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 #include <linux/if_vlan.h>
28
29 #include <net/sock.h>
30
31 #include "vhost.h"
32
33 static int experimental_zcopytx;
34 module_param(experimental_zcopytx, int, 0444);
35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
36
37 /* Max number of bytes transferred before requeueing the job.
38 * Using this limit prevents one virtqueue from starving others. */
39 #define VHOST_NET_WEIGHT 0x80000
40
41 /* MAX number of TX used buffers for outstanding zerocopy */
42 #define VHOST_MAX_PEND 128
43 #define VHOST_GOODCOPY_LEN 256
44
45 enum {
46 VHOST_NET_VQ_RX = 0,
47 VHOST_NET_VQ_TX = 1,
48 VHOST_NET_VQ_MAX = 2,
49 };
50
51 enum vhost_net_poll_state {
52 VHOST_NET_POLL_DISABLED = 0,
53 VHOST_NET_POLL_STARTED = 1,
54 VHOST_NET_POLL_STOPPED = 2,
55 };
56
57 struct vhost_net {
58 struct vhost_dev dev;
59 struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
60 struct vhost_poll poll[VHOST_NET_VQ_MAX];
61 /* Tells us whether we are polling a socket for TX.
62 * We only do this when socket buffer fills up.
63 * Protected by tx vq lock. */
64 enum vhost_net_poll_state tx_poll_state;
65 };
66
vhost_sock_zcopy(struct socket * sock)67 static bool vhost_sock_zcopy(struct socket *sock)
68 {
69 return unlikely(experimental_zcopytx) &&
70 sock_flag(sock->sk, SOCK_ZEROCOPY);
71 }
72
73 /* Pop first len bytes from iovec. Return number of segments used. */
move_iovec_hdr(struct iovec * from,struct iovec * to,size_t len,int iov_count)74 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
75 size_t len, int iov_count)
76 {
77 int seg = 0;
78 size_t size;
79
80 while (len && seg < iov_count) {
81 size = min(from->iov_len, len);
82 to->iov_base = from->iov_base;
83 to->iov_len = size;
84 from->iov_len -= size;
85 from->iov_base += size;
86 len -= size;
87 ++from;
88 ++to;
89 ++seg;
90 }
91 return seg;
92 }
93 /* Copy iovec entries for len bytes from iovec. */
copy_iovec_hdr(const struct iovec * from,struct iovec * to,size_t len,int iovcount)94 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
95 size_t len, int iovcount)
96 {
97 int seg = 0;
98 size_t size;
99
100 while (len && seg < iovcount) {
101 size = min(from->iov_len, len);
102 to->iov_base = from->iov_base;
103 to->iov_len = size;
104 len -= size;
105 ++from;
106 ++to;
107 ++seg;
108 }
109 }
110
111 /* Caller must have TX VQ lock */
tx_poll_stop(struct vhost_net * net)112 static void tx_poll_stop(struct vhost_net *net)
113 {
114 if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
115 return;
116 vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
117 net->tx_poll_state = VHOST_NET_POLL_STOPPED;
118 }
119
120 /* Caller must have TX VQ lock */
tx_poll_start(struct vhost_net * net,struct socket * sock)121 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
122 {
123 if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
124 return;
125 vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
126 net->tx_poll_state = VHOST_NET_POLL_STARTED;
127 }
128
129 /* Expects to be always run from workqueue - which acts as
130 * read-size critical section for our kind of RCU. */
handle_tx(struct vhost_net * net)131 static void handle_tx(struct vhost_net *net)
132 {
133 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
134 unsigned out, in, s;
135 int head;
136 struct msghdr msg = {
137 .msg_name = NULL,
138 .msg_namelen = 0,
139 .msg_control = NULL,
140 .msg_controllen = 0,
141 .msg_iov = vq->iov,
142 .msg_flags = MSG_DONTWAIT,
143 };
144 size_t len, total_len = 0;
145 int err, wmem;
146 size_t hdr_size;
147 struct socket *sock;
148 struct vhost_ubuf_ref *uninitialized_var(ubufs);
149 bool zcopy;
150
151 /* TODO: check that we are running from vhost_worker? */
152 sock = rcu_dereference_check(vq->private_data, 1);
153 if (!sock)
154 return;
155
156 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
157 if (wmem >= sock->sk->sk_sndbuf) {
158 mutex_lock(&vq->mutex);
159 tx_poll_start(net, sock);
160 mutex_unlock(&vq->mutex);
161 return;
162 }
163
164 mutex_lock(&vq->mutex);
165 vhost_disable_notify(&net->dev, vq);
166
167 if (wmem < sock->sk->sk_sndbuf / 2)
168 tx_poll_stop(net);
169 hdr_size = vq->vhost_hlen;
170 zcopy = vhost_sock_zcopy(sock);
171
172 for (;;) {
173 /* Release DMAs done buffers first */
174 if (zcopy)
175 vhost_zerocopy_signal_used(vq);
176
177 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
178 ARRAY_SIZE(vq->iov),
179 &out, &in,
180 NULL, NULL);
181 /* On error, stop handling until the next kick. */
182 if (unlikely(head < 0))
183 break;
184 /* Nothing new? Wait for eventfd to tell us they refilled. */
185 if (head == vq->num) {
186 int num_pends;
187
188 wmem = atomic_read(&sock->sk->sk_wmem_alloc);
189 if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
190 tx_poll_start(net, sock);
191 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
192 break;
193 }
194 /* If more outstanding DMAs, queue the work.
195 * Handle upend_idx wrap around
196 */
197 num_pends = likely(vq->upend_idx >= vq->done_idx) ?
198 (vq->upend_idx - vq->done_idx) :
199 (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
200 if (unlikely(num_pends > VHOST_MAX_PEND)) {
201 tx_poll_start(net, sock);
202 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
203 break;
204 }
205 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
206 vhost_disable_notify(&net->dev, vq);
207 continue;
208 }
209 break;
210 }
211 if (in) {
212 vq_err(vq, "Unexpected descriptor format for TX: "
213 "out %d, int %d\n", out, in);
214 break;
215 }
216 /* Skip header. TODO: support TSO. */
217 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
218 msg.msg_iovlen = out;
219 len = iov_length(vq->iov, out);
220 /* Sanity check */
221 if (!len) {
222 vq_err(vq, "Unexpected header len for TX: "
223 "%zd expected %zd\n",
224 iov_length(vq->hdr, s), hdr_size);
225 break;
226 }
227 /* use msg_control to pass vhost zerocopy ubuf info to skb */
228 if (zcopy) {
229 vq->heads[vq->upend_idx].id = head;
230 if (len < VHOST_GOODCOPY_LEN) {
231 /* copy don't need to wait for DMA done */
232 vq->heads[vq->upend_idx].len =
233 VHOST_DMA_DONE_LEN;
234 msg.msg_control = NULL;
235 msg.msg_controllen = 0;
236 ubufs = NULL;
237 } else {
238 struct ubuf_info *ubuf;
239 ubuf = vq->ubuf_info + vq->upend_idx;
240
241 vq->heads[vq->upend_idx].len = len;
242 ubuf->callback = vhost_zerocopy_callback;
243 ubuf->ctx = vq->ubufs;
244 ubuf->desc = vq->upend_idx;
245 msg.msg_control = ubuf;
246 msg.msg_controllen = sizeof(ubuf);
247 ubufs = vq->ubufs;
248 kref_get(&ubufs->kref);
249 }
250 vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
251 }
252 /* TODO: Check specific error and bomb out unless ENOBUFS? */
253 err = sock->ops->sendmsg(NULL, sock, &msg, len);
254 if (unlikely(err < 0)) {
255 if (zcopy) {
256 if (ubufs)
257 vhost_ubuf_put(ubufs);
258 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
259 UIO_MAXIOV;
260 }
261 vhost_discard_vq_desc(vq, 1);
262 tx_poll_start(net, sock);
263 break;
264 }
265 if (err != len)
266 pr_debug("Truncated TX packet: "
267 " len %d != %zd\n", err, len);
268 if (!zcopy)
269 vhost_add_used_and_signal(&net->dev, vq, head, 0);
270 total_len += len;
271 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
272 vhost_poll_queue(&vq->poll);
273 break;
274 }
275 }
276
277 mutex_unlock(&vq->mutex);
278 }
279
peek_head_len(struct sock * sk)280 static int peek_head_len(struct sock *sk)
281 {
282 struct sk_buff *head;
283 int len = 0;
284 unsigned long flags;
285
286 spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
287 head = skb_peek(&sk->sk_receive_queue);
288 if (likely(head)) {
289 len = head->len;
290 if (vlan_tx_tag_present(head))
291 len += VLAN_HLEN;
292 }
293
294 spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
295 return len;
296 }
297
298 /* This is a multi-buffer version of vhost_get_desc, that works if
299 * vq has read descriptors only.
300 * @vq - the relevant virtqueue
301 * @datalen - data length we'll be reading
302 * @iovcount - returned count of io vectors we fill
303 * @log - vhost log
304 * @log_num - log offset
305 * @quota - headcount quota, 1 for big buffer
306 * returns number of buffer heads allocated, negative on error
307 */
get_rx_bufs(struct vhost_virtqueue * vq,struct vring_used_elem * heads,int datalen,unsigned * iovcount,struct vhost_log * log,unsigned * log_num,unsigned int quota)308 static int get_rx_bufs(struct vhost_virtqueue *vq,
309 struct vring_used_elem *heads,
310 int datalen,
311 unsigned *iovcount,
312 struct vhost_log *log,
313 unsigned *log_num,
314 unsigned int quota)
315 {
316 unsigned int out, in;
317 int seg = 0;
318 int headcount = 0;
319 unsigned d;
320 int r, nlogs = 0;
321
322 while (datalen > 0 && headcount < quota) {
323 if (unlikely(seg >= UIO_MAXIOV)) {
324 r = -ENOBUFS;
325 goto err;
326 }
327 r = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
328 ARRAY_SIZE(vq->iov) - seg, &out,
329 &in, log, log_num);
330 if (unlikely(r < 0))
331 goto err;
332
333 d = r;
334 if (d == vq->num) {
335 r = 0;
336 goto err;
337 }
338 if (unlikely(out || in <= 0)) {
339 vq_err(vq, "unexpected descriptor format for RX: "
340 "out %d, in %d\n", out, in);
341 r = -EINVAL;
342 goto err;
343 }
344 if (unlikely(log)) {
345 nlogs += *log_num;
346 log += *log_num;
347 }
348 heads[headcount].id = d;
349 heads[headcount].len = iov_length(vq->iov + seg, in);
350 datalen -= heads[headcount].len;
351 ++headcount;
352 seg += in;
353 }
354 heads[headcount - 1].len += datalen;
355 *iovcount = seg;
356 if (unlikely(log))
357 *log_num = nlogs;
358
359 /* Detect overrun */
360 if (unlikely(datalen > 0)) {
361 r = UIO_MAXIOV + 1;
362 goto err;
363 }
364 return headcount;
365 err:
366 vhost_discard_vq_desc(vq, headcount);
367 return r;
368 }
369
370 /* Expects to be always run from workqueue - which acts as
371 * read-size critical section for our kind of RCU. */
handle_rx(struct vhost_net * net)372 static void handle_rx(struct vhost_net *net)
373 {
374 struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
375 unsigned uninitialized_var(in), log;
376 struct vhost_log *vq_log;
377 struct msghdr msg = {
378 .msg_name = NULL,
379 .msg_namelen = 0,
380 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
381 .msg_controllen = 0,
382 .msg_iov = vq->iov,
383 .msg_flags = MSG_DONTWAIT,
384 };
385 struct virtio_net_hdr_mrg_rxbuf hdr = {
386 .hdr.flags = 0,
387 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
388 };
389 size_t total_len = 0;
390 int err, mergeable;
391 s16 headcount;
392 size_t vhost_hlen, sock_hlen;
393 size_t vhost_len, sock_len;
394 /* TODO: check that we are running from vhost_worker? */
395 struct socket *sock = rcu_dereference_check(vq->private_data, 1);
396
397 if (!sock)
398 return;
399
400 mutex_lock(&vq->mutex);
401 vhost_disable_notify(&net->dev, vq);
402 vhost_hlen = vq->vhost_hlen;
403 sock_hlen = vq->sock_hlen;
404
405 vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
406 vq->log : NULL;
407 mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
408
409 while ((sock_len = peek_head_len(sock->sk))) {
410 sock_len += sock_hlen;
411 vhost_len = sock_len + vhost_hlen;
412 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
413 &in, vq_log, &log,
414 likely(mergeable) ? UIO_MAXIOV : 1);
415 /* On error, stop handling until the next kick. */
416 if (unlikely(headcount < 0))
417 break;
418 /* On overrun, truncate and discard */
419 if (unlikely(headcount > UIO_MAXIOV)) {
420 msg.msg_iovlen = 1;
421 err = sock->ops->recvmsg(NULL, sock, &msg,
422 1, MSG_DONTWAIT | MSG_TRUNC);
423 pr_debug("Discarded rx packet: len %zd\n", sock_len);
424 continue;
425 }
426 /* OK, now we need to know about added descriptors. */
427 if (!headcount) {
428 if (unlikely(vhost_enable_notify(&net->dev, vq))) {
429 /* They have slipped one in as we were
430 * doing that: check again. */
431 vhost_disable_notify(&net->dev, vq);
432 continue;
433 }
434 /* Nothing new? Wait for eventfd to tell us
435 * they refilled. */
436 break;
437 }
438 /* We don't need to be notified again. */
439 if (unlikely((vhost_hlen)))
440 /* Skip header. TODO: support TSO. */
441 move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
442 else
443 /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
444 * needed because recvmsg can modify msg_iov. */
445 copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
446 msg.msg_iovlen = in;
447 err = sock->ops->recvmsg(NULL, sock, &msg,
448 sock_len, MSG_DONTWAIT | MSG_TRUNC);
449 /* Userspace might have consumed the packet meanwhile:
450 * it's not supposed to do this usually, but might be hard
451 * to prevent. Discard data we got (if any) and keep going. */
452 if (unlikely(err != sock_len)) {
453 pr_debug("Discarded rx packet: "
454 " len %d, expected %zd\n", err, sock_len);
455 vhost_discard_vq_desc(vq, headcount);
456 continue;
457 }
458 if (unlikely(vhost_hlen) &&
459 memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
460 vhost_hlen)) {
461 vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
462 vq->iov->iov_base);
463 break;
464 }
465 /* TODO: Should check and handle checksum. */
466 if (likely(mergeable) &&
467 memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
468 offsetof(typeof(hdr), num_buffers),
469 sizeof hdr.num_buffers)) {
470 vq_err(vq, "Failed num_buffers write");
471 vhost_discard_vq_desc(vq, headcount);
472 break;
473 }
474 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
475 headcount);
476 if (unlikely(vq_log))
477 vhost_log_write(vq, vq_log, log, vhost_len);
478 total_len += vhost_len;
479 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
480 vhost_poll_queue(&vq->poll);
481 break;
482 }
483 }
484
485 mutex_unlock(&vq->mutex);
486 }
487
handle_tx_kick(struct vhost_work * work)488 static void handle_tx_kick(struct vhost_work *work)
489 {
490 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
491 poll.work);
492 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
493
494 handle_tx(net);
495 }
496
handle_rx_kick(struct vhost_work * work)497 static void handle_rx_kick(struct vhost_work *work)
498 {
499 struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
500 poll.work);
501 struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
502
503 handle_rx(net);
504 }
505
handle_tx_net(struct vhost_work * work)506 static void handle_tx_net(struct vhost_work *work)
507 {
508 struct vhost_net *net = container_of(work, struct vhost_net,
509 poll[VHOST_NET_VQ_TX].work);
510 handle_tx(net);
511 }
512
handle_rx_net(struct vhost_work * work)513 static void handle_rx_net(struct vhost_work *work)
514 {
515 struct vhost_net *net = container_of(work, struct vhost_net,
516 poll[VHOST_NET_VQ_RX].work);
517 handle_rx(net);
518 }
519
vhost_net_open(struct inode * inode,struct file * f)520 static int vhost_net_open(struct inode *inode, struct file *f)
521 {
522 struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
523 struct vhost_dev *dev;
524 int r;
525
526 if (!n)
527 return -ENOMEM;
528
529 dev = &n->dev;
530 n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
531 n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
532 r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
533 if (r < 0) {
534 kfree(n);
535 return r;
536 }
537
538 vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
539 vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
540 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
541
542 f->private_data = n;
543
544 return 0;
545 }
546
vhost_net_disable_vq(struct vhost_net * n,struct vhost_virtqueue * vq)547 static void vhost_net_disable_vq(struct vhost_net *n,
548 struct vhost_virtqueue *vq)
549 {
550 if (!vq->private_data)
551 return;
552 if (vq == n->vqs + VHOST_NET_VQ_TX) {
553 tx_poll_stop(n);
554 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
555 } else
556 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
557 }
558
vhost_net_enable_vq(struct vhost_net * n,struct vhost_virtqueue * vq)559 static void vhost_net_enable_vq(struct vhost_net *n,
560 struct vhost_virtqueue *vq)
561 {
562 struct socket *sock;
563
564 sock = rcu_dereference_protected(vq->private_data,
565 lockdep_is_held(&vq->mutex));
566 if (!sock)
567 return;
568 if (vq == n->vqs + VHOST_NET_VQ_TX) {
569 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
570 tx_poll_start(n, sock);
571 } else
572 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
573 }
574
vhost_net_stop_vq(struct vhost_net * n,struct vhost_virtqueue * vq)575 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
576 struct vhost_virtqueue *vq)
577 {
578 struct socket *sock;
579
580 mutex_lock(&vq->mutex);
581 sock = rcu_dereference_protected(vq->private_data,
582 lockdep_is_held(&vq->mutex));
583 vhost_net_disable_vq(n, vq);
584 rcu_assign_pointer(vq->private_data, NULL);
585 mutex_unlock(&vq->mutex);
586 return sock;
587 }
588
vhost_net_stop(struct vhost_net * n,struct socket ** tx_sock,struct socket ** rx_sock)589 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
590 struct socket **rx_sock)
591 {
592 *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
593 *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
594 }
595
vhost_net_flush_vq(struct vhost_net * n,int index)596 static void vhost_net_flush_vq(struct vhost_net *n, int index)
597 {
598 vhost_poll_flush(n->poll + index);
599 vhost_poll_flush(&n->dev.vqs[index].poll);
600 }
601
vhost_net_flush(struct vhost_net * n)602 static void vhost_net_flush(struct vhost_net *n)
603 {
604 vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
605 vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
606 }
607
vhost_net_release(struct inode * inode,struct file * f)608 static int vhost_net_release(struct inode *inode, struct file *f)
609 {
610 struct vhost_net *n = f->private_data;
611 struct socket *tx_sock;
612 struct socket *rx_sock;
613
614 vhost_net_stop(n, &tx_sock, &rx_sock);
615 vhost_net_flush(n);
616 vhost_dev_cleanup(&n->dev, false);
617 if (tx_sock)
618 fput(tx_sock->file);
619 if (rx_sock)
620 fput(rx_sock->file);
621 /* We do an extra flush before freeing memory,
622 * since jobs can re-queue themselves. */
623 vhost_net_flush(n);
624 kfree(n);
625 return 0;
626 }
627
get_raw_socket(int fd)628 static struct socket *get_raw_socket(int fd)
629 {
630 struct {
631 struct sockaddr_ll sa;
632 char buf[MAX_ADDR_LEN];
633 } uaddr;
634 int uaddr_len = sizeof uaddr, r;
635 struct socket *sock = sockfd_lookup(fd, &r);
636
637 if (!sock)
638 return ERR_PTR(-ENOTSOCK);
639
640 /* Parameter checking */
641 if (sock->sk->sk_type != SOCK_RAW) {
642 r = -ESOCKTNOSUPPORT;
643 goto err;
644 }
645
646 r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
647 &uaddr_len, 0);
648 if (r)
649 goto err;
650
651 if (uaddr.sa.sll_family != AF_PACKET) {
652 r = -EPFNOSUPPORT;
653 goto err;
654 }
655 return sock;
656 err:
657 fput(sock->file);
658 return ERR_PTR(r);
659 }
660
get_tap_socket(int fd)661 static struct socket *get_tap_socket(int fd)
662 {
663 struct file *file = fget(fd);
664 struct socket *sock;
665
666 if (!file)
667 return ERR_PTR(-EBADF);
668 sock = tun_get_socket(file);
669 if (!IS_ERR(sock))
670 return sock;
671 sock = macvtap_get_socket(file);
672 if (IS_ERR(sock))
673 fput(file);
674 return sock;
675 }
676
get_socket(int fd)677 static struct socket *get_socket(int fd)
678 {
679 struct socket *sock;
680
681 /* special case to disable backend */
682 if (fd == -1)
683 return NULL;
684 sock = get_raw_socket(fd);
685 if (!IS_ERR(sock))
686 return sock;
687 sock = get_tap_socket(fd);
688 if (!IS_ERR(sock))
689 return sock;
690 return ERR_PTR(-ENOTSOCK);
691 }
692
vhost_net_set_backend(struct vhost_net * n,unsigned index,int fd)693 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
694 {
695 struct socket *sock, *oldsock;
696 struct vhost_virtqueue *vq;
697 struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
698 int r;
699
700 mutex_lock(&n->dev.mutex);
701 r = vhost_dev_check_owner(&n->dev);
702 if (r)
703 goto err;
704
705 if (index >= VHOST_NET_VQ_MAX) {
706 r = -ENOBUFS;
707 goto err;
708 }
709 vq = n->vqs + index;
710 mutex_lock(&vq->mutex);
711
712 /* Verify that ring has been setup correctly. */
713 if (!vhost_vq_access_ok(vq)) {
714 r = -EFAULT;
715 goto err_vq;
716 }
717 sock = get_socket(fd);
718 if (IS_ERR(sock)) {
719 r = PTR_ERR(sock);
720 goto err_vq;
721 }
722
723 /* start polling new socket */
724 oldsock = rcu_dereference_protected(vq->private_data,
725 lockdep_is_held(&vq->mutex));
726 if (sock != oldsock) {
727 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
728 if (IS_ERR(ubufs)) {
729 r = PTR_ERR(ubufs);
730 goto err_ubufs;
731 }
732 oldubufs = vq->ubufs;
733 vq->ubufs = ubufs;
734 vhost_net_disable_vq(n, vq);
735 rcu_assign_pointer(vq->private_data, sock);
736 vhost_net_enable_vq(n, vq);
737
738 r = vhost_init_used(vq);
739 if (r)
740 goto err_vq;
741 }
742
743 mutex_unlock(&vq->mutex);
744
745 if (oldubufs) {
746 vhost_ubuf_put_and_wait(oldubufs);
747 mutex_lock(&vq->mutex);
748 vhost_zerocopy_signal_used(vq);
749 mutex_unlock(&vq->mutex);
750 }
751
752 if (oldsock) {
753 vhost_net_flush_vq(n, index);
754 fput(oldsock->file);
755 }
756
757 mutex_unlock(&n->dev.mutex);
758 return 0;
759
760 err_ubufs:
761 fput(sock->file);
762 err_vq:
763 mutex_unlock(&vq->mutex);
764 err:
765 mutex_unlock(&n->dev.mutex);
766 return r;
767 }
768
vhost_net_reset_owner(struct vhost_net * n)769 static long vhost_net_reset_owner(struct vhost_net *n)
770 {
771 struct socket *tx_sock = NULL;
772 struct socket *rx_sock = NULL;
773 long err;
774
775 mutex_lock(&n->dev.mutex);
776 err = vhost_dev_check_owner(&n->dev);
777 if (err)
778 goto done;
779 vhost_net_stop(n, &tx_sock, &rx_sock);
780 vhost_net_flush(n);
781 err = vhost_dev_reset_owner(&n->dev);
782 done:
783 mutex_unlock(&n->dev.mutex);
784 if (tx_sock)
785 fput(tx_sock->file);
786 if (rx_sock)
787 fput(rx_sock->file);
788 return err;
789 }
790
vhost_net_set_features(struct vhost_net * n,u64 features)791 static int vhost_net_set_features(struct vhost_net *n, u64 features)
792 {
793 size_t vhost_hlen, sock_hlen, hdr_len;
794 int i;
795
796 hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
797 sizeof(struct virtio_net_hdr_mrg_rxbuf) :
798 sizeof(struct virtio_net_hdr);
799 if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
800 /* vhost provides vnet_hdr */
801 vhost_hlen = hdr_len;
802 sock_hlen = 0;
803 } else {
804 /* socket provides vnet_hdr */
805 vhost_hlen = 0;
806 sock_hlen = hdr_len;
807 }
808 mutex_lock(&n->dev.mutex);
809 if ((features & (1 << VHOST_F_LOG_ALL)) &&
810 !vhost_log_access_ok(&n->dev)) {
811 mutex_unlock(&n->dev.mutex);
812 return -EFAULT;
813 }
814 n->dev.acked_features = features;
815 smp_wmb();
816 for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
817 mutex_lock(&n->vqs[i].mutex);
818 n->vqs[i].vhost_hlen = vhost_hlen;
819 n->vqs[i].sock_hlen = sock_hlen;
820 mutex_unlock(&n->vqs[i].mutex);
821 }
822 vhost_net_flush(n);
823 mutex_unlock(&n->dev.mutex);
824 return 0;
825 }
826
vhost_net_ioctl(struct file * f,unsigned int ioctl,unsigned long arg)827 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
828 unsigned long arg)
829 {
830 struct vhost_net *n = f->private_data;
831 void __user *argp = (void __user *)arg;
832 u64 __user *featurep = argp;
833 struct vhost_vring_file backend;
834 u64 features;
835 int r;
836
837 switch (ioctl) {
838 case VHOST_NET_SET_BACKEND:
839 if (copy_from_user(&backend, argp, sizeof backend))
840 return -EFAULT;
841 return vhost_net_set_backend(n, backend.index, backend.fd);
842 case VHOST_GET_FEATURES:
843 features = VHOST_FEATURES;
844 if (copy_to_user(featurep, &features, sizeof features))
845 return -EFAULT;
846 return 0;
847 case VHOST_SET_FEATURES:
848 if (copy_from_user(&features, featurep, sizeof features))
849 return -EFAULT;
850 if (features & ~VHOST_FEATURES)
851 return -EOPNOTSUPP;
852 return vhost_net_set_features(n, features);
853 case VHOST_RESET_OWNER:
854 return vhost_net_reset_owner(n);
855 default:
856 mutex_lock(&n->dev.mutex);
857 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
858 vhost_net_flush(n);
859 mutex_unlock(&n->dev.mutex);
860 return r;
861 }
862 }
863
864 #ifdef CONFIG_COMPAT
vhost_net_compat_ioctl(struct file * f,unsigned int ioctl,unsigned long arg)865 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
866 unsigned long arg)
867 {
868 return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
869 }
870 #endif
871
872 static const struct file_operations vhost_net_fops = {
873 .owner = THIS_MODULE,
874 .release = vhost_net_release,
875 .unlocked_ioctl = vhost_net_ioctl,
876 #ifdef CONFIG_COMPAT
877 .compat_ioctl = vhost_net_compat_ioctl,
878 #endif
879 .open = vhost_net_open,
880 .llseek = noop_llseek,
881 };
882
883 static struct miscdevice vhost_net_misc = {
884 .minor = VHOST_NET_MINOR,
885 .name = "vhost-net",
886 .fops = &vhost_net_fops,
887 };
888
vhost_net_init(void)889 static int vhost_net_init(void)
890 {
891 if (experimental_zcopytx)
892 vhost_enable_zcopy(VHOST_NET_VQ_TX);
893 return misc_register(&vhost_net_misc);
894 }
895 module_init(vhost_net_init);
896
vhost_net_exit(void)897 static void vhost_net_exit(void)
898 {
899 misc_deregister(&vhost_net_misc);
900 }
901 module_exit(vhost_net_exit);
902
903 MODULE_VERSION("0.0.1");
904 MODULE_LICENSE("GPL v2");
905 MODULE_AUTHOR("Michael S. Tsirkin");
906 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
907 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
908 MODULE_ALIAS("devname:vhost-net");
909