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