Revert "ARM: tegra: tegratab: dummy change"
[linux-2.6.git] / drivers / vhost / net.c
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
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. */
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. */
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 */
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 */
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. */
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
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  */
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                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
328                                       ARRAY_SIZE(vq->iov) - seg, &out,
329                                       &in, log, log_num);
330                 if (d == vq->num) {
331                         r = 0;
332                         goto err;
333                 }
334                 if (unlikely(out || in <= 0)) {
335                         vq_err(vq, "unexpected descriptor format for RX: "
336                                 "out %d, in %d\n", out, in);
337                         r = -EINVAL;
338                         goto err;
339                 }
340                 if (unlikely(log)) {
341                         nlogs += *log_num;
342                         log += *log_num;
343                 }
344                 heads[headcount].id = d;
345                 heads[headcount].len = iov_length(vq->iov + seg, in);
346                 datalen -= heads[headcount].len;
347                 ++headcount;
348                 seg += in;
349         }
350         heads[headcount - 1].len += datalen;
351         *iovcount = seg;
352         if (unlikely(log))
353                 *log_num = nlogs;
354         return headcount;
355 err:
356         vhost_discard_vq_desc(vq, headcount);
357         return r;
358 }
359
360 /* Expects to be always run from workqueue - which acts as
361  * read-size critical section for our kind of RCU. */
362 static void handle_rx(struct vhost_net *net)
363 {
364         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
365         unsigned uninitialized_var(in), log;
366         struct vhost_log *vq_log;
367         struct msghdr msg = {
368                 .msg_name = NULL,
369                 .msg_namelen = 0,
370                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
371                 .msg_controllen = 0,
372                 .msg_iov = vq->iov,
373                 .msg_flags = MSG_DONTWAIT,
374         };
375         struct virtio_net_hdr_mrg_rxbuf hdr = {
376                 .hdr.flags = 0,
377                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
378         };
379         size_t total_len = 0;
380         int err, mergeable;
381         s16 headcount;
382         size_t vhost_hlen, sock_hlen;
383         size_t vhost_len, sock_len;
384         /* TODO: check that we are running from vhost_worker? */
385         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
386
387         if (!sock)
388                 return;
389
390         mutex_lock(&vq->mutex);
391         vhost_disable_notify(&net->dev, vq);
392         vhost_hlen = vq->vhost_hlen;
393         sock_hlen = vq->sock_hlen;
394
395         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
396                 vq->log : NULL;
397         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
398
399         while ((sock_len = peek_head_len(sock->sk))) {
400                 sock_len += sock_hlen;
401                 vhost_len = sock_len + vhost_hlen;
402                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
403                                         &in, vq_log, &log,
404                                         likely(mergeable) ? UIO_MAXIOV : 1);
405                 /* On error, stop handling until the next kick. */
406                 if (unlikely(headcount < 0))
407                         break;
408                 /* OK, now we need to know about added descriptors. */
409                 if (!headcount) {
410                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
411                                 /* They have slipped one in as we were
412                                  * doing that: check again. */
413                                 vhost_disable_notify(&net->dev, vq);
414                                 continue;
415                         }
416                         /* Nothing new?  Wait for eventfd to tell us
417                          * they refilled. */
418                         break;
419                 }
420                 /* We don't need to be notified again. */
421                 if (unlikely((vhost_hlen)))
422                         /* Skip header. TODO: support TSO. */
423                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
424                 else
425                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
426                          * needed because recvmsg can modify msg_iov. */
427                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
428                 msg.msg_iovlen = in;
429                 err = sock->ops->recvmsg(NULL, sock, &msg,
430                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
431                 /* Userspace might have consumed the packet meanwhile:
432                  * it's not supposed to do this usually, but might be hard
433                  * to prevent. Discard data we got (if any) and keep going. */
434                 if (unlikely(err != sock_len)) {
435                         pr_debug("Discarded rx packet: "
436                                  " len %d, expected %zd\n", err, sock_len);
437                         vhost_discard_vq_desc(vq, headcount);
438                         continue;
439                 }
440                 if (unlikely(vhost_hlen) &&
441                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
442                                       vhost_hlen)) {
443                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
444                                vq->iov->iov_base);
445                         break;
446                 }
447                 /* TODO: Should check and handle checksum. */
448                 if (likely(mergeable) &&
449                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
450                                       offsetof(typeof(hdr), num_buffers),
451                                       sizeof hdr.num_buffers)) {
452                         vq_err(vq, "Failed num_buffers write");
453                         vhost_discard_vq_desc(vq, headcount);
454                         break;
455                 }
456                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
457                                             headcount);
458                 if (unlikely(vq_log))
459                         vhost_log_write(vq, vq_log, log, vhost_len);
460                 total_len += vhost_len;
461                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
462                         vhost_poll_queue(&vq->poll);
463                         break;
464                 }
465         }
466
467         mutex_unlock(&vq->mutex);
468 }
469
470 static void handle_tx_kick(struct vhost_work *work)
471 {
472         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
473                                                   poll.work);
474         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
475
476         handle_tx(net);
477 }
478
479 static void handle_rx_kick(struct vhost_work *work)
480 {
481         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
482                                                   poll.work);
483         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
484
485         handle_rx(net);
486 }
487
488 static void handle_tx_net(struct vhost_work *work)
489 {
490         struct vhost_net *net = container_of(work, struct vhost_net,
491                                              poll[VHOST_NET_VQ_TX].work);
492         handle_tx(net);
493 }
494
495 static void handle_rx_net(struct vhost_work *work)
496 {
497         struct vhost_net *net = container_of(work, struct vhost_net,
498                                              poll[VHOST_NET_VQ_RX].work);
499         handle_rx(net);
500 }
501
502 static int vhost_net_open(struct inode *inode, struct file *f)
503 {
504         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
505         struct vhost_dev *dev;
506         int r;
507
508         if (!n)
509                 return -ENOMEM;
510
511         dev = &n->dev;
512         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
513         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
514         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
515         if (r < 0) {
516                 kfree(n);
517                 return r;
518         }
519
520         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
521         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
522         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
523
524         f->private_data = n;
525
526         return 0;
527 }
528
529 static void vhost_net_disable_vq(struct vhost_net *n,
530                                  struct vhost_virtqueue *vq)
531 {
532         if (!vq->private_data)
533                 return;
534         if (vq == n->vqs + VHOST_NET_VQ_TX) {
535                 tx_poll_stop(n);
536                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
537         } else
538                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
539 }
540
541 static void vhost_net_enable_vq(struct vhost_net *n,
542                                 struct vhost_virtqueue *vq)
543 {
544         struct socket *sock;
545
546         sock = rcu_dereference_protected(vq->private_data,
547                                          lockdep_is_held(&vq->mutex));
548         if (!sock)
549                 return;
550         if (vq == n->vqs + VHOST_NET_VQ_TX) {
551                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
552                 tx_poll_start(n, sock);
553         } else
554                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
555 }
556
557 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
558                                         struct vhost_virtqueue *vq)
559 {
560         struct socket *sock;
561
562         mutex_lock(&vq->mutex);
563         sock = rcu_dereference_protected(vq->private_data,
564                                          lockdep_is_held(&vq->mutex));
565         vhost_net_disable_vq(n, vq);
566         rcu_assign_pointer(vq->private_data, NULL);
567         mutex_unlock(&vq->mutex);
568         return sock;
569 }
570
571 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
572                            struct socket **rx_sock)
573 {
574         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
575         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
576 }
577
578 static void vhost_net_flush_vq(struct vhost_net *n, int index)
579 {
580         vhost_poll_flush(n->poll + index);
581         vhost_poll_flush(&n->dev.vqs[index].poll);
582 }
583
584 static void vhost_net_flush(struct vhost_net *n)
585 {
586         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
587         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
588 }
589
590 static int vhost_net_release(struct inode *inode, struct file *f)
591 {
592         struct vhost_net *n = f->private_data;
593         struct socket *tx_sock;
594         struct socket *rx_sock;
595
596         vhost_net_stop(n, &tx_sock, &rx_sock);
597         vhost_net_flush(n);
598         vhost_dev_cleanup(&n->dev, false);
599         if (tx_sock)
600                 fput(tx_sock->file);
601         if (rx_sock)
602                 fput(rx_sock->file);
603         /* We do an extra flush before freeing memory,
604          * since jobs can re-queue themselves. */
605         vhost_net_flush(n);
606         kfree(n);
607         return 0;
608 }
609
610 static struct socket *get_raw_socket(int fd)
611 {
612         struct {
613                 struct sockaddr_ll sa;
614                 char  buf[MAX_ADDR_LEN];
615         } uaddr;
616         int uaddr_len = sizeof uaddr, r;
617         struct socket *sock = sockfd_lookup(fd, &r);
618
619         if (!sock)
620                 return ERR_PTR(-ENOTSOCK);
621
622         /* Parameter checking */
623         if (sock->sk->sk_type != SOCK_RAW) {
624                 r = -ESOCKTNOSUPPORT;
625                 goto err;
626         }
627
628         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
629                                &uaddr_len, 0);
630         if (r)
631                 goto err;
632
633         if (uaddr.sa.sll_family != AF_PACKET) {
634                 r = -EPFNOSUPPORT;
635                 goto err;
636         }
637         return sock;
638 err:
639         fput(sock->file);
640         return ERR_PTR(r);
641 }
642
643 static struct socket *get_tap_socket(int fd)
644 {
645         struct file *file = fget(fd);
646         struct socket *sock;
647
648         if (!file)
649                 return ERR_PTR(-EBADF);
650         sock = tun_get_socket(file);
651         if (!IS_ERR(sock))
652                 return sock;
653         sock = macvtap_get_socket(file);
654         if (IS_ERR(sock))
655                 fput(file);
656         return sock;
657 }
658
659 static struct socket *get_socket(int fd)
660 {
661         struct socket *sock;
662
663         /* special case to disable backend */
664         if (fd == -1)
665                 return NULL;
666         sock = get_raw_socket(fd);
667         if (!IS_ERR(sock))
668                 return sock;
669         sock = get_tap_socket(fd);
670         if (!IS_ERR(sock))
671                 return sock;
672         return ERR_PTR(-ENOTSOCK);
673 }
674
675 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
676 {
677         struct socket *sock, *oldsock;
678         struct vhost_virtqueue *vq;
679         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
680         int r;
681
682         mutex_lock(&n->dev.mutex);
683         r = vhost_dev_check_owner(&n->dev);
684         if (r)
685                 goto err;
686
687         if (index >= VHOST_NET_VQ_MAX) {
688                 r = -ENOBUFS;
689                 goto err;
690         }
691         vq = n->vqs + index;
692         mutex_lock(&vq->mutex);
693
694         /* Verify that ring has been setup correctly. */
695         if (!vhost_vq_access_ok(vq)) {
696                 r = -EFAULT;
697                 goto err_vq;
698         }
699         sock = get_socket(fd);
700         if (IS_ERR(sock)) {
701                 r = PTR_ERR(sock);
702                 goto err_vq;
703         }
704
705         /* start polling new socket */
706         oldsock = rcu_dereference_protected(vq->private_data,
707                                             lockdep_is_held(&vq->mutex));
708         if (sock != oldsock) {
709                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
710                 if (IS_ERR(ubufs)) {
711                         r = PTR_ERR(ubufs);
712                         goto err_ubufs;
713                 }
714                 oldubufs = vq->ubufs;
715                 vq->ubufs = ubufs;
716                 vhost_net_disable_vq(n, vq);
717                 rcu_assign_pointer(vq->private_data, sock);
718                 vhost_net_enable_vq(n, vq);
719
720                 r = vhost_init_used(vq);
721                 if (r)
722                         goto err_vq;
723         }
724
725         mutex_unlock(&vq->mutex);
726
727         if (oldubufs) {
728                 vhost_ubuf_put_and_wait(oldubufs);
729                 mutex_lock(&vq->mutex);
730                 vhost_zerocopy_signal_used(vq);
731                 mutex_unlock(&vq->mutex);
732         }
733
734         if (oldsock) {
735                 vhost_net_flush_vq(n, index);
736                 fput(oldsock->file);
737         }
738
739         mutex_unlock(&n->dev.mutex);
740         return 0;
741
742 err_ubufs:
743         fput(sock->file);
744 err_vq:
745         mutex_unlock(&vq->mutex);
746 err:
747         mutex_unlock(&n->dev.mutex);
748         return r;
749 }
750
751 static long vhost_net_reset_owner(struct vhost_net *n)
752 {
753         struct socket *tx_sock = NULL;
754         struct socket *rx_sock = NULL;
755         long err;
756
757         mutex_lock(&n->dev.mutex);
758         err = vhost_dev_check_owner(&n->dev);
759         if (err)
760                 goto done;
761         vhost_net_stop(n, &tx_sock, &rx_sock);
762         vhost_net_flush(n);
763         err = vhost_dev_reset_owner(&n->dev);
764 done:
765         mutex_unlock(&n->dev.mutex);
766         if (tx_sock)
767                 fput(tx_sock->file);
768         if (rx_sock)
769                 fput(rx_sock->file);
770         return err;
771 }
772
773 static int vhost_net_set_features(struct vhost_net *n, u64 features)
774 {
775         size_t vhost_hlen, sock_hlen, hdr_len;
776         int i;
777
778         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
779                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
780                         sizeof(struct virtio_net_hdr);
781         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
782                 /* vhost provides vnet_hdr */
783                 vhost_hlen = hdr_len;
784                 sock_hlen = 0;
785         } else {
786                 /* socket provides vnet_hdr */
787                 vhost_hlen = 0;
788                 sock_hlen = hdr_len;
789         }
790         mutex_lock(&n->dev.mutex);
791         if ((features & (1 << VHOST_F_LOG_ALL)) &&
792             !vhost_log_access_ok(&n->dev)) {
793                 mutex_unlock(&n->dev.mutex);
794                 return -EFAULT;
795         }
796         n->dev.acked_features = features;
797         smp_wmb();
798         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
799                 mutex_lock(&n->vqs[i].mutex);
800                 n->vqs[i].vhost_hlen = vhost_hlen;
801                 n->vqs[i].sock_hlen = sock_hlen;
802                 mutex_unlock(&n->vqs[i].mutex);
803         }
804         vhost_net_flush(n);
805         mutex_unlock(&n->dev.mutex);
806         return 0;
807 }
808
809 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
810                             unsigned long arg)
811 {
812         struct vhost_net *n = f->private_data;
813         void __user *argp = (void __user *)arg;
814         u64 __user *featurep = argp;
815         struct vhost_vring_file backend;
816         u64 features;
817         int r;
818
819         switch (ioctl) {
820         case VHOST_NET_SET_BACKEND:
821                 if (copy_from_user(&backend, argp, sizeof backend))
822                         return -EFAULT;
823                 return vhost_net_set_backend(n, backend.index, backend.fd);
824         case VHOST_GET_FEATURES:
825                 features = VHOST_FEATURES;
826                 if (copy_to_user(featurep, &features, sizeof features))
827                         return -EFAULT;
828                 return 0;
829         case VHOST_SET_FEATURES:
830                 if (copy_from_user(&features, featurep, sizeof features))
831                         return -EFAULT;
832                 if (features & ~VHOST_FEATURES)
833                         return -EOPNOTSUPP;
834                 return vhost_net_set_features(n, features);
835         case VHOST_RESET_OWNER:
836                 return vhost_net_reset_owner(n);
837         default:
838                 mutex_lock(&n->dev.mutex);
839                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
840                 vhost_net_flush(n);
841                 mutex_unlock(&n->dev.mutex);
842                 return r;
843         }
844 }
845
846 #ifdef CONFIG_COMPAT
847 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
848                                    unsigned long arg)
849 {
850         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
851 }
852 #endif
853
854 static const struct file_operations vhost_net_fops = {
855         .owner          = THIS_MODULE,
856         .release        = vhost_net_release,
857         .unlocked_ioctl = vhost_net_ioctl,
858 #ifdef CONFIG_COMPAT
859         .compat_ioctl   = vhost_net_compat_ioctl,
860 #endif
861         .open           = vhost_net_open,
862         .llseek         = noop_llseek,
863 };
864
865 static struct miscdevice vhost_net_misc = {
866         .minor = VHOST_NET_MINOR,
867         .name = "vhost-net",
868         .fops = &vhost_net_fops,
869 };
870
871 static int vhost_net_init(void)
872 {
873         if (experimental_zcopytx)
874                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
875         return misc_register(&vhost_net_misc);
876 }
877 module_init(vhost_net_init);
878
879 static void vhost_net_exit(void)
880 {
881         misc_deregister(&vhost_net_misc);
882 }
883 module_exit(vhost_net_exit);
884
885 MODULE_VERSION("0.0.1");
886 MODULE_LICENSE("GPL v2");
887 MODULE_AUTHOR("Michael S. Tsirkin");
888 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
889 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
890 MODULE_ALIAS("devname:vhost-net");