drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <[email protected]>
   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 = &vq->ubuf_info[head];
 239
 240                                 vq->heads[vq->upend_idx].len = len;
 241                                 ubuf->callback = vhost_zerocopy_callback;
 242                                 ubuf->ctx = vq->ubufs;
 243                                 ubuf->desc = vq->upend_idx;
 244                                 msg.msg_control = ubuf;
 245                                 msg.msg_controllen = sizeof(ubuf);
 246                                 ubufs = vq->ubufs;
 247                                 kref_get(&ubufs->kref);
 248                         }
 249                         vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
 250                 }
 251                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 252                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 253                 if (unlikely(err < 0)) {
 254                         if (zcopy) {
 255                                 if (ubufs)
 256                                         vhost_ubuf_put(ubufs);
 257                                 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
 258                                         UIO_MAXIOV;
 259                         }
 260                         vhost_discard_vq_desc(vq, 1);
 261                         tx_poll_start(net, sock);
 262                         break;
 263                 }
 264                 if (err != len)
 265                         pr_debug("Truncated TX packet: "
 266                                  " len %d != %zd\n", err, len);
 267                 if (!zcopy)
 268                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 269                 total_len += len;
 270                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 271                         vhost_poll_queue(&vq->poll);
 272                         break;
 273                 }
 274         }
 275
 276         mutex_unlock(&vq->mutex);
 277 }
 278
 279 static int peek_head_len(struct sock *sk)
 280 {
 281         struct sk_buff *head;
 282         int len = 0;
 283         unsigned long flags;
 284
 285         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 286         head = skb_peek(&sk->sk_receive_queue);
 287         if (likely(head)) {
 288                 len = head->len;
 289                 if (vlan_tx_tag_present(head))
 290                         len += VLAN_HLEN;
 291         }
 292
 293         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 294         return len;
 295 }
 296
 297 /* This is a multi-buffer version of vhost_get_desc, that works if
 298  *      vq has read descriptors only.
 299  * @vq          - the relevant virtqueue
 300  * @datalen     - data length we'll be reading
 301  * @iovcount    - returned count of io vectors we fill
 302  * @log         - vhost log
 303  * @log_num     - log offset
 304  * @quota       - headcount quota, 1 for big buffer
 305  *      returns number of buffer heads allocated, negative on error
 306  */
 307 static int get_rx_bufs(struct vhost_virtqueue *vq,
 308                        struct vring_used_elem *heads,
 309                        int datalen,
 310                        unsigned *iovcount,
 311                        struct vhost_log *log,
 312                        unsigned *log_num,
 313                        unsigned int quota)
 314 {
 315         unsigned int out, in;
 316         int seg = 0;
 317         int headcount = 0;
 318         unsigned d;
 319         int r, nlogs = 0;
 320
 321         while (datalen > 0 && headcount < quota) {
 322                 if (unlikely(seg >= UIO_MAXIOV)) {
 323                         r = -ENOBUFS;
 324                         goto err;
 325                 }
 326                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 327                                       ARRAY_SIZE(vq->iov) - seg, &out,
 328                                       &in, log, log_num);
 329                 if (d == vq->num) {
 330                         r = 0;
 331                         goto err;
 332                 }
 333                 if (unlikely(out || in <= 0)) {
 334                         vq_err(vq, "unexpected descriptor format for RX: "
 335                                 "out %d, in %d\n", out, in);
 336                         r = -EINVAL;
 337                         goto err;
 338                 }
 339                 if (unlikely(log)) {
 340                         nlogs += *log_num;
 341                         log += *log_num;
 342                 }
 343                 heads[headcount].id = d;
 344                 heads[headcount].len = iov_length(vq->iov + seg, in);
 345                 datalen -= heads[headcount].len;
 346                 ++headcount;
 347                 seg += in;
 348         }
 349         heads[headcount - 1].len += datalen;
 350         *iovcount = seg;
 351         if (unlikely(log))
 352                 *log_num = nlogs;
 353         return headcount;
 354 err:
 355         vhost_discard_vq_desc(vq, headcount);
 356         return r;
 357 }
 358
 359 /* Expects to be always run from workqueue - which acts as
 360  * read-size critical section for our kind of RCU. */
 361 static void handle_rx(struct vhost_net *net)
 362 {
 363         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 364         unsigned uninitialized_var(in), log;
 365         struct vhost_log *vq_log;
 366         struct msghdr msg = {
 367                 .msg_name = NULL,
 368                 .msg_namelen = 0,
 369                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 370                 .msg_controllen = 0,
 371                 .msg_iov = vq->iov,
 372                 .msg_flags = MSG_DONTWAIT,
 373         };
 374         struct virtio_net_hdr_mrg_rxbuf hdr = {
 375                 .hdr.flags = 0,
 376                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 377         };
 378         size_t total_len = 0;
 379         int err, headcount, mergeable;
 380         size_t vhost_hlen, sock_hlen;
 381         size_t vhost_len, sock_len;
 382         /* TODO: check that we are running from vhost_worker? */
 383         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 384
 385         if (!sock)
 386                 return;
 387
 388         mutex_lock(&vq->mutex);
 389         vhost_disable_notify(&net->dev, vq);
 390         vhost_hlen = vq->vhost_hlen;
 391         sock_hlen = vq->sock_hlen;
 392
 393         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 394                 vq->log : NULL;
 395         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 396
 397         while ((sock_len = peek_head_len(sock->sk))) {
 398                 sock_len += sock_hlen;
 399                 vhost_len = sock_len + vhost_hlen;
 400                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 401                                         &in, vq_log, &log,
 402                                         likely(mergeable) ? UIO_MAXIOV : 1);
 403                 /* On error, stop handling until the next kick. */
 404                 if (unlikely(headcount < 0))
 405                         break;
 406                 /* OK, now we need to know about added descriptors. */
 407                 if (!headcount) {
 408                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 409                                 /* They have slipped one in as we were
 410                                  * doing that: check again. */
 411                                 vhost_disable_notify(&net->dev, vq);
 412                                 continue;
 413                         }
 414                         /* Nothing new?  Wait for eventfd to tell us
 415                          * they refilled. */
 416                         break;
 417                 }
 418                 /* We don't need to be notified again. */
 419                 if (unlikely((vhost_hlen)))
 420                         /* Skip header. TODO: support TSO. */
 421                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 422                 else
 423                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 424                          * needed because recvmsg can modify msg_iov. */
 425                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 426                 msg.msg_iovlen = in;
 427                 err = sock->ops->recvmsg(NULL, sock, &msg,
 428                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 429                 /* Userspace might have consumed the packet meanwhile:
 430                  * it's not supposed to do this usually, but might be hard
 431                  * to prevent. Discard data we got (if any) and keep going. */
 432                 if (unlikely(err != sock_len)) {
 433                         pr_debug("Discarded rx packet: "
 434                                  " len %d, expected %zd\n", err, sock_len);
 435                         vhost_discard_vq_desc(vq, headcount);
 436                         continue;
 437                 }
 438                 if (unlikely(vhost_hlen) &&
 439                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 440                                       vhost_hlen)) {
 441                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 442                                vq->iov->iov_base);
 443                         break;
 444                 }
 445                 /* TODO: Should check and handle checksum. */
 446                 if (likely(mergeable) &&
 447                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 448                                       offsetof(typeof(hdr), num_buffers),
 449                                       sizeof hdr.num_buffers)) {
 450                         vq_err(vq, "Failed num_buffers write");
 451                         vhost_discard_vq_desc(vq, headcount);
 452                         break;
 453                 }
 454                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 455                                             headcount);
 456                 if (unlikely(vq_log))
 457                         vhost_log_write(vq, vq_log, log, vhost_len);
 458                 total_len += vhost_len;
 459                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 460                         vhost_poll_queue(&vq->poll);
 461                         break;
 462                 }
 463         }
 464
 465         mutex_unlock(&vq->mutex);
 466 }
 467
 468 static void handle_tx_kick(struct vhost_work *work)
 469 {
 470         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 471                                                   poll.work);
 472         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 473
 474         handle_tx(net);
 475 }
 476
 477 static void handle_rx_kick(struct vhost_work *work)
 478 {
 479         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 480                                                   poll.work);
 481         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 482
 483         handle_rx(net);
 484 }
 485
 486 static void handle_tx_net(struct vhost_work *work)
 487 {
 488         struct vhost_net *net = container_of(work, struct vhost_net,
 489                                              poll[VHOST_NET_VQ_TX].work);
 490         handle_tx(net);
 491 }
 492
 493 static void handle_rx_net(struct vhost_work *work)
 494 {
 495         struct vhost_net *net = container_of(work, struct vhost_net,
 496                                              poll[VHOST_NET_VQ_RX].work);
 497         handle_rx(net);
 498 }
 499
 500 static int vhost_net_open(struct inode *inode, struct file *f)
 501 {
 502         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 503         struct vhost_dev *dev;
 504         int r;
 505
 506         if (!n)
 507                 return -ENOMEM;
 508
 509         dev = &n->dev;
 510         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 511         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 512         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 513         if (r < 0) {
 514                 kfree(n);
 515                 return r;
 516         }
 517
 518         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 519         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 520         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 521
 522         f->private_data = n;
 523
 524         return 0;
 525 }
 526
 527 static void vhost_net_disable_vq(struct vhost_net *n,
 528                                  struct vhost_virtqueue *vq)
 529 {
 530         if (!vq->private_data)
 531                 return;
 532         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 533                 tx_poll_stop(n);
 534                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 535         } else
 536                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 537 }
 538
 539 static void vhost_net_enable_vq(struct vhost_net *n,
 540                                 struct vhost_virtqueue *vq)
 541 {
 542         struct socket *sock;
 543
 544         sock = rcu_dereference_protected(vq->private_data,
 545                                          lockdep_is_held(&vq->mutex));
 546         if (!sock)
 547                 return;
 548         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 549                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 550                 tx_poll_start(n, sock);
 551         } else
 552                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 553 }
 554
 555 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 556                                         struct vhost_virtqueue *vq)
 557 {
 558         struct socket *sock;
 559
 560         mutex_lock(&vq->mutex);
 561         sock = rcu_dereference_protected(vq->private_data,
 562                                          lockdep_is_held(&vq->mutex));
 563         vhost_net_disable_vq(n, vq);
 564         rcu_assign_pointer(vq->private_data, NULL);
 565         mutex_unlock(&vq->mutex);
 566         return sock;
 567 }
 568
 569 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 570                            struct socket **rx_sock)
 571 {
 572         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 573         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 574 }
 575
 576 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 577 {
 578         vhost_poll_flush(n->poll + index);
 579         vhost_poll_flush(&n->dev.vqs[index].poll);
 580 }
 581
 582 static void vhost_net_flush(struct vhost_net *n)
 583 {
 584         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 585         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 586 }
 587
 588 static int vhost_net_release(struct inode *inode, struct file *f)
 589 {
 590         struct vhost_net *n = f->private_data;
 591         struct socket *tx_sock;
 592         struct socket *rx_sock;
 593
 594         vhost_net_stop(n, &tx_sock, &rx_sock);
 595         vhost_net_flush(n);
 596         vhost_dev_cleanup(&n->dev, false);
 597         if (tx_sock)
 598                 fput(tx_sock->file);
 599         if (rx_sock)
 600                 fput(rx_sock->file);
 601         /* We do an extra flush before freeing memory,
 602          * since jobs can re-queue themselves. */
 603         vhost_net_flush(n);
 604         kfree(n);
 605         return 0;
 606 }
 607
 608 static struct socket *get_raw_socket(int fd)
 609 {
 610         struct {
 611                 struct sockaddr_ll sa;
 612                 char  buf[MAX_ADDR_LEN];
 613         } uaddr;
 614         int uaddr_len = sizeof uaddr, r;
 615         struct socket *sock = sockfd_lookup(fd, &r);
 616
 617         if (!sock)
 618                 return ERR_PTR(-ENOTSOCK);
 619
 620         /* Parameter checking */
 621         if (sock->sk->sk_type != SOCK_RAW) {
 622                 r = -ESOCKTNOSUPPORT;
 623                 goto err;
 624         }
 625
 626         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 627                                &uaddr_len, 0);
 628         if (r)
 629                 goto err;
 630
 631         if (uaddr.sa.sll_family != AF_PACKET) {
 632                 r = -EPFNOSUPPORT;
 633                 goto err;
 634         }
 635         return sock;
 636 err:
 637         fput(sock->file);
 638         return ERR_PTR(r);
 639 }
 640
 641 static struct socket *get_tap_socket(int fd)
 642 {
 643         struct file *file = fget(fd);
 644         struct socket *sock;
 645
 646         if (!file)
 647                 return ERR_PTR(-EBADF);
 648         sock = tun_get_socket(file);
 649         if (!IS_ERR(sock))
 650                 return sock;
 651         sock = macvtap_get_socket(file);
 652         if (IS_ERR(sock))
 653                 fput(file);
 654         return sock;
 655 }
 656
 657 static struct socket *get_socket(int fd)
 658 {
 659         struct socket *sock;
 660
 661         /* special case to disable backend */
 662         if (fd == -1)
 663                 return NULL;
 664         sock = get_raw_socket(fd);
 665         if (!IS_ERR(sock))
 666                 return sock;
 667         sock = get_tap_socket(fd);
 668         if (!IS_ERR(sock))
 669                 return sock;
 670         return ERR_PTR(-ENOTSOCK);
 671 }
 672
 673 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 674 {
 675         struct socket *sock, *oldsock;
 676         struct vhost_virtqueue *vq;
 677         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 678         int r;
 679
 680         mutex_lock(&n->dev.mutex);
 681         r = vhost_dev_check_owner(&n->dev);
 682         if (r)
 683                 goto err;
 684
 685         if (index >= VHOST_NET_VQ_MAX) {
 686                 r = -ENOBUFS;
 687                 goto err;
 688         }
 689         vq = n->vqs + index;
 690         mutex_lock(&vq->mutex);
 691
 692         /* Verify that ring has been setup correctly. */
 693         if (!vhost_vq_access_ok(vq)) {
 694                 r = -EFAULT;
 695                 goto err_vq;
 696         }
 697         sock = get_socket(fd);
 698         if (IS_ERR(sock)) {
 699                 r = PTR_ERR(sock);
 700                 goto err_vq;
 701         }
 702
 703         /* start polling new socket */
 704         oldsock = rcu_dereference_protected(vq->private_data,
 705                                             lockdep_is_held(&vq->mutex));
 706         if (sock != oldsock) {
 707                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 708                 if (IS_ERR(ubufs)) {
 709                         r = PTR_ERR(ubufs);
 710                         goto err_ubufs;
 711                 }
 712                 oldubufs = vq->ubufs;
 713                 vq->ubufs = ubufs;
 714                 vhost_net_disable_vq(n, vq);
 715                 rcu_assign_pointer(vq->private_data, sock);
 716                 vhost_net_enable_vq(n, vq);
 717
 718                 r = vhost_init_used(vq);
 719                 if (r)
 720                         goto err_vq;
 721         }
 722
 723         mutex_unlock(&vq->mutex);
 724
 725         if (oldubufs) {
 726                 vhost_ubuf_put_and_wait(oldubufs);
 727                 mutex_lock(&vq->mutex);
 728                 vhost_zerocopy_signal_used(vq);
 729                 mutex_unlock(&vq->mutex);
 730         }
 731
 732         if (oldsock) {
 733                 vhost_net_flush_vq(n, index);
 734                 fput(oldsock->file);
 735         }
 736
 737         mutex_unlock(&n->dev.mutex);
 738         return 0;
 739
 740 err_ubufs:
 741         fput(sock->file);
 742 err_vq:
 743         mutex_unlock(&vq->mutex);
 744 err:
 745         mutex_unlock(&n->dev.mutex);
 746         return r;
 747 }
 748
 749 static long vhost_net_reset_owner(struct vhost_net *n)
 750 {
 751         struct socket *tx_sock = NULL;
 752         struct socket *rx_sock = NULL;
 753         long err;
 754
 755         mutex_lock(&n->dev.mutex);
 756         err = vhost_dev_check_owner(&n->dev);
 757         if (err)
 758                 goto done;
 759         vhost_net_stop(n, &tx_sock, &rx_sock);
 760         vhost_net_flush(n);
 761         err = vhost_dev_reset_owner(&n->dev);
 762 done:
 763         mutex_unlock(&n->dev.mutex);
 764         if (tx_sock)
 765                 fput(tx_sock->file);
 766         if (rx_sock)
 767                 fput(rx_sock->file);
 768         return err;
 769 }
 770
 771 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 772 {
 773         size_t vhost_hlen, sock_hlen, hdr_len;
 774         int i;
 775
 776         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 777                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 778                         sizeof(struct virtio_net_hdr);
 779         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 780                 /* vhost provides vnet_hdr */
 781                 vhost_hlen = hdr_len;
 782                 sock_hlen = 0;
 783         } else {
 784                 /* socket provides vnet_hdr */
 785                 vhost_hlen = 0;
 786                 sock_hlen = hdr_len;
 787         }
 788         mutex_lock(&n->dev.mutex);
 789         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 790             !vhost_log_access_ok(&n->dev)) {
 791                 mutex_unlock(&n->dev.mutex);
 792                 return -EFAULT;
 793         }
 794         n->dev.acked_features = features;
 795         smp_wmb();
 796         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 797                 mutex_lock(&n->vqs[i].mutex);
 798                 n->vqs[i].vhost_hlen = vhost_hlen;
 799                 n->vqs[i].sock_hlen = sock_hlen;
 800                 mutex_unlock(&n->vqs[i].mutex);
 801         }
 802         vhost_net_flush(n);
 803         mutex_unlock(&n->dev.mutex);
 804         return 0;
 805 }
 806
 807 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 808                             unsigned long arg)
 809 {
 810         struct vhost_net *n = f->private_data;
 811         void __user *argp = (void __user *)arg;
 812         u64 __user *featurep = argp;
 813         struct vhost_vring_file backend;
 814         u64 features;
 815         int r;
 816
 817         switch (ioctl) {
 818         case VHOST_NET_SET_BACKEND:
 819                 if (copy_from_user(&backend, argp, sizeof backend))
 820                         return -EFAULT;
 821                 return vhost_net_set_backend(n, backend.index, backend.fd);
 822         case VHOST_GET_FEATURES:
 823                 features = VHOST_FEATURES;
 824                 if (copy_to_user(featurep, &features, sizeof features))
 825                         return -EFAULT;
 826                 return 0;
 827         case VHOST_SET_FEATURES:
 828                 if (copy_from_user(&features, featurep, sizeof features))
 829                         return -EFAULT;
 830                 if (features & ~VHOST_FEATURES)
 831                         return -EOPNOTSUPP;
 832                 return vhost_net_set_features(n, features);
 833         case VHOST_RESET_OWNER:
 834                 return vhost_net_reset_owner(n);
 835         default:
 836                 mutex_lock(&n->dev.mutex);
 837                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 838                 vhost_net_flush(n);
 839                 mutex_unlock(&n->dev.mutex);
 840                 return r;
 841         }
 842 }
 843
 844 #ifdef CONFIG_COMPAT
 845 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 846                                    unsigned long arg)
 847 {
 848         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 849 }
 850 #endif
 851
 852 static const struct file_operations vhost_net_fops = {
 853         .owner          = THIS_MODULE,
 854         .release        = vhost_net_release,
 855         .unlocked_ioctl = vhost_net_ioctl,
 856 #ifdef CONFIG_COMPAT
 857         .compat_ioctl   = vhost_net_compat_ioctl,
 858 #endif
 859         .open           = vhost_net_open,
 860         .llseek         = noop_llseek,
 861 };
 862
 863 static struct miscdevice vhost_net_misc = {
 864         .minor = VHOST_NET_MINOR,
 865         .name = "vhost-net",
 866         .fops = &vhost_net_fops,
 867 };
 868
 869 static int vhost_net_init(void)
 870 {
 871         if (experimental_zcopytx)
 872                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
 873         return misc_register(&vhost_net_misc);
 874 }
 875 module_init(vhost_net_init);
 876
 877 static void vhost_net_exit(void)
 878 {
 879         misc_deregister(&vhost_net_misc);
 880 }
 881 module_exit(vhost_net_exit);
 882
 883 MODULE_VERSION("0.0.1");
 884 MODULE_LICENSE("GPL v2");
 885 MODULE_AUTHOR("Michael S. Tsirkin");
 886 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
 887 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
 888 MODULE_ALIAS("devname:vhost-net");