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