blob: ad37da2b6cb5bba78ac91664d93088f7ec2fb59c [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);
Sridhar Samudrala39286fa2010-02-28 19:39:16 +0200117 if (wmem >= sock->sk->sk_sndbuf) {
118 mutex_lock(&vq->mutex);
119 tx_poll_start(net, sock);
120 mutex_unlock(&vq->mutex);
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +0000121 return;
Sridhar Samudrala39286fa2010-02-28 19:39:16 +0200122 }
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +0000123
124 use_mm(net->dev.mm);
125 mutex_lock(&vq->mutex);
126 vhost_disable_notify(vq);
127
128 if (wmem < sock->sk->sk_sndbuf * 2)
129 tx_poll_stop(net);
130 hdr_size = vq->hdr_size;
131
132 for (;;) {
133 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
134 ARRAY_SIZE(vq->iov),
135 &out, &in,
136 NULL, NULL);
137 /* Nothing new? Wait for eventfd to tell us they refilled. */
138 if (head == vq->num) {
139 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
140 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
141 tx_poll_start(net, sock);
142 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
143 break;
144 }
145 if (unlikely(vhost_enable_notify(vq))) {
146 vhost_disable_notify(vq);
147 continue;
148 }
149 break;
150 }
151 if (in) {
152 vq_err(vq, "Unexpected descriptor format for TX: "
153 "out %d, int %d\n", out, in);
154 break;
155 }
156 /* Skip header. TODO: support TSO. */
157 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
158 msg.msg_iovlen = out;
159 len = iov_length(vq->iov, out);
160 /* Sanity check */
161 if (!len) {
162 vq_err(vq, "Unexpected header len for TX: "
163 "%zd expected %zd\n",
164 iov_length(vq->hdr, s), hdr_size);
165 break;
166 }
167 /* TODO: Check specific error and bomb out unless ENOBUFS? */
168 err = sock->ops->sendmsg(NULL, sock, &msg, len);
169 if (unlikely(err < 0)) {
170 vhost_discard_vq_desc(vq);
171 tx_poll_start(net, sock);
172 break;
173 }
174 if (err != len)
175 pr_err("Truncated TX packet: "
176 " len %d != %zd\n", err, len);
177 vhost_add_used_and_signal(&net->dev, vq, head, 0);
178 total_len += len;
179 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
180 vhost_poll_queue(&vq->poll);
181 break;
182 }
183 }
184
185 mutex_unlock(&vq->mutex);
186 unuse_mm(net->dev.mm);
187}
188
189/* Expects to be always run from workqueue - which acts as
190 * read-size critical section for our kind of RCU. */
191static void handle_rx(struct vhost_net *net)
192{
193 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
194 unsigned head, out, in, log, s;
195 struct vhost_log *vq_log;
196 struct msghdr msg = {
197 .msg_name = NULL,
198 .msg_namelen = 0,
199 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
200 .msg_controllen = 0,
201 .msg_iov = vq->iov,
202 .msg_flags = MSG_DONTWAIT,
203 };
204
205 struct virtio_net_hdr hdr = {
206 .flags = 0,
207 .gso_type = VIRTIO_NET_HDR_GSO_NONE
208 };
209
210 size_t len, total_len = 0;
211 int err;
212 size_t hdr_size;
213 struct socket *sock = rcu_dereference(vq->private_data);
214 if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
215 return;
216
217 use_mm(net->dev.mm);
218 mutex_lock(&vq->mutex);
219 vhost_disable_notify(vq);
220 hdr_size = vq->hdr_size;
221
222 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
223 vq->log : NULL;
224
225 for (;;) {
226 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
227 ARRAY_SIZE(vq->iov),
228 &out, &in,
229 vq_log, &log);
230 /* OK, now we need to know about added descriptors. */
231 if (head == vq->num) {
232 if (unlikely(vhost_enable_notify(vq))) {
233 /* They have slipped one in as we were
234 * doing that: check again. */
235 vhost_disable_notify(vq);
236 continue;
237 }
238 /* Nothing new? Wait for eventfd to tell us
239 * they refilled. */
240 break;
241 }
242 /* We don't need to be notified again. */
243 if (out) {
244 vq_err(vq, "Unexpected descriptor format for RX: "
245 "out %d, int %d\n",
246 out, in);
247 break;
248 }
249 /* Skip header. TODO: support TSO/mergeable rx buffers. */
250 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, in);
251 msg.msg_iovlen = in;
252 len = iov_length(vq->iov, in);
253 /* Sanity check */
254 if (!len) {
255 vq_err(vq, "Unexpected header len for RX: "
256 "%zd expected %zd\n",
257 iov_length(vq->hdr, s), hdr_size);
258 break;
259 }
260 err = sock->ops->recvmsg(NULL, sock, &msg,
261 len, MSG_DONTWAIT | MSG_TRUNC);
262 /* TODO: Check specific error and bomb out unless EAGAIN? */
263 if (err < 0) {
264 vhost_discard_vq_desc(vq);
265 break;
266 }
267 /* TODO: Should check and handle checksum. */
268 if (err > len) {
269 pr_err("Discarded truncated rx packet: "
270 " len %d > %zd\n", err, len);
271 vhost_discard_vq_desc(vq);
272 continue;
273 }
274 len = err;
275 err = memcpy_toiovec(vq->hdr, (unsigned char *)&hdr, hdr_size);
276 if (err) {
277 vq_err(vq, "Unable to write vnet_hdr at addr %p: %d\n",
278 vq->iov->iov_base, err);
279 break;
280 }
281 len += hdr_size;
282 vhost_add_used_and_signal(&net->dev, vq, head, len);
283 if (unlikely(vq_log))
284 vhost_log_write(vq, vq_log, log, len);
285 total_len += len;
286 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
287 vhost_poll_queue(&vq->poll);
288 break;
289 }
290 }
291
292 mutex_unlock(&vq->mutex);
293 unuse_mm(net->dev.mm);
294}
295
296static void handle_tx_kick(struct work_struct *work)
297{
298 struct vhost_virtqueue *vq;
299 struct vhost_net *net;
300 vq = container_of(work, struct vhost_virtqueue, poll.work);
301 net = container_of(vq->dev, struct vhost_net, dev);
302 handle_tx(net);
303}
304
305static void handle_rx_kick(struct work_struct *work)
306{
307 struct vhost_virtqueue *vq;
308 struct vhost_net *net;
309 vq = container_of(work, struct vhost_virtqueue, poll.work);
310 net = container_of(vq->dev, struct vhost_net, dev);
311 handle_rx(net);
312}
313
314static void handle_tx_net(struct work_struct *work)
315{
316 struct vhost_net *net;
317 net = container_of(work, struct vhost_net, poll[VHOST_NET_VQ_TX].work);
318 handle_tx(net);
319}
320
321static void handle_rx_net(struct work_struct *work)
322{
323 struct vhost_net *net;
324 net = container_of(work, struct vhost_net, poll[VHOST_NET_VQ_RX].work);
325 handle_rx(net);
326}
327
328static int vhost_net_open(struct inode *inode, struct file *f)
329{
330 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
331 int r;
332 if (!n)
333 return -ENOMEM;
334 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
335 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
336 r = vhost_dev_init(&n->dev, n->vqs, VHOST_NET_VQ_MAX);
337 if (r < 0) {
338 kfree(n);
339 return r;
340 }
341
342 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT);
343 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN);
344 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
345
346 f->private_data = n;
347
348 return 0;
349}
350
351static void vhost_net_disable_vq(struct vhost_net *n,
352 struct vhost_virtqueue *vq)
353{
354 if (!vq->private_data)
355 return;
356 if (vq == n->vqs + VHOST_NET_VQ_TX) {
357 tx_poll_stop(n);
358 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
359 } else
360 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
361}
362
363static void vhost_net_enable_vq(struct vhost_net *n,
364 struct vhost_virtqueue *vq)
365{
366 struct socket *sock = vq->private_data;
367 if (!sock)
368 return;
369 if (vq == n->vqs + VHOST_NET_VQ_TX) {
370 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
371 tx_poll_start(n, sock);
372 } else
373 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
374}
375
376static struct socket *vhost_net_stop_vq(struct vhost_net *n,
377 struct vhost_virtqueue *vq)
378{
379 struct socket *sock;
380
381 mutex_lock(&vq->mutex);
382 sock = vq->private_data;
383 vhost_net_disable_vq(n, vq);
384 rcu_assign_pointer(vq->private_data, NULL);
385 mutex_unlock(&vq->mutex);
386 return sock;
387}
388
389static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
390 struct socket **rx_sock)
391{
392 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
393 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
394}
395
396static void vhost_net_flush_vq(struct vhost_net *n, int index)
397{
398 vhost_poll_flush(n->poll + index);
399 vhost_poll_flush(&n->dev.vqs[index].poll);
400}
401
402static void vhost_net_flush(struct vhost_net *n)
403{
404 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
405 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
406}
407
408static int vhost_net_release(struct inode *inode, struct file *f)
409{
410 struct vhost_net *n = f->private_data;
411 struct socket *tx_sock;
412 struct socket *rx_sock;
413
414 vhost_net_stop(n, &tx_sock, &rx_sock);
415 vhost_net_flush(n);
416 vhost_dev_cleanup(&n->dev);
417 if (tx_sock)
418 fput(tx_sock->file);
419 if (rx_sock)
420 fput(rx_sock->file);
421 /* We do an extra flush before freeing memory,
422 * since jobs can re-queue themselves. */
423 vhost_net_flush(n);
424 kfree(n);
425 return 0;
426}
427
428static struct socket *get_raw_socket(int fd)
429{
430 struct {
431 struct sockaddr_ll sa;
432 char buf[MAX_ADDR_LEN];
433 } uaddr;
434 int uaddr_len = sizeof uaddr, r;
435 struct socket *sock = sockfd_lookup(fd, &r);
436 if (!sock)
437 return ERR_PTR(-ENOTSOCK);
438
439 /* Parameter checking */
440 if (sock->sk->sk_type != SOCK_RAW) {
441 r = -ESOCKTNOSUPPORT;
442 goto err;
443 }
444
445 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
446 &uaddr_len, 0);
447 if (r)
448 goto err;
449
450 if (uaddr.sa.sll_family != AF_PACKET) {
451 r = -EPFNOSUPPORT;
452 goto err;
453 }
454 return sock;
455err:
456 fput(sock->file);
457 return ERR_PTR(r);
458}
459
Arnd Bergmann501c7742010-02-18 05:46:50 +0000460static struct socket *get_tap_socket(int fd)
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +0000461{
462 struct file *file = fget(fd);
463 struct socket *sock;
464 if (!file)
465 return ERR_PTR(-EBADF);
466 sock = tun_get_socket(file);
Arnd Bergmann501c7742010-02-18 05:46:50 +0000467 if (!IS_ERR(sock))
468 return sock;
469 sock = macvtap_get_socket(file);
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +0000470 if (IS_ERR(sock))
471 fput(file);
472 return sock;
473}
474
475static struct socket *get_socket(int fd)
476{
477 struct socket *sock;
478 /* special case to disable backend */
479 if (fd == -1)
480 return NULL;
481 sock = get_raw_socket(fd);
482 if (!IS_ERR(sock))
483 return sock;
Arnd Bergmann501c7742010-02-18 05:46:50 +0000484 sock = get_tap_socket(fd);
Michael S. Tsirkin3a4d5c92010-01-14 06:17:27 +0000485 if (!IS_ERR(sock))
486 return sock;
487 return ERR_PTR(-ENOTSOCK);
488}
489
490static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
491{
492 struct socket *sock, *oldsock;
493 struct vhost_virtqueue *vq;
494 int r;
495
496 mutex_lock(&n->dev.mutex);
497 r = vhost_dev_check_owner(&n->dev);
498 if (r)
499 goto err;
500
501 if (index >= VHOST_NET_VQ_MAX) {
502 r = -ENOBUFS;
503 goto err;
504 }
505 vq = n->vqs + index;
506 mutex_lock(&vq->mutex);
507
508 /* Verify that ring has been setup correctly. */
509 if (!vhost_vq_access_ok(vq)) {
510 r = -EFAULT;
511 goto err;
512 }
513 sock = get_socket(fd);
514 if (IS_ERR(sock)) {
515 r = PTR_ERR(sock);
516 goto err;
517 }
518
519 /* start polling new socket */
520 oldsock = vq->private_data;
521 if (sock == oldsock)
522 goto done;
523
524 vhost_net_disable_vq(n, vq);
525 rcu_assign_pointer(vq->private_data, sock);
526 vhost_net_enable_vq(n, vq);
527 mutex_unlock(&vq->mutex);
528done:
529 if (oldsock) {
530 vhost_net_flush_vq(n, index);
531 fput(oldsock->file);
532 }
533err:
534 mutex_unlock(&n->dev.mutex);
535 return r;
536}
537
538static long vhost_net_reset_owner(struct vhost_net *n)
539{
540 struct socket *tx_sock = NULL;
541 struct socket *rx_sock = NULL;
542 long err;
543 mutex_lock(&n->dev.mutex);
544 err = vhost_dev_check_owner(&n->dev);
545 if (err)
546 goto done;
547 vhost_net_stop(n, &tx_sock, &rx_sock);
548 vhost_net_flush(n);
549 err = vhost_dev_reset_owner(&n->dev);
550done:
551 mutex_unlock(&n->dev.mutex);
552 if (tx_sock)
553 fput(tx_sock->file);
554 if (rx_sock)
555 fput(rx_sock->file);
556 return err;
557}
558
559static int vhost_net_set_features(struct vhost_net *n, u64 features)
560{
561 size_t hdr_size = features & (1 << VHOST_NET_F_VIRTIO_NET_HDR) ?
562 sizeof(struct virtio_net_hdr) : 0;
563 int i;
564 mutex_lock(&n->dev.mutex);
565 if ((features & (1 << VHOST_F_LOG_ALL)) &&
566 !vhost_log_access_ok(&n->dev)) {
567 mutex_unlock(&n->dev.mutex);
568 return -EFAULT;
569 }
570 n->dev.acked_features = features;
571 smp_wmb();
572 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
573 mutex_lock(&n->vqs[i].mutex);
574 n->vqs[i].hdr_size = hdr_size;
575 mutex_unlock(&n->vqs[i].mutex);
576 }
577 vhost_net_flush(n);
578 mutex_unlock(&n->dev.mutex);
579 return 0;
580}
581
582static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
583 unsigned long arg)
584{
585 struct vhost_net *n = f->private_data;
586 void __user *argp = (void __user *)arg;
587 u64 __user *featurep = argp;
588 struct vhost_vring_file backend;
589 u64 features;
590 int r;
591 switch (ioctl) {
592 case VHOST_NET_SET_BACKEND:
593 r = copy_from_user(&backend, argp, sizeof backend);
594 if (r < 0)
595 return r;
596 return vhost_net_set_backend(n, backend.index, backend.fd);
597 case VHOST_GET_FEATURES:
598 features = VHOST_FEATURES;
599 return copy_to_user(featurep, &features, sizeof features);
600 case VHOST_SET_FEATURES:
601 r = copy_from_user(&features, featurep, sizeof features);
602 if (r < 0)
603 return r;
604 if (features & ~VHOST_FEATURES)
605 return -EOPNOTSUPP;
606 return vhost_net_set_features(n, features);
607 case VHOST_RESET_OWNER:
608 return vhost_net_reset_owner(n);
609 default:
610 mutex_lock(&n->dev.mutex);
611 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
612 vhost_net_flush(n);
613 mutex_unlock(&n->dev.mutex);
614 return r;
615 }
616}
617
618#ifdef CONFIG_COMPAT
619static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
620 unsigned long arg)
621{
622 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
623}
624#endif
625
626const static struct file_operations vhost_net_fops = {
627 .owner = THIS_MODULE,
628 .release = vhost_net_release,
629 .unlocked_ioctl = vhost_net_ioctl,
630#ifdef CONFIG_COMPAT
631 .compat_ioctl = vhost_net_compat_ioctl,
632#endif
633 .open = vhost_net_open,
634};
635
636static struct miscdevice vhost_net_misc = {
637 VHOST_NET_MINOR,
638 "vhost-net",
639 &vhost_net_fops,
640};
641
642int vhost_net_init(void)
643{
644 int r = vhost_init();
645 if (r)
646 goto err_init;
647 r = misc_register(&vhost_net_misc);
648 if (r)
649 goto err_reg;
650 return 0;
651err_reg:
652 vhost_cleanup();
653err_init:
654 return r;
655
656}
657module_init(vhost_net_init);
658
659void vhost_net_exit(void)
660{
661 misc_deregister(&vhost_net_misc);
662 vhost_cleanup();
663}
664module_exit(vhost_net_exit);
665
666MODULE_VERSION("0.0.1");
667MODULE_LICENSE("GPL v2");
668MODULE_AUTHOR("Michael S. Tsirkin");
669MODULE_DESCRIPTION("Host kernel accelerator for virtio net");