blob: 91a324cc22982510bbb27420fdefd6bbd95dc37c [file] [log] [blame]
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +00001/* 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/mmu_context.h>
14#include <linux/miscdevice.h>
15#include <linux/module.h>
16#include <linux/mutex.h>
17#include <linux/workqueue.h>
18#include <linux/rcupdate.h>
19#include <linux/file.h>
20
21#include <linux/net.h>
22#include <linux/if_packet.h>
23#include <linux/if_arp.h>
24#include <linux/if_tun.h>
Arnd Bergmann501c7742010-02-18 05:46:50 +000025#include <linux/if_macvlan.h>
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +000026
27#include <net/sock.h>
28
29#include "vhost.h"
30
31/* Max number of bytes transferred before requeueing the job.
32 * Using this limit prevents one virtqueue from starving others. */
33#define VHOST_NET_WEIGHT 0x80000
34
35enum {
36 VHOST_NET_VQ_RX = 0,
37 VHOST_NET_VQ_TX = 1,
38 VHOST_NET_VQ_MAX = 2,
39};
40
41enum vhost_net_poll_state {
42 VHOST_NET_POLL_DISABLED = 0,
43 VHOST_NET_POLL_STARTED = 1,
44 VHOST_NET_POLL_STOPPED = 2,
45};
46
47struct vhost_net {
48 struct vhost_dev dev;
49 struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
50 struct vhost_poll poll[VHOST_NET_VQ_MAX];
51 /* Tells us whether we are polling a socket for TX.
52 * We only do this when socket buffer fills up.
53 * Protected by tx vq lock. */
54 enum vhost_net_poll_state tx_poll_state;
55};
56
57/* Pop first len bytes from iovec. Return number of segments used. */
58static int move_iovec_hdr(struct iovec *from, struct iovec *to,
59 size_t len, int iov_count)
60{
61 int seg = 0;
62 size_t size;
63 while (len && seg < iov_count) {
64 size = min(from->iov_len, len);
65 to->iov_base = from->iov_base;
66 to->iov_len = size;
67 from->iov_len -= size;
68 from->iov_base += size;
69 len -= size;
70 ++from;
71 ++to;
72 ++seg;
73 }
74 return seg;
75}
76
77/* Caller must have TX VQ lock */
78static void tx_poll_stop(struct vhost_net *net)
79{
80 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
81 return;
82 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
83 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
84}
85
86/* Caller must have TX VQ lock */
87static void tx_poll_start(struct vhost_net *net, struct socket *sock)
88{
89 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
90 return;
91 vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
92 net->tx_poll_state = VHOST_NET_POLL_STARTED;
93}
94
95/* Expects to be always run from workqueue - which acts as
96 * read-size critical section for our kind of RCU. */
97static void handle_tx(struct vhost_net *net)
98{
99 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
100 unsigned head, out, in, s;
101 struct msghdr msg = {
102 .msg_name = NULL,
103 .msg_namelen = 0,
104 .msg_control = NULL,
105 .msg_controllen = 0,
106 .msg_iov = vq->iov,
107 .msg_flags = MSG_DONTWAIT,
108 };
109 size_t len, total_len = 0;
110 int err, wmem;
111 size_t hdr_size;
112 struct socket *sock = rcu_dereference(vq->private_data);
113 if (!sock)
114 return;
115
116 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
117 if (wmem >= sock->sk->sk_sndbuf)
118 return;
119
120 use_mm(net->dev.mm);
121 mutex_lock(&vq->mutex);
122 vhost_disable_notify(vq);
123
124 if (wmem < sock->sk->sk_sndbuf * 2)
125 tx_poll_stop(net);
126 hdr_size = vq->hdr_size;
127
128 for (;;) {
129 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
130 ARRAY_SIZE(vq->iov),
131 &out, &in,
132 NULL, NULL);
133 /* Nothing new? Wait for eventfd to tell us they refilled. */
134 if (head == vq->num) {
135 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
136 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
137 tx_poll_start(net, sock);
138 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
139 break;
140 }
141 if (unlikely(vhost_enable_notify(vq))) {
142 vhost_disable_notify(vq);
143 continue;
144 }
145 break;
146 }
147 if (in) {
148 vq_err(vq, "Unexpected descriptor format for TX: "
149 "out %d, int %d\n", out, in);
150 break;
151 }
152 /* Skip header. TODO: support TSO. */
153 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
154 msg.msg_iovlen = out;
155 len = iov_length(vq->iov, out);
156 /* Sanity check */
157 if (!len) {
158 vq_err(vq, "Unexpected header len for TX: "
159 "%zd expected %zd\n",
160 iov_length(vq->hdr, s), hdr_size);
161 break;
162 }
163 /* TODO: Check specific error and bomb out unless ENOBUFS? */
164 err = sock->ops->sendmsg(NULL, sock, &msg, len);
165 if (unlikely(err < 0)) {
166 vhost_discard_vq_desc(vq);
167 tx_poll_start(net, sock);
168 break;
169 }
170 if (err != len)
171 pr_err("Truncated TX packet: "
172 " len %d != %zd\n", err, len);
173 vhost_add_used_and_signal(&net->dev, vq, head, 0);
174 total_len += len;
175 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
176 vhost_poll_queue(&vq->poll);
177 break;
178 }
179 }
180
181 mutex_unlock(&vq->mutex);
182 unuse_mm(net->dev.mm);
183}
184
185/* Expects to be always run from workqueue - which acts as
186 * read-size critical section for our kind of RCU. */
187static void handle_rx(struct vhost_net *net)
188{
189 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
190 unsigned head, out, in, log, s;
191 struct vhost_log *vq_log;
192 struct msghdr msg = {
193 .msg_name = NULL,
194 .msg_namelen = 0,
195 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
196 .msg_controllen = 0,
197 .msg_iov = vq->iov,
198 .msg_flags = MSG_DONTWAIT,
199 };
200
201 struct virtio_net_hdr hdr = {
202 .flags = 0,
203 .gso_type = VIRTIO_NET_HDR_GSO_NONE
204 };
205
206 size_t len, total_len = 0;
207 int err;
208 size_t hdr_size;
209 struct socket *sock = rcu_dereference(vq->private_data);
210 if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
211 return;
212
213 use_mm(net->dev.mm);
214 mutex_lock(&vq->mutex);
215 vhost_disable_notify(vq);
216 hdr_size = vq->hdr_size;
217
218 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
219 vq->log : NULL;
220
221 for (;;) {
222 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
223 ARRAY_SIZE(vq->iov),
224 &out, &in,
225 vq_log, &log);
226 /* OK, now we need to know about added descriptors. */
227 if (head == vq->num) {
228 if (unlikely(vhost_enable_notify(vq))) {
229 /* They have slipped one in as we were
230 * doing that: check again. */
231 vhost_disable_notify(vq);
232 continue;
233 }
234 /* Nothing new? Wait for eventfd to tell us
235 * they refilled. */
236 break;
237 }
238 /* We don't need to be notified again. */
239 if (out) {
240 vq_err(vq, "Unexpected descriptor format for RX: "
241 "out %d, int %d\n",
242 out, in);
243 break;
244 }
245 /* Skip header. TODO: support TSO/mergeable rx buffers. */
246 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, in);
247 msg.msg_iovlen = in;
248 len = iov_length(vq->iov, in);
249 /* Sanity check */
250 if (!len) {
251 vq_err(vq, "Unexpected header len for RX: "
252 "%zd expected %zd\n",
253 iov_length(vq->hdr, s), hdr_size);
254 break;
255 }
256 err = sock->ops->recvmsg(NULL, sock, &msg,
257 len, MSG_DONTWAIT | MSG_TRUNC);
258 /* TODO: Check specific error and bomb out unless EAGAIN? */
259 if (err < 0) {
260 vhost_discard_vq_desc(vq);
261 break;
262 }
263 /* TODO: Should check and handle checksum. */
264 if (err > len) {
265 pr_err("Discarded truncated rx packet: "
266 " len %d > %zd\n", err, len);
267 vhost_discard_vq_desc(vq);
268 continue;
269 }
270 len = err;
271 err = memcpy_toiovec(vq->hdr, (unsigned char *)&hdr, hdr_size);
272 if (err) {
273 vq_err(vq, "Unable to write vnet_hdr at addr %p: %d\n",
274 vq->iov->iov_base, err);
275 break;
276 }
277 len += hdr_size;
278 vhost_add_used_and_signal(&net->dev, vq, head, len);
279 if (unlikely(vq_log))
280 vhost_log_write(vq, vq_log, log, len);
281 total_len += len;
282 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
283 vhost_poll_queue(&vq->poll);
284 break;
285 }
286 }
287
288 mutex_unlock(&vq->mutex);
289 unuse_mm(net->dev.mm);
290}
291
292static void handle_tx_kick(struct work_struct *work)
293{
294 struct vhost_virtqueue *vq;
295 struct vhost_net *net;
296 vq = container_of(work, struct vhost_virtqueue, poll.work);
297 net = container_of(vq->dev, struct vhost_net, dev);
298 handle_tx(net);
299}
300
301static void handle_rx_kick(struct work_struct *work)
302{
303 struct vhost_virtqueue *vq;
304 struct vhost_net *net;
305 vq = container_of(work, struct vhost_virtqueue, poll.work);
306 net = container_of(vq->dev, struct vhost_net, dev);
307 handle_rx(net);
308}
309
310static void handle_tx_net(struct work_struct *work)
311{
312 struct vhost_net *net;
313 net = container_of(work, struct vhost_net, poll[VHOST_NET_VQ_TX].work);
314 handle_tx(net);
315}
316
317static void handle_rx_net(struct work_struct *work)
318{
319 struct vhost_net *net;
320 net = container_of(work, struct vhost_net, poll[VHOST_NET_VQ_RX].work);
321 handle_rx(net);
322}
323
324static int vhost_net_open(struct inode *inode, struct file *f)
325{
326 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
327 int r;
328 if (!n)
329 return -ENOMEM;
330 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
331 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
332 r = vhost_dev_init(&n->dev, n->vqs, VHOST_NET_VQ_MAX);
333 if (r < 0) {
334 kfree(n);
335 return r;
336 }
337
338 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT);
339 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN);
340 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
341
342 f->private_data = n;
343
344 return 0;
345}
346
347static void vhost_net_disable_vq(struct vhost_net *n,
348 struct vhost_virtqueue *vq)
349{
350 if (!vq->private_data)
351 return;
352 if (vq == n->vqs + VHOST_NET_VQ_TX) {
353 tx_poll_stop(n);
354 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
355 } else
356 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
357}
358
359static void vhost_net_enable_vq(struct vhost_net *n,
360 struct vhost_virtqueue *vq)
361{
362 struct socket *sock = vq->private_data;
363 if (!sock)
364 return;
365 if (vq == n->vqs + VHOST_NET_VQ_TX) {
366 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
367 tx_poll_start(n, sock);
368 } else
369 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
370}
371
372static struct socket *vhost_net_stop_vq(struct vhost_net *n,
373 struct vhost_virtqueue *vq)
374{
375 struct socket *sock;
376
377 mutex_lock(&vq->mutex);
378 sock = vq->private_data;
379 vhost_net_disable_vq(n, vq);
380 rcu_assign_pointer(vq->private_data, NULL);
381 mutex_unlock(&vq->mutex);
382 return sock;
383}
384
385static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
386 struct socket **rx_sock)
387{
388 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
389 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
390}
391
392static void vhost_net_flush_vq(struct vhost_net *n, int index)
393{
394 vhost_poll_flush(n->poll + index);
395 vhost_poll_flush(&n->dev.vqs[index].poll);
396}
397
398static void vhost_net_flush(struct vhost_net *n)
399{
400 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
401 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
402}
403
404static int vhost_net_release(struct inode *inode, struct file *f)
405{
406 struct vhost_net *n = f->private_data;
407 struct socket *tx_sock;
408 struct socket *rx_sock;
409
410 vhost_net_stop(n, &tx_sock, &rx_sock);
411 vhost_net_flush(n);
412 vhost_dev_cleanup(&n->dev);
413 if (tx_sock)
414 fput(tx_sock->file);
415 if (rx_sock)
416 fput(rx_sock->file);
417 /* We do an extra flush before freeing memory,
418 * since jobs can re-queue themselves. */
419 vhost_net_flush(n);
420 kfree(n);
421 return 0;
422}
423
424static struct socket *get_raw_socket(int fd)
425{
426 struct {
427 struct sockaddr_ll sa;
428 char buf[MAX_ADDR_LEN];
429 } uaddr;
430 int uaddr_len = sizeof uaddr, r;
431 struct socket *sock = sockfd_lookup(fd, &r);
432 if (!sock)
433 return ERR_PTR(-ENOTSOCK);
434
435 /* Parameter checking */
436 if (sock->sk->sk_type != SOCK_RAW) {
437 r = -ESOCKTNOSUPPORT;
438 goto err;
439 }
440
441 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
442 &uaddr_len, 0);
443 if (r)
444 goto err;
445
446 if (uaddr.sa.sll_family != AF_PACKET) {
447 r = -EPFNOSUPPORT;
448 goto err;
449 }
450 return sock;
451err:
452 fput(sock->file);
453 return ERR_PTR(r);
454}
455
Arnd Bergmann501c7742010-02-18 05:46:50 +0000456static struct socket *get_tap_socket(int fd)
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +0000457{
458 struct file *file = fget(fd);
459 struct socket *sock;
460 if (!file)
461 return ERR_PTR(-EBADF);
462 sock = tun_get_socket(file);
Arnd Bergmann501c7742010-02-18 05:46:50 +0000463 if (!IS_ERR(sock))
464 return sock;
465 sock = macvtap_get_socket(file);
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +0000466 if (IS_ERR(sock))
467 fput(file);
468 return sock;
469}
470
471static struct socket *get_socket(int fd)
472{
473 struct socket *sock;
474 /* special case to disable backend */
475 if (fd == -1)
476 return NULL;
477 sock = get_raw_socket(fd);
478 if (!IS_ERR(sock))
479 return sock;
Arnd Bergmann501c7742010-02-18 05:46:50 +0000480 sock = get_tap_socket(fd);
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +0000481 if (!IS_ERR(sock))
482 return sock;
483 return ERR_PTR(-ENOTSOCK);
484}
485
486static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
487{
488 struct socket *sock, *oldsock;
489 struct vhost_virtqueue *vq;
490 int r;
491
492 mutex_lock(&n->dev.mutex);
493 r = vhost_dev_check_owner(&n->dev);
494 if (r)
495 goto err;
496
497 if (index >= VHOST_NET_VQ_MAX) {
498 r = -ENOBUFS;
499 goto err;
500 }
501 vq = n->vqs + index;
502 mutex_lock(&vq->mutex);
503
504 /* Verify that ring has been setup correctly. */
505 if (!vhost_vq_access_ok(vq)) {
506 r = -EFAULT;
507 goto err;
508 }
509 sock = get_socket(fd);
510 if (IS_ERR(sock)) {
511 r = PTR_ERR(sock);
512 goto err;
513 }
514
515 /* start polling new socket */
516 oldsock = vq->private_data;
517 if (sock == oldsock)
518 goto done;
519
520 vhost_net_disable_vq(n, vq);
521 rcu_assign_pointer(vq->private_data, sock);
522 vhost_net_enable_vq(n, vq);
523 mutex_unlock(&vq->mutex);
524done:
525 if (oldsock) {
526 vhost_net_flush_vq(n, index);
527 fput(oldsock->file);
528 }
529err:
530 mutex_unlock(&n->dev.mutex);
531 return r;
532}
533
534static long vhost_net_reset_owner(struct vhost_net *n)
535{
536 struct socket *tx_sock = NULL;
537 struct socket *rx_sock = NULL;
538 long err;
539 mutex_lock(&n->dev.mutex);
540 err = vhost_dev_check_owner(&n->dev);
541 if (err)
542 goto done;
543 vhost_net_stop(n, &tx_sock, &rx_sock);
544 vhost_net_flush(n);
545 err = vhost_dev_reset_owner(&n->dev);
546done:
547 mutex_unlock(&n->dev.mutex);
548 if (tx_sock)
549 fput(tx_sock->file);
550 if (rx_sock)
551 fput(rx_sock->file);
552 return err;
553}
554
555static int vhost_net_set_features(struct vhost_net *n, u64 features)
556{
557 size_t hdr_size = features & (1 << VHOST_NET_F_VIRTIO_NET_HDR) ?
558 sizeof(struct virtio_net_hdr) : 0;
559 int i;
560 mutex_lock(&n->dev.mutex);
561 if ((features & (1 << VHOST_F_LOG_ALL)) &&
562 !vhost_log_access_ok(&n->dev)) {
563 mutex_unlock(&n->dev.mutex);
564 return -EFAULT;
565 }
566 n->dev.acked_features = features;
567 smp_wmb();
568 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
569 mutex_lock(&n->vqs[i].mutex);
570 n->vqs[i].hdr_size = hdr_size;
571 mutex_unlock(&n->vqs[i].mutex);
572 }
573 vhost_net_flush(n);
574 mutex_unlock(&n->dev.mutex);
575 return 0;
576}
577
578static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
579 unsigned long arg)
580{
581 struct vhost_net *n = f->private_data;
582 void __user *argp = (void __user *)arg;
583 u64 __user *featurep = argp;
584 struct vhost_vring_file backend;
585 u64 features;
586 int r;
587 switch (ioctl) {
588 case VHOST_NET_SET_BACKEND:
589 r = copy_from_user(&backend, argp, sizeof backend);
590 if (r < 0)
591 return r;
592 return vhost_net_set_backend(n, backend.index, backend.fd);
593 case VHOST_GET_FEATURES:
594 features = VHOST_FEATURES;
595 return copy_to_user(featurep, &features, sizeof features);
596 case VHOST_SET_FEATURES:
597 r = copy_from_user(&features, featurep, sizeof features);
598 if (r < 0)
599 return r;
600 if (features & ~VHOST_FEATURES)
601 return -EOPNOTSUPP;
602 return vhost_net_set_features(n, features);
603 case VHOST_RESET_OWNER:
604 return vhost_net_reset_owner(n);
605 default:
606 mutex_lock(&n->dev.mutex);
607 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
608 vhost_net_flush(n);
609 mutex_unlock(&n->dev.mutex);
610 return r;
611 }
612}
613
614#ifdef CONFIG_COMPAT
615static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
616 unsigned long arg)
617{
618 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
619}
620#endif
621
622const static struct file_operations vhost_net_fops = {
623 .owner = THIS_MODULE,
624 .release = vhost_net_release,
625 .unlocked_ioctl = vhost_net_ioctl,
626#ifdef CONFIG_COMPAT
627 .compat_ioctl = vhost_net_compat_ioctl,
628#endif
629 .open = vhost_net_open,
630};
631
632static struct miscdevice vhost_net_misc = {
633 VHOST_NET_MINOR,
634 "vhost-net",
635 &vhost_net_fops,
636};
637
638int vhost_net_init(void)
639{
640 int r = vhost_init();
641 if (r)
642 goto err_init;
643 r = misc_register(&vhost_net_misc);
644 if (r)
645 goto err_reg;
646 return 0;
647err_reg:
648 vhost_cleanup();
649err_init:
650 return r;
651
652}
653module_init(vhost_net_init);
654
655void vhost_net_exit(void)
656{
657 misc_deregister(&vhost_net_misc);
658 vhost_cleanup();
659}
660module_exit(vhost_net_exit);
661
662MODULE_VERSION("0.0.1");
663MODULE_LICENSE("GPL v2");
664MODULE_AUTHOR("Michael S. Tsirkin");
665MODULE_DESCRIPTION("Host kernel accelerator for virtio net");