918 struct vhost_net *net = container_of(work, struct vhost_net,
919 poll[VHOST_NET_VQ_TX].work);
920 handle_tx(net);
921}
922
923static void handle_rx_net(struct vhost_work *work)
924{
925 struct vhost_net *net = container_of(work, struct vhost_net,
926 poll[VHOST_NET_VQ_RX].work);
927 handle_rx(net);
928}
929
930static int vhost_net_open(struct inode *inode, struct file *f)
931{
932 struct vhost_net *n;
933 struct vhost_dev *dev;
934 struct vhost_virtqueue **vqs;
935 void **queue;
936 int i;
937
938 n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
939 if (!n)
940 return -ENOMEM;
941 vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
942 if (!vqs) {
943 kvfree(n);
944 return -ENOMEM;
945 }
946
947 queue = kmalloc_array(VHOST_RX_BATCH, sizeof(void *),
948 GFP_KERNEL);
949 if (!queue) {
950 kfree(vqs);
951 kvfree(n);
952 return -ENOMEM;
953 }
954 n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
955
956 dev = &n->dev;
957 vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
958 vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
959 n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
960 n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
961 for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
962 n->vqs[i].ubufs = NULL;
963 n->vqs[i].ubuf_info = NULL;
964 n->vqs[i].upend_idx = 0;
965 n->vqs[i].done_idx = 0;
966 n->vqs[i].vhost_hlen = 0;
967 n->vqs[i].sock_hlen = 0;
968 n->vqs[i].rx_ring = NULL;
969 vhost_net_buf_init(&n->vqs[i].rxq);
970 }
971 vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
972
973 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
974 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
975
976 f->private_data = n;
977
978 return 0;
979}
980
981static struct socket *vhost_net_stop_vq(struct vhost_net *n,
982 struct vhost_virtqueue *vq)
983{
984 struct socket *sock;
985 struct vhost_net_virtqueue *nvq =
986 container_of(vq, struct vhost_net_virtqueue, vq);
987
988 mutex_lock(&vq->mutex);
989 sock = vq->private_data;
990 vhost_net_disable_vq(n, vq);
991 vq->private_data = NULL;
992 vhost_net_buf_unproduce(nvq);
993 nvq->rx_ring = NULL;
994 mutex_unlock(&vq->mutex);
995 return sock;
996}
997
998static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
999 struct socket **rx_sock)
1000{
1001 *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1002 *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1003}
1004
1005static void vhost_net_flush_vq(struct vhost_net *n, int index)
1006{
1007 vhost_poll_flush(n->poll + index);
1008 vhost_poll_flush(&n->vqs[index].vq.poll);
1009}
1010
1011static void vhost_net_flush(struct vhost_net *n)
1012{
1013 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1014 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1015 if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1016 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1017 n->tx_flush = true;
1018 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1019 /* Wait for all lower device DMAs done. */
1020 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1021 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1022 n->tx_flush = false;
1023 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1024 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1025 }
1026}
1027
1028static int vhost_net_release(struct inode *inode, struct file *f)
1029{
1030 struct vhost_net *n = f->private_data;
1031 struct socket *tx_sock;
1032 struct socket *rx_sock;
1033
1034 vhost_net_stop(n, &tx_sock, &rx_sock);
1035 vhost_net_flush(n);
1036 vhost_dev_stop(&n->dev);
1037 vhost_dev_cleanup(&n->dev);
1038 vhost_net_vq_reset(n);
1039 if (tx_sock)
1040 sockfd_put(tx_sock);
1041 if (rx_sock)
1042 sockfd_put(rx_sock);
1043 /* Make sure no callbacks are outstanding */
1044 synchronize_rcu_bh();
1045 /* We do an extra flush before freeing memory,
1046 * since jobs can re-queue themselves. */
1047 vhost_net_flush(n);
1048 kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1049 kfree(n->dev.vqs);
1050 kvfree(n);
1051 return 0;
1052}
1053
1054static struct socket *get_raw_socket(int fd)
1055{
1056 struct {
1057 struct sockaddr_ll sa;
1058 char buf[MAX_ADDR_LEN];
1059 } uaddr;
1060 int r;
1061 struct socket *sock = sockfd_lookup(fd, &r);
1062
1063 if (!sock)
1064 return ERR_PTR(-ENOTSOCK);
1065
1066 /* Parameter checking */
1067 if (sock->sk->sk_type != SOCK_RAW) {
1068 r = -ESOCKTNOSUPPORT;
1069 goto err;
1070 }
1071
1072 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 0);
1073 if (r < 0)
1074 goto err;
1075
1076 if (uaddr.sa.sll_family != AF_PACKET) {
1077 r = -EPFNOSUPPORT;
1078 goto err;
1079 }
1080 return sock;
1081err:
1082 sockfd_put(sock);
1083 return ERR_PTR(r);
1084}
1085
1086static struct ptr_ring *get_tap_ptr_ring(int fd)
1087{
1088 struct ptr_ring *ring;
1089 struct file *file = fget(fd);
1090
1091 if (!file)
1092 return NULL;
1093 ring = tun_get_tx_ring(file);
1094 if (!IS_ERR(ring))
1095 goto out;
1096 ring = tap_get_ptr_ring(file);
1097 if (!IS_ERR(ring))
1098 goto out;
1099 ring = NULL;
1100out:
1101 fput(file);
1102 return ring;
1103}
1104
1105static struct socket *get_tap_socket(int fd)
1106{
1107 struct file *file = fget(fd);
1108 struct socket *sock;
1109
1110 if (!file)
1111 return ERR_PTR(-EBADF);
1112 sock = tun_get_socket(file);
1113 if (!IS_ERR(sock))
1114 return sock;
1115 sock = tap_get_socket(file);
1116 if (IS_ERR(sock))
1117 fput(file);
1118 return sock;
1119}
1120
1121static struct socket *get_socket(int fd)
1122{
1123 struct socket *sock;
1124
1125 /* special case to disable backend */
1126 if (fd == -1)
1127 return NULL;
1128 sock = get_raw_socket(fd);
1129 if (!IS_ERR(sock))
1130 return sock;
1131 sock = get_tap_socket(fd);
1132 if (!IS_ERR(sock))
1133 return sock;
1134 return ERR_PTR(-ENOTSOCK);
1135}
1136
1137static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1138{
1139 struct socket *sock, *oldsock;
1140 struct vhost_virtqueue *vq;
1141 struct vhost_net_virtqueue *nvq;
1142 struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1143 int r;
1144
1145 mutex_lock(&n->dev.mutex);
1146 r = vhost_dev_check_owner(&n->dev);
1147 if (r)
1148 goto err;
1149
1150 if (index >= VHOST_NET_VQ_MAX) {
1151 r = -ENOBUFS;
1152 goto err;
1153 }
1154 vq = &n->vqs[index].vq;
1155 nvq = &n->vqs[index];
1156 mutex_lock(&vq->mutex);
1157
1158 /* Verify that ring has been setup correctly. */
1159 if (!vhost_vq_access_ok(vq)) {
1160 r = -EFAULT;
1161 goto err_vq;
1162 }
1163 sock = get_socket(fd);
1164 if (IS_ERR(sock)) {
1165 r = PTR_ERR(sock);
1166 goto err_vq;
1167 }
1168
1169 /* start polling new socket */
1170 oldsock = vq->private_data;
1171 if (sock != oldsock) {
1172 ubufs = vhost_net_ubuf_alloc(vq,
1173 sock && vhost_sock_zcopy(sock));
1174 if (IS_ERR(ubufs)) {
1175 r = PTR_ERR(ubufs);
1176 goto err_ubufs;
1177 }
1178
1179 vhost_net_disable_vq(n, vq);
1180 vq->private_data = sock;
1181 vhost_net_buf_unproduce(nvq);
1182 r = vhost_vq_init_access(vq);
1183 if (r)
1184 goto err_used;
1185 r = vhost_net_enable_vq(n, vq);
1186 if (r)
1187 goto err_used;
1188 if (index == VHOST_NET_VQ_RX)
1189 nvq->rx_ring = get_tap_ptr_ring(fd);
1190
1191 oldubufs = nvq->ubufs;
1192 nvq->ubufs = ubufs;
1193
1194 n->tx_packets = 0;
1195 n->tx_zcopy_err = 0;
1196 n->tx_flush = false;
1197 }
1198
1199 mutex_unlock(&vq->mutex);
1200
1201 if (oldubufs) {
1202 vhost_net_ubuf_put_wait_and_free(oldubufs);
1203 mutex_lock(&vq->mutex);
1204 vhost_zerocopy_signal_used(n, vq);
1205 mutex_unlock(&vq->mutex);
1206 }
1207
1208 if (oldsock) {
1209 vhost_net_flush_vq(n, index);
1210 sockfd_put(oldsock);
1211 }
1212
1213 mutex_unlock(&n->dev.mutex);
1214 return 0;
1215
1216err_used:
1217 vq->private_data = oldsock;
1218 vhost_net_enable_vq(n, vq);
1219 if (ubufs)
1220 vhost_net_ubuf_put_wait_and_free(ubufs);
1221err_ubufs:
1222 sockfd_put(sock);
1223err_vq:
1224 mutex_unlock(&vq->mutex);
1225err:
1226 mutex_unlock(&n->dev.mutex);
1227 return r;
1228}
1229
1230static long vhost_net_reset_owner(struct vhost_net *n)
1231{
1232 struct socket *tx_sock = NULL;
1233 struct socket *rx_sock = NULL;
1234 long err;
1235 struct vhost_umem *umem;
1236
1237 mutex_lock(&n->dev.mutex);
1238 err = vhost_dev_check_owner(&n->dev);
1239 if (err)
1240 goto done;
1241 umem = vhost_dev_reset_owner_prepare();
1242 if (!umem) {
1243 err = -ENOMEM;
1244 goto done;
1245 }
1246 vhost_net_stop(n, &tx_sock, &rx_sock);
1247 vhost_net_flush(n);
1248 vhost_dev_stop(&n->dev);
1249 vhost_dev_reset_owner(&n->dev, umem);
1250 vhost_net_vq_reset(n);
1251done:
1252 mutex_unlock(&n->dev.mutex);
1253 if (tx_sock)
1254 sockfd_put(tx_sock);
1255 if (rx_sock)
1256 sockfd_put(rx_sock);
1257 return err;
1258}
1259
1260static int vhost_net_set_features(struct vhost_net *n, u64 features)
1261{
1262 size_t vhost_hlen, sock_hlen, hdr_len;
1263 int i;
1264
1265 hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1266 (1ULL << VIRTIO_F_VERSION_1))) ?
1267 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1268 sizeof(struct virtio_net_hdr);
1269 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1270 /* vhost provides vnet_hdr */
1271 vhost_hlen = hdr_len;
1272 sock_hlen = 0;
1273 } else {
1274 /* socket provides vnet_hdr */
1275 vhost_hlen = 0;
1276 sock_hlen = hdr_len;
1277 }
1278 mutex_lock(&n->dev.mutex);
1279 if ((features & (1 << VHOST_F_LOG_ALL)) &&
1280 !vhost_log_access_ok(&n->dev))
1281 goto out_unlock;
1282
1283 if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1284 if (vhost_init_device_iotlb(&n->dev, true))
1285 goto out_unlock;
1286 }
1287
1288 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1289 mutex_lock(&n->vqs[i].vq.mutex);
1290 n->vqs[i].vq.acked_features = features;
1291 n->vqs[i].vhost_hlen = vhost_hlen;
1292 n->vqs[i].sock_hlen = sock_hlen;
1293 mutex_unlock(&n->vqs[i].vq.mutex);
1294 }
1295 mutex_unlock(&n->dev.mutex);
1296 return 0;
1297
1298out_unlock:
1299 mutex_unlock(&n->dev.mutex);
1300 return -EFAULT;
1301}
1302
1303static long vhost_net_set_owner(struct vhost_net *n)
1304{
1305 int r;
1306
1307 mutex_lock(&n->dev.mutex);
1308 if (vhost_dev_has_owner(&n->dev)) {
1309 r = -EBUSY;
1310 goto out;
1311 }
1312 r = vhost_net_set_ubuf_info(n);
1313 if (r)
1314 goto out;
1315 r = vhost_dev_set_owner(&n->dev);
1316 if (r)
1317 vhost_net_clear_ubuf_info(n);
1318 vhost_net_flush(n);
1319out:
1320 mutex_unlock(&n->dev.mutex);
1321 return r;
1322}
1323
1324static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1325 unsigned long arg)
1326{
1327 struct vhost_net *n = f->private_data;
1328 void __user *argp = (void __user *)arg;
1329 u64 __user *featurep = argp;
1330 struct vhost_vring_file backend;
1331 u64 features;
1332 int r;
1333
1334 switch (ioctl) {
1335 case VHOST_NET_SET_BACKEND:
1336 if (copy_from_user(&backend, argp, sizeof backend))
1337 return -EFAULT;
1338 return vhost_net_set_backend(n, backend.index, backend.fd);
1339 case VHOST_GET_FEATURES:
1340 features = VHOST_NET_FEATURES;
1341 if (copy_to_user(featurep, &features, sizeof features))
1342 return -EFAULT;
1343 return 0;
1344 case VHOST_SET_FEATURES:
1345 if (copy_from_user(&features, featurep, sizeof features))
1346 return -EFAULT;
1347 if (features & ~VHOST_NET_FEATURES)
1348 return -EOPNOTSUPP;
1349 return vhost_net_set_features(n, features);
1350 case VHOST_RESET_OWNER:
1351 return vhost_net_reset_owner(n);
1352 case VHOST_SET_OWNER:
1353 return vhost_net_set_owner(n);
1354 default:
1355 mutex_lock(&n->dev.mutex);
1356 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1357 if (r == -ENOIOCTLCMD)
1358 r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1359 else
1360 vhost_net_flush(n);
1361 mutex_unlock(&n->dev.mutex);
1362 return r;
1363 }
1364}
1365
1366#ifdef CONFIG_COMPAT
1367static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1368 unsigned long arg)
1369{
1370 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1371}
1372#endif
1373
1374static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1375{
1376 struct file *file = iocb->ki_filp;
1377 struct vhost_net *n = file->private_data;
1378 struct vhost_dev *dev = &n->dev;
1379 int noblock = file->f_flags & O_NONBLOCK;
1380
1381 return vhost_chr_read_iter(dev, to, noblock);
1382}
1383
1384static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1385 struct iov_iter *from)
1386{
1387 struct file *file = iocb->ki_filp;
1388 struct vhost_net *n = file->private_data;
1389 struct vhost_dev *dev = &n->dev;
1390
1391 return vhost_chr_write_iter(dev, from);
1392}
1393
1394static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1395{
1396 struct vhost_net *n = file->private_data;
1397 struct vhost_dev *dev = &n->dev;
1398
1399 return vhost_chr_poll(file, dev, wait);
1400}
1401
1402static const struct file_operations vhost_net_fops = {
1403 .owner = THIS_MODULE,
1404 .release = vhost_net_release,
1405 .read_iter = vhost_net_chr_read_iter,
1406 .write_iter = vhost_net_chr_write_iter,
1407 .poll = vhost_net_chr_poll,
1408 .unlocked_ioctl = vhost_net_ioctl,
1409#ifdef CONFIG_COMPAT
1410 .compat_ioctl = vhost_net_compat_ioctl,
1411#endif
1412 .open = vhost_net_open,
1413 .llseek = noop_llseek,
1414};
1415
1416static struct miscdevice vhost_net_misc = {
1417 .minor = VHOST_NET_MINOR,
1418 .name = "vhost-net",
1419 .fops = &vhost_net_fops,
1420};
1421
1422static int vhost_net_init(void)
1423{
1424 if (experimental_zcopytx)
1425 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1426 return misc_register(&vhost_net_misc);
1427}
1428module_init(vhost_net_init);
1429
1430static void vhost_net_exit(void)
1431{
1432 misc_deregister(&vhost_net_misc);
1433}
1434module_exit(vhost_net_exit);
1435
1436MODULE_VERSION("0.0.1");
1437MODULE_LICENSE("GPL v2");
1438MODULE_AUTHOR("Michael S. Tsirkin");
1439MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1440MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1441MODULE_ALIAS("devname:vhost-net");