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/mutex.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/rcupdate.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.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>
  25 #include <linux/if_macvlan.h>
  26
  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
  35 enum {
  36         VHOST_NET_VQ_RX = 0,
  37         VHOST_NET_VQ_TX = 1,
  38         VHOST_NET_VQ_MAX = 2,
  39 };
  40
  41 enum vhost_net_poll_state {
  42         VHOST_NET_POLL_DISABLED = 0,
  43         VHOST_NET_POLL_STARTED = 1,
  44         VHOST_NET_POLL_STOPPED = 2,
  45 };
  46
  47 struct 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. */
  58 static 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
  64         while (len && seg < iov_count) {
  65                 size = min(from->iov_len, len);
  66                 to->iov_base = from->iov_base;
  67                 to->iov_len = size;
  68                 from->iov_len -= size;
  69                 from->iov_base += size;
  70                 len -= size;
  71                 ++from;
  72                 ++to;
  73                 ++seg;
  74         }
  75         return seg;
  76 }
  77 /* Copy iovec entries for len bytes from iovec. */
  78 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
  79                            size_t len, int iovcount)
  80 {
  81         int seg = 0;
  82         size_t size;
  83
  84         while (len && seg < iovcount) {
  85                 size = min(from->iov_len, len);
  86                 to->iov_base = from->iov_base;
  87                 to->iov_len = size;
  88                 len -= size;
  89                 ++from;
  90                 ++to;
  91                 ++seg;
  92         }
  93 }
  94
  95 /* Caller must have TX VQ lock */
  96 static void tx_poll_stop(struct vhost_net *net)
  97 {
  98         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
  99                 return;
 100         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 101         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 102 }
 103
 104 /* Caller must have TX VQ lock */
 105 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 106 {
 107         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 108                 return;
 109         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 110         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 111 }
 112
 113 /* Expects to be always run from workqueue - which acts as
 114  * read-size critical section for our kind of RCU. */
 115 static void handle_tx(struct vhost_net *net)
 116 {
 117         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 118         unsigned out, in, s;
 119         int head;
 120         struct msghdr msg = {
 121                 .msg_name = NULL,
 122                 .msg_namelen = 0,
 123                 .msg_control = NULL,
 124                 .msg_controllen = 0,
 125                 .msg_iov = vq->iov,
 126                 .msg_flags = MSG_DONTWAIT,
 127         };
 128         size_t len, total_len = 0;
 129         int err, wmem;
 130         size_t hdr_size;
 131         struct socket *sock;
 132
 133         /* TODO: check that we are running from vhost_worker? */
 134         sock = rcu_dereference_check(vq->private_data, 1);
 135         if (!sock)
 136                 return;
 137
 138         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 139         if (wmem >= sock->sk->sk_sndbuf) {
 140                 mutex_lock(&vq->mutex);
 141                 tx_poll_start(net, sock);
 142                 mutex_unlock(&vq->mutex);
 143                 return;
 144         }
 145
 146         mutex_lock(&vq->mutex);
 147         vhost_disable_notify(vq);
 148
 149         if (wmem < sock->sk->sk_sndbuf / 2)
 150                 tx_poll_stop(net);
 151         hdr_size = vq->vhost_hlen;
 152
 153         for (;;) {
 154                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 155                                          ARRAY_SIZE(vq->iov),
 156                                          &out, &in,
 157                                          NULL, NULL);
 158                 /* On error, stop handling until the next kick. */
 159                 if (unlikely(head < 0))
 160                         break;
 161                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 162                 if (head == vq->num) {
 163                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 164                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 165                                 tx_poll_start(net, sock);
 166                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 167                                 break;
 168                         }
 169                         if (unlikely(vhost_enable_notify(vq))) {
 170                                 vhost_disable_notify(vq);
 171                                 continue;
 172                         }
 173                         break;
 174                 }
 175                 if (in) {
 176                         vq_err(vq, "Unexpected descriptor format for TX: "
 177                                "out %d, int %d\n", out, in);
 178                         break;
 179                 }
 180                 /* Skip header. TODO: support TSO. */
 181                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 182                 msg.msg_iovlen = out;
 183                 len = iov_length(vq->iov, out);
 184                 /* Sanity check */
 185                 if (!len) {
 186                         vq_err(vq, "Unexpected header len for TX: "
 187                                "%zd expected %zd\n",
 188                                iov_length(vq->hdr, s), hdr_size);
 189                         break;
 190                 }
 191                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 192                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 193                 if (unlikely(err < 0)) {
 194                         vhost_discard_vq_desc(vq, 1);
 195                         tx_poll_start(net, sock);
 196                         break;
 197                 }
 198                 if (err != len)
 199                         pr_debug("Truncated TX packet: "
 200                                  " len %d != %zd\n", err, len);
 201                 vhost_add_used_and_signal(&net->dev, vq, head, 0);
 202                 total_len += len;
 203                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 204                         vhost_poll_queue(&vq->poll);
 205                         break;
 206                 }
 207         }
 208
 209         mutex_unlock(&vq->mutex);
 210 }
 211
 212 static int peek_head_len(struct sock *sk)
 213 {
 214         struct sk_buff *head;
 215         int len = 0;
 216
 217         lock_sock(sk);
 218         head = skb_peek(&sk->sk_receive_queue);
 219         if (head)
 220                 len = head->len;
 221         release_sock(sk);
 222         return len;
 223 }
 224
 225 /* This is a multi-buffer version of vhost_get_desc, that works if
 226  *      vq has read descriptors only.
 227  * @vq          - the relevant virtqueue
 228  * @datalen     - data length we'll be reading
 229  * @iovcount    - returned count of io vectors we fill
 230  * @log         - vhost log
 231  * @log_num     - log offset
 232  * @quota       - headcount quota, 1 for big buffer
 233  *      returns number of buffer heads allocated, negative on error
 234  */
 235 static int get_rx_bufs(struct vhost_virtqueue *vq,
 236                        struct vring_used_elem *heads,
 237                        int datalen,
 238                        unsigned *iovcount,
 239                        struct vhost_log *log,
 240                        unsigned *log_num,
 241                        unsigned int quota)
 242 {
 243         unsigned int out, in;
 244         int seg = 0;
 245         int headcount = 0;
 246         unsigned d;
 247         int r, nlogs = 0;
 248
 249         while (datalen > 0 && headcount < quota) {
 250                 if (unlikely(seg >= UIO_MAXIOV)) {
 251                         r = -ENOBUFS;
 252                         goto err;
 253                 }
 254                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 255                                       ARRAY_SIZE(vq->iov) - seg, &out,
 256                                       &in, log, log_num);
 257                 if (d == vq->num) {
 258                         r = 0;
 259                         goto err;
 260                 }
 261                 if (unlikely(out || in <= 0)) {
 262                         vq_err(vq, "unexpected descriptor format for RX: "
 263                                 "out %d, in %d\n", out, in);
 264                         r = -EINVAL;
 265                         goto err;
 266                 }
 267                 if (unlikely(log)) {
 268                         nlogs += *log_num;
 269                         log += *log_num;
 270                 }
 271                 heads[headcount].id = d;
 272                 heads[headcount].len = iov_length(vq->iov + seg, in);
 273                 datalen -= heads[headcount].len;
 274                 ++headcount;
 275                 seg += in;
 276         }
 277         heads[headcount - 1].len += datalen;
 278         *iovcount = seg;
 279         if (unlikely(log))
 280                 *log_num = nlogs;
 281         return headcount;
 282 err:
 283         vhost_discard_vq_desc(vq, headcount);
 284         return r;
 285 }
 286
 287 /* Expects to be always run from workqueue - which acts as
 288  * read-size critical section for our kind of RCU. */
 289 static void handle_rx(struct vhost_net *net)
 290 {
 291         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 292         unsigned uninitialized_var(in), log;
 293         struct vhost_log *vq_log;
 294         struct msghdr msg = {
 295                 .msg_name = NULL,
 296                 .msg_namelen = 0,
 297                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 298                 .msg_controllen = 0,
 299                 .msg_iov = vq->iov,
 300                 .msg_flags = MSG_DONTWAIT,
 301         };
 302         struct virtio_net_hdr_mrg_rxbuf hdr = {
 303                 .hdr.flags = 0,
 304                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 305         };
 306         size_t total_len = 0;
 307         int err, headcount, mergeable;
 308         size_t vhost_hlen, sock_hlen;
 309         size_t vhost_len, sock_len;
 310         /* TODO: check that we are running from vhost_worker? */
 311         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 312
 313         if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
 314                 return;
 315
 316         mutex_lock(&vq->mutex);
 317         vhost_disable_notify(vq);
 318         vhost_hlen = vq->vhost_hlen;
 319         sock_hlen = vq->sock_hlen;
 320
 321         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 322                 vq->log : NULL;
 323         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 324
 325         while ((sock_len = peek_head_len(sock->sk))) {
 326                 sock_len += sock_hlen;
 327                 vhost_len = sock_len + vhost_hlen;
 328                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 329                                         &in, vq_log, &log,
 330                                         likely(mergeable) ? UIO_MAXIOV : 1);
 331                 /* On error, stop handling until the next kick. */
 332                 if (unlikely(headcount < 0))
 333                         break;
 334                 /* OK, now we need to know about added descriptors. */
 335                 if (!headcount) {
 336                         if (unlikely(vhost_enable_notify(vq))) {
 337                                 /* They have slipped one in as we were
 338                                  * doing that: check again. */
 339                                 vhost_disable_notify(vq);
 340                                 continue;
 341                         }
 342                         /* Nothing new?  Wait for eventfd to tell us
 343                          * they refilled. */
 344                         break;
 345                 }
 346                 /* We don't need to be notified again. */
 347                 if (unlikely((vhost_hlen)))
 348                         /* Skip header. TODO: support TSO. */
 349                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 350                 else
 351                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 352                          * needed because recvmsg can modify msg_iov. */
 353                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 354                 msg.msg_iovlen = in;
 355                 err = sock->ops->recvmsg(NULL, sock, &msg,
 356                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 357                 /* Userspace might have consumed the packet meanwhile:
 358                  * it's not supposed to do this usually, but might be hard
 359                  * to prevent. Discard data we got (if any) and keep going. */
 360                 if (unlikely(err != sock_len)) {
 361                         pr_debug("Discarded rx packet: "
 362                                  " len %d, expected %zd\n", err, sock_len);
 363                         vhost_discard_vq_desc(vq, headcount);
 364                         continue;
 365                 }
 366                 if (unlikely(vhost_hlen) &&
 367                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 368                                       vhost_hlen)) {
 369                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 370                                vq->iov->iov_base);
 371                         break;
 372                 }
 373                 /* TODO: Should check and handle checksum. */
 374                 if (likely(mergeable) &&
 375                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 376                                       offsetof(typeof(hdr), num_buffers),
 377                                       sizeof hdr.num_buffers)) {
 378                         vq_err(vq, "Failed num_buffers write");
 379                         vhost_discard_vq_desc(vq, headcount);
 380                         break;
 381                 }
 382                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 383                                             headcount);
 384                 if (unlikely(vq_log))
 385                         vhost_log_write(vq, vq_log, log, vhost_len);
 386                 total_len += vhost_len;
 387                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 388                         vhost_poll_queue(&vq->poll);
 389                         break;
 390                 }
 391         }
 392
 393         mutex_unlock(&vq->mutex);
 394 }
 395
 396 static void handle_tx_kick(struct vhost_work *work)
 397 {
 398         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 399                                                   poll.work);
 400         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 401
 402         handle_tx(net);
 403 }
 404
 405 static void handle_rx_kick(struct vhost_work *work)
 406 {
 407         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 408                                                   poll.work);
 409         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 410
 411         handle_rx(net);
 412 }
 413
 414 static void handle_tx_net(struct vhost_work *work)
 415 {
 416         struct vhost_net *net = container_of(work, struct vhost_net,
 417                                              poll[VHOST_NET_VQ_TX].work);
 418         handle_tx(net);
 419 }
 420
 421 static void handle_rx_net(struct vhost_work *work)
 422 {
 423         struct vhost_net *net = container_of(work, struct vhost_net,
 424                                              poll[VHOST_NET_VQ_RX].work);
 425         handle_rx(net);
 426 }
 427
 428 static int vhost_net_open(struct inode *inode, struct file *f)
 429 {
 430         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 431         struct vhost_dev *dev;
 432         int r;
 433
 434         if (!n)
 435                 return -ENOMEM;
 436
 437         dev = &n->dev;
 438         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 439         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 440         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 441         if (r < 0) {
 442                 kfree(n);
 443                 return r;
 444         }
 445
 446         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 447         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 448         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 449
 450         f->private_data = n;
 451
 452         return 0;
 453 }
 454
 455 static void vhost_net_disable_vq(struct vhost_net *n,
 456                                  struct vhost_virtqueue *vq)
 457 {
 458         if (!vq->private_data)
 459                 return;
 460         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 461                 tx_poll_stop(n);
 462                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 463         } else
 464                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 465 }
 466
 467 static void vhost_net_enable_vq(struct vhost_net *n,
 468                                 struct vhost_virtqueue *vq)
 469 {
 470         struct socket *sock;
 471
 472         sock = rcu_dereference_protected(vq->private_data,
 473                                          lockdep_is_held(&vq->mutex));
 474         if (!sock)
 475                 return;
 476         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 477                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 478                 tx_poll_start(n, sock);
 479         } else
 480                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 481 }
 482
 483 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 484                                         struct vhost_virtqueue *vq)
 485 {
 486         struct socket *sock;
 487
 488         mutex_lock(&vq->mutex);
 489         sock = rcu_dereference_protected(vq->private_data,
 490                                          lockdep_is_held(&vq->mutex));
 491         vhost_net_disable_vq(n, vq);
 492         rcu_assign_pointer(vq->private_data, NULL);
 493         mutex_unlock(&vq->mutex);
 494         return sock;
 495 }
 496
 497 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 498                            struct socket **rx_sock)
 499 {
 500         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 501         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 502 }
 503
 504 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 505 {
 506         vhost_poll_flush(n->poll + index);
 507         vhost_poll_flush(&n->dev.vqs[index].poll);
 508 }
 509
 510 static void vhost_net_flush(struct vhost_net *n)
 511 {
 512         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 513         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 514 }
 515
 516 static int vhost_net_release(struct inode *inode, struct file *f)
 517 {
 518         struct vhost_net *n = f->private_data;
 519         struct socket *tx_sock;
 520         struct socket *rx_sock;
 521
 522         vhost_net_stop(n, &tx_sock, &rx_sock);
 523         vhost_net_flush(n);
 524         vhost_dev_cleanup(&n->dev);
 525         if (tx_sock)
 526                 fput(tx_sock->file);
 527         if (rx_sock)
 528                 fput(rx_sock->file);
 529         /* We do an extra flush before freeing memory,
 530          * since jobs can re-queue themselves. */
 531         vhost_net_flush(n);
 532         kfree(n);
 533         return 0;
 534 }
 535
 536 static struct socket *get_raw_socket(int fd)
 537 {
 538         struct {
 539                 struct sockaddr_ll sa;
 540                 char  buf[MAX_ADDR_LEN];
 541         } uaddr;
 542         int uaddr_len = sizeof uaddr, r;
 543         struct socket *sock = sockfd_lookup(fd, &r);
 544
 545         if (!sock)
 546                 return ERR_PTR(-ENOTSOCK);
 547
 548         /* Parameter checking */
 549         if (sock->sk->sk_type != SOCK_RAW) {
 550                 r = -ESOCKTNOSUPPORT;
 551                 goto err;
 552         }
 553
 554         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 555                                &uaddr_len, 0);
 556         if (r)
 557                 goto err;
 558
 559         if (uaddr.sa.sll_family != AF_PACKET) {
 560                 r = -EPFNOSUPPORT;
 561                 goto err;
 562         }
 563         return sock;
 564 err:
 565         fput(sock->file);
 566         return ERR_PTR(r);
 567 }
 568
 569 static struct socket *get_tap_socket(int fd)
 570 {
 571         struct file *file = fget(fd);
 572         struct socket *sock;
 573
 574         if (!file)
 575                 return ERR_PTR(-EBADF);
 576         sock = tun_get_socket(file);
 577         if (!IS_ERR(sock))
 578                 return sock;
 579         sock = macvtap_get_socket(file);
 580         if (IS_ERR(sock))
 581                 fput(file);
 582         return sock;
 583 }
 584
 585 static struct socket *get_socket(int fd)
 586 {
 587         struct socket *sock;
 588
 589         /* special case to disable backend */
 590         if (fd == -1)
 591                 return NULL;
 592         sock = get_raw_socket(fd);
 593         if (!IS_ERR(sock))
 594                 return sock;
 595         sock = get_tap_socket(fd);
 596         if (!IS_ERR(sock))
 597                 return sock;
 598         return ERR_PTR(-ENOTSOCK);
 599 }
 600
 601 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 602 {
 603         struct socket *sock, *oldsock;
 604         struct vhost_virtqueue *vq;
 605         int r;
 606
 607         mutex_lock(&n->dev.mutex);
 608         r = vhost_dev_check_owner(&n->dev);
 609         if (r)
 610                 goto err;
 611
 612         if (index >= VHOST_NET_VQ_MAX) {
 613                 r = -ENOBUFS;
 614                 goto err;
 615         }
 616         vq = n->vqs + index;
 617         mutex_lock(&vq->mutex);
 618
 619         /* Verify that ring has been setup correctly. */
 620         if (!vhost_vq_access_ok(vq)) {
 621                 r = -EFAULT;
 622                 goto err_vq;
 623         }
 624         sock = get_socket(fd);
 625         if (IS_ERR(sock)) {
 626                 r = PTR_ERR(sock);
 627                 goto err_vq;
 628         }
 629
 630         /* start polling new socket */
 631         oldsock = rcu_dereference_protected(vq->private_data,
 632                                             lockdep_is_held(&vq->mutex));
 633         if (sock != oldsock) {
 634                 vhost_net_disable_vq(n, vq);
 635                 rcu_assign_pointer(vq->private_data, sock);
 636                 vhost_net_enable_vq(n, vq);
 637         }
 638
 639         mutex_unlock(&vq->mutex);
 640
 641         if (oldsock) {
 642                 vhost_net_flush_vq(n, index);
 643                 fput(oldsock->file);
 644         }
 645
 646         mutex_unlock(&n->dev.mutex);
 647         return 0;
 648
 649 err_vq:
 650         mutex_unlock(&vq->mutex);
 651 err:
 652         mutex_unlock(&n->dev.mutex);
 653         return r;
 654 }
 655
 656 static long vhost_net_reset_owner(struct vhost_net *n)
 657 {
 658         struct socket *tx_sock = NULL;
 659         struct socket *rx_sock = NULL;
 660         long err;
 661
 662         mutex_lock(&n->dev.mutex);
 663         err = vhost_dev_check_owner(&n->dev);
 664         if (err)
 665                 goto done;
 666         vhost_net_stop(n, &tx_sock, &rx_sock);
 667         vhost_net_flush(n);
 668         err = vhost_dev_reset_owner(&n->dev);
 669 done:
 670         mutex_unlock(&n->dev.mutex);
 671         if (tx_sock)
 672                 fput(tx_sock->file);
 673         if (rx_sock)
 674                 fput(rx_sock->file);
 675         return err;
 676 }
 677
 678 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 679 {
 680         size_t vhost_hlen, sock_hlen, hdr_len;
 681         int i;
 682
 683         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 684                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 685                         sizeof(struct virtio_net_hdr);
 686         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 687                 /* vhost provides vnet_hdr */
 688                 vhost_hlen = hdr_len;
 689                 sock_hlen = 0;
 690         } else {
 691                 /* socket provides vnet_hdr */
 692                 vhost_hlen = 0;
 693                 sock_hlen = hdr_len;
 694         }
 695         mutex_lock(&n->dev.mutex);
 696         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 697             !vhost_log_access_ok(&n->dev)) {
 698                 mutex_unlock(&n->dev.mutex);
 699                 return -EFAULT;
 700         }
 701         n->dev.acked_features = features;
 702         smp_wmb();
 703         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 704                 mutex_lock(&n->vqs[i].mutex);
 705                 n->vqs[i].vhost_hlen = vhost_hlen;
 706                 n->vqs[i].sock_hlen = sock_hlen;
 707                 mutex_unlock(&n->vqs[i].mutex);
 708         }
 709         vhost_net_flush(n);
 710         mutex_unlock(&n->dev.mutex);
 711         return 0;
 712 }
 713
 714 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 715                             unsigned long arg)
 716 {
 717         struct vhost_net *n = f->private_data;
 718         void __user *argp = (void __user *)arg;
 719         u64 __user *featurep = argp;
 720         struct vhost_vring_file backend;
 721         u64 features;
 722         int r;
 723
 724         switch (ioctl) {
 725         case VHOST_NET_SET_BACKEND:
 726                 if (copy_from_user(&backend, argp, sizeof backend))
 727                         return -EFAULT;
 728                 return vhost_net_set_backend(n, backend.index, backend.fd);
 729         case VHOST_GET_FEATURES:
 730                 features = VHOST_FEATURES;
 731                 if (copy_to_user(featurep, &features, sizeof features))
 732                         return -EFAULT;
 733                 return 0;
 734         case VHOST_SET_FEATURES:
 735                 if (copy_from_user(&features, featurep, sizeof features))
 736                         return -EFAULT;
 737                 if (features & ~VHOST_FEATURES)
 738                         return -EOPNOTSUPP;
 739                 return vhost_net_set_features(n, features);
 740         case VHOST_RESET_OWNER:
 741                 return vhost_net_reset_owner(n);
 742         default:
 743                 mutex_lock(&n->dev.mutex);
 744                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 745                 vhost_net_flush(n);
 746                 mutex_unlock(&n->dev.mutex);
 747                 return r;
 748         }
 749 }
 750
 751 #ifdef CONFIG_COMPAT
 752 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 753                                    unsigned long arg)
 754 {
 755         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 756 }
 757 #endif
 758
 759 static const struct file_operations vhost_net_fops = {
 760         .owner          = THIS_MODULE,
 761         .release        = vhost_net_release,
 762         .unlocked_ioctl = vhost_net_ioctl,
 763 #ifdef CONFIG_COMPAT
 764         .compat_ioctl   = vhost_net_compat_ioctl,
 765 #endif
 766         .open           = vhost_net_open,
 767         .llseek         = noop_llseek,
 768 };
 769
 770 static struct miscdevice vhost_net_misc = {
 771         MISC_DYNAMIC_MINOR,
 772         "vhost-net",
 773         &vhost_net_fops,
 774 };
 775
 776 static int vhost_net_init(void)
 777 {
 778         return misc_register(&vhost_net_misc);
 779 }
 780 module_init(vhost_net_init);
 781
 782 static void vhost_net_exit(void)
 783 {
 784         misc_deregister(&vhost_net_misc);
 785 }
 786 module_exit(vhost_net_exit);
 787
 788 MODULE_VERSION("0.0.1");
 789 MODULE_LICENSE("GPL v2");
 790 MODULE_AUTHOR("Michael S. Tsirkin");
 791 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");