drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/rcupdate.h>
  19 #include <linux/file.h>
  20 #include <linux/slab.h>
  21
  22 #include <linux/net.h>
  23 #include <linux/if_packet.h>
  24 #include <linux/if_arp.h>
  25 #include <linux/if_tun.h>
  26 #include <linux/if_macvlan.h>
  27 #include <linux/if_vlan.h>
  28
  29 #include <net/sock.h>
  30
  31 #include "vhost.h"
  32
  33 static int experimental_zcopytx;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
  36
  37 /* Max number of bytes transferred before requeueing the job.
  38  * Using this limit prevents one virtqueue from starving others. */
  39 #define VHOST_NET_WEIGHT 0x80000
  40
  41 /* MAX number of TX used buffers for outstanding zerocopy */
  42 #define VHOST_MAX_PEND 128
  43 #define VHOST_GOODCOPY_LEN 256
  44
  45 enum {
  46         VHOST_NET_VQ_RX = 0,
  47         VHOST_NET_VQ_TX = 1,
  48         VHOST_NET_VQ_MAX = 2,
  49 };
  50
  51 enum vhost_net_poll_state {
  52         VHOST_NET_POLL_DISABLED = 0,
  53         VHOST_NET_POLL_STARTED = 1,
  54         VHOST_NET_POLL_STOPPED = 2,
  55 };
  56
  57 struct vhost_net {
  58         struct vhost_dev dev;
  59         struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  60         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  61         /* Tells us whether we are polling a socket for TX.
  62          * We only do this when socket buffer fills up.
  63          * Protected by tx vq lock. */
  64         enum vhost_net_poll_state tx_poll_state;
  65 };
  66
  67 static bool vhost_sock_zcopy(struct socket *sock)
  68 {
  69         return unlikely(experimental_zcopytx) &&
  70                 sock_flag(sock->sk, SOCK_ZEROCOPY);
  71 }
  72
  73 /* Pop first len bytes from iovec. Return number of segments used. */
  74 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
  75                           size_t len, int iov_count)
  76 {
  77         int seg = 0;
  78         size_t size;
  79
  80         while (len && seg < iov_count) {
  81                 size = min(from->iov_len, len);
  82                 to->iov_base = from->iov_base;
  83                 to->iov_len = size;
  84                 from->iov_len -= size;
  85                 from->iov_base += size;
  86                 len -= size;
  87                 ++from;
  88                 ++to;
  89                 ++seg;
  90         }
  91         return seg;
  92 }
  93 /* Copy iovec entries for len bytes from iovec. */
  94 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
  95                            size_t len, int iovcount)
  96 {
  97         int seg = 0;
  98         size_t size;
  99
 100         while (len && seg < iovcount) {
 101                 size = min(from->iov_len, len);
 102                 to->iov_base = from->iov_base;
 103                 to->iov_len = size;
 104                 len -= size;
 105                 ++from;
 106                 ++to;
 107                 ++seg;
 108         }
 109 }
 110
 111 /* Caller must have TX VQ lock */
 112 static void tx_poll_stop(struct vhost_net *net)
 113 {
 114         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
 115                 return;
 116         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 117         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 118 }
 119
 120 /* Caller must have TX VQ lock */
 121 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 122 {
 123         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 124                 return;
 125         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 126         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 127 }
 128
 129 /* Expects to be always run from workqueue - which acts as
 130  * read-size critical section for our kind of RCU. */
 131 static void handle_tx(struct vhost_net *net)
 132 {
 133         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 134         unsigned out, in, s;
 135         int head;
 136         struct msghdr msg = {
 137                 .msg_name = NULL,
 138                 .msg_namelen = 0,
 139                 .msg_control = NULL,
 140                 .msg_controllen = 0,
 141                 .msg_iov = vq->iov,
 142                 .msg_flags = MSG_DONTWAIT,
 143         };
 144         size_t len, total_len = 0;
 145         int err, wmem;
 146         size_t hdr_size;
 147         struct socket *sock;
 148         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 149         bool zcopy;
 150
 151         /* TODO: check that we are running from vhost_worker? */
 152         sock = rcu_dereference_check(vq->private_data, 1);
 153         if (!sock)
 154                 return;
 155
 156         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 157         if (wmem >= sock->sk->sk_sndbuf) {
 158                 mutex_lock(&vq->mutex);
 159                 tx_poll_start(net, sock);
 160                 mutex_unlock(&vq->mutex);
 161                 return;
 162         }
 163
 164         mutex_lock(&vq->mutex);
 165         vhost_disable_notify(&net->dev, vq);
 166
 167         if (wmem < sock->sk->sk_sndbuf / 2)
 168                 tx_poll_stop(net);
 169         hdr_size = vq->vhost_hlen;
 170         zcopy = vq->ubufs;
 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                         if (err == -EAGAIN || err == -ENOBUFS)
 262                                 tx_poll_start(net, sock);
 263                         break;
 264                 }
 265                 if (err != len)
 266                         pr_debug("Truncated TX packet: "
 267                                  " len %d != %zd\n", err, len);
 268                 if (!zcopy)
 269                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 270                 else
 271                         vhost_zerocopy_signal_used(vq);
 272                 total_len += len;
 273                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 274                         vhost_poll_queue(&vq->poll);
 275                         break;
 276                 }
 277         }
 278
 279         mutex_unlock(&vq->mutex);
 280 }
 281
 282 static int peek_head_len(struct sock *sk)
 283 {
 284         struct sk_buff *head;
 285         int len = 0;
 286         unsigned long flags;
 287
 288         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 289         head = skb_peek(&sk->sk_receive_queue);
 290         if (likely(head)) {
 291                 len = head->len;
 292                 if (vlan_tx_tag_present(head))
 293                         len += VLAN_HLEN;
 294         }
 295
 296         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 297         return len;
 298 }
 299
 300 /* This is a multi-buffer version of vhost_get_desc, that works if
 301  *      vq has read descriptors only.
 302  * @vq          - the relevant virtqueue
 303  * @datalen     - data length we'll be reading
 304  * @iovcount    - returned count of io vectors we fill
 305  * @log         - vhost log
 306  * @log_num     - log offset
 307  * @quota       - headcount quota, 1 for big buffer
 308  *      returns number of buffer heads allocated, negative on error
 309  */
 310 static int get_rx_bufs(struct vhost_virtqueue *vq,
 311                        struct vring_used_elem *heads,
 312                        int datalen,
 313                        unsigned *iovcount,
 314                        struct vhost_log *log,
 315                        unsigned *log_num,
 316                        unsigned int quota)
 317 {
 318         unsigned int out, in;
 319         int seg = 0;
 320         int headcount = 0;
 321         unsigned d;
 322         int r, nlogs = 0;
 323
 324         while (datalen > 0 && headcount < quota) {
 325                 if (unlikely(seg >= UIO_MAXIOV)) {
 326                         r = -ENOBUFS;
 327                         goto err;
 328                 }
 329                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 330                                       ARRAY_SIZE(vq->iov) - seg, &out,
 331                                       &in, log, log_num);
 332                 if (d == vq->num) {
 333                         r = 0;
 334                         goto err;
 335                 }
 336                 if (unlikely(out || in <= 0)) {
 337                         vq_err(vq, "unexpected descriptor format for RX: "
 338                                 "out %d, in %d\n", out, in);
 339                         r = -EINVAL;
 340                         goto err;
 341                 }
 342                 if (unlikely(log)) {
 343                         nlogs += *log_num;
 344                         log += *log_num;
 345                 }
 346                 heads[headcount].id = d;
 347                 heads[headcount].len = iov_length(vq->iov + seg, in);
 348                 datalen -= heads[headcount].len;
 349                 ++headcount;
 350                 seg += in;
 351         }
 352         heads[headcount - 1].len += datalen;
 353         *iovcount = seg;
 354         if (unlikely(log))
 355                 *log_num = nlogs;
 356         return headcount;
 357 err:
 358         vhost_discard_vq_desc(vq, headcount);
 359         return r;
 360 }
 361
 362 /* Expects to be always run from workqueue - which acts as
 363  * read-size critical section for our kind of RCU. */
 364 static void handle_rx(struct vhost_net *net)
 365 {
 366         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 367         unsigned uninitialized_var(in), log;
 368         struct vhost_log *vq_log;
 369         struct msghdr msg = {
 370                 .msg_name = NULL,
 371                 .msg_namelen = 0,
 372                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 373                 .msg_controllen = 0,
 374                 .msg_iov = vq->iov,
 375                 .msg_flags = MSG_DONTWAIT,
 376         };
 377         struct virtio_net_hdr_mrg_rxbuf hdr = {
 378                 .hdr.flags = 0,
 379                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 380         };
 381         size_t total_len = 0;
 382         int err, mergeable;
 383         s16 headcount;
 384         size_t vhost_hlen, sock_hlen;
 385         size_t vhost_len, sock_len;
 386         /* TODO: check that we are running from vhost_worker? */
 387         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 388
 389         if (!sock)
 390                 return;
 391
 392         mutex_lock(&vq->mutex);
 393         vhost_disable_notify(&net->dev, vq);
 394         vhost_hlen = vq->vhost_hlen;
 395         sock_hlen = vq->sock_hlen;
 396
 397         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 398                 vq->log : NULL;
 399         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 400
 401         while ((sock_len = peek_head_len(sock->sk))) {
 402                 sock_len += sock_hlen;
 403                 vhost_len = sock_len + vhost_hlen;
 404                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 405                                         &in, vq_log, &log,
 406                                         likely(mergeable) ? UIO_MAXIOV : 1);
 407                 /* On error, stop handling until the next kick. */
 408                 if (unlikely(headcount < 0))
 409                         break;
 410                 /* OK, now we need to know about added descriptors. */
 411                 if (!headcount) {
 412                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 413                                 /* They have slipped one in as we were
 414                                  * doing that: check again. */
 415                                 vhost_disable_notify(&net->dev, vq);
 416                                 continue;
 417                         }
 418                         /* Nothing new?  Wait for eventfd to tell us
 419                          * they refilled. */
 420                         break;
 421                 }
 422                 /* We don't need to be notified again. */
 423                 if (unlikely((vhost_hlen)))
 424                         /* Skip header. TODO: support TSO. */
 425                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 426                 else
 427                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 428                          * needed because recvmsg can modify msg_iov. */
 429                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 430                 msg.msg_iovlen = in;
 431                 err = sock->ops->recvmsg(NULL, sock, &msg,
 432                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 433                 /* Userspace might have consumed the packet meanwhile:
 434                  * it's not supposed to do this usually, but might be hard
 435                  * to prevent. Discard data we got (if any) and keep going. */
 436                 if (unlikely(err != sock_len)) {
 437                         pr_debug("Discarded rx packet: "
 438                                  " len %d, expected %zd\n", err, sock_len);
 439                         vhost_discard_vq_desc(vq, headcount);
 440                         continue;
 441                 }
 442                 if (unlikely(vhost_hlen) &&
 443                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 444                                       vhost_hlen)) {
 445                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 446                                vq->iov->iov_base);
 447                         break;
 448                 }
 449                 /* TODO: Should check and handle checksum. */
 450                 if (likely(mergeable) &&
 451                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 452                                       offsetof(typeof(hdr), num_buffers),
 453                                       sizeof hdr.num_buffers)) {
 454                         vq_err(vq, "Failed num_buffers write");
 455                         vhost_discard_vq_desc(vq, headcount);
 456                         break;
 457                 }
 458                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 459                                             headcount);
 460                 if (unlikely(vq_log))
 461                         vhost_log_write(vq, vq_log, log, vhost_len);
 462                 total_len += vhost_len;
 463                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 464                         vhost_poll_queue(&vq->poll);
 465                         break;
 466                 }
 467         }
 468
 469         mutex_unlock(&vq->mutex);
 470 }
 471
 472 static void handle_tx_kick(struct vhost_work *work)
 473 {
 474         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 475                                                   poll.work);
 476         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 477
 478         handle_tx(net);
 479 }
 480
 481 static void handle_rx_kick(struct vhost_work *work)
 482 {
 483         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 484                                                   poll.work);
 485         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 486
 487         handle_rx(net);
 488 }
 489
 490 static void handle_tx_net(struct vhost_work *work)
 491 {
 492         struct vhost_net *net = container_of(work, struct vhost_net,
 493                                              poll[VHOST_NET_VQ_TX].work);
 494         handle_tx(net);
 495 }
 496
 497 static void handle_rx_net(struct vhost_work *work)
 498 {
 499         struct vhost_net *net = container_of(work, struct vhost_net,
 500                                              poll[VHOST_NET_VQ_RX].work);
 501         handle_rx(net);
 502 }
 503
 504 static int vhost_net_open(struct inode *inode, struct file *f)
 505 {
 506         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 507         struct vhost_dev *dev;
 508         int r;
 509
 510         if (!n)
 511                 return -ENOMEM;
 512
 513         dev = &n->dev;
 514         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 515         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 516         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 517         if (r < 0) {
 518                 kfree(n);
 519                 return r;
 520         }
 521
 522         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 523         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 524         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 525
 526         f->private_data = n;
 527
 528         return 0;
 529 }
 530
 531 static void vhost_net_disable_vq(struct vhost_net *n,
 532                                  struct vhost_virtqueue *vq)
 533 {
 534         if (!vq->private_data)
 535                 return;
 536         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 537                 tx_poll_stop(n);
 538                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 539         } else
 540                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 541 }
 542
 543 static void vhost_net_enable_vq(struct vhost_net *n,
 544                                 struct vhost_virtqueue *vq)
 545 {
 546         struct socket *sock;
 547
 548         sock = rcu_dereference_protected(vq->private_data,
 549                                          lockdep_is_held(&vq->mutex));
 550         if (!sock)
 551                 return;
 552         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 553                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 554                 tx_poll_start(n, sock);
 555         } else
 556                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 557 }
 558
 559 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 560                                         struct vhost_virtqueue *vq)
 561 {
 562         struct socket *sock;
 563
 564         mutex_lock(&vq->mutex);
 565         sock = rcu_dereference_protected(vq->private_data,
 566                                          lockdep_is_held(&vq->mutex));
 567         vhost_net_disable_vq(n, vq);
 568         rcu_assign_pointer(vq->private_data, NULL);
 569         mutex_unlock(&vq->mutex);
 570         return sock;
 571 }
 572
 573 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 574                            struct socket **rx_sock)
 575 {
 576         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 577         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 578 }
 579
 580 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 581 {
 582         vhost_poll_flush(n->poll + index);
 583         vhost_poll_flush(&n->dev.vqs[index].poll);
 584 }
 585
 586 static void vhost_net_flush(struct vhost_net *n)
 587 {
 588         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 589         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 590 }
 591
 592 static int vhost_net_release(struct inode *inode, struct file *f)
 593 {
 594         struct vhost_net *n = f->private_data;
 595         struct socket *tx_sock;
 596         struct socket *rx_sock;
 597
 598         vhost_net_stop(n, &tx_sock, &rx_sock);
 599         vhost_net_flush(n);
 600         vhost_dev_cleanup(&n->dev, false);
 601         if (tx_sock)
 602                 fput(tx_sock->file);
 603         if (rx_sock)
 604                 fput(rx_sock->file);
 605         /* We do an extra flush before freeing memory,
 606          * since jobs can re-queue themselves. */
 607         vhost_net_flush(n);
 608         kfree(n);
 609         return 0;
 610 }
 611
 612 static struct socket *get_raw_socket(int fd)
 613 {
 614         struct {
 615                 struct sockaddr_ll sa;
 616                 char  buf[MAX_ADDR_LEN];
 617         } uaddr;
 618         int uaddr_len = sizeof uaddr, r;
 619         struct socket *sock = sockfd_lookup(fd, &r);
 620
 621         if (!sock)
 622                 return ERR_PTR(-ENOTSOCK);
 623
 624         /* Parameter checking */
 625         if (sock->sk->sk_type != SOCK_RAW) {
 626                 r = -ESOCKTNOSUPPORT;
 627                 goto err;
 628         }
 629
 630         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 631                                &uaddr_len, 0);
 632         if (r)
 633                 goto err;
 634
 635         if (uaddr.sa.sll_family != AF_PACKET) {
 636                 r = -EPFNOSUPPORT;
 637                 goto err;
 638         }
 639         return sock;
 640 err:
 641         fput(sock->file);
 642         return ERR_PTR(r);
 643 }
 644
 645 static struct socket *get_tap_socket(int fd)
 646 {
 647         struct file *file = fget(fd);
 648         struct socket *sock;
 649
 650         if (!file)
 651                 return ERR_PTR(-EBADF);
 652         sock = tun_get_socket(file);
 653         if (!IS_ERR(sock))
 654                 return sock;
 655         sock = macvtap_get_socket(file);
 656         if (IS_ERR(sock))
 657                 fput(file);
 658         return sock;
 659 }
 660
 661 static struct socket *get_socket(int fd)
 662 {
 663         struct socket *sock;
 664
 665         /* special case to disable backend */
 666         if (fd == -1)
 667                 return NULL;
 668         sock = get_raw_socket(fd);
 669         if (!IS_ERR(sock))
 670                 return sock;
 671         sock = get_tap_socket(fd);
 672         if (!IS_ERR(sock))
 673                 return sock;
 674         return ERR_PTR(-ENOTSOCK);
 675 }
 676
 677 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 678 {
 679         struct socket *sock, *oldsock;
 680         struct vhost_virtqueue *vq;
 681         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 682         int r;
 683
 684         mutex_lock(&n->dev.mutex);
 685         r = vhost_dev_check_owner(&n->dev);
 686         if (r)
 687                 goto err;
 688
 689         if (index >= VHOST_NET_VQ_MAX) {
 690                 r = -ENOBUFS;
 691                 goto err;
 692         }
 693         vq = n->vqs + index;
 694         mutex_lock(&vq->mutex);
 695
 696         /* Verify that ring has been setup correctly. */
 697         if (!vhost_vq_access_ok(vq)) {
 698                 r = -EFAULT;
 699                 goto err_vq;
 700         }
 701         sock = get_socket(fd);
 702         if (IS_ERR(sock)) {
 703                 r = PTR_ERR(sock);
 704                 goto err_vq;
 705         }
 706
 707         /* start polling new socket */
 708         oldsock = rcu_dereference_protected(vq->private_data,
 709                                             lockdep_is_held(&vq->mutex));
 710         if (sock != oldsock) {
 711                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 712                 if (IS_ERR(ubufs)) {
 713                         r = PTR_ERR(ubufs);
 714                         goto err_ubufs;
 715                 }
 716                 oldubufs = vq->ubufs;
 717                 vq->ubufs = ubufs;
 718                 vhost_net_disable_vq(n, vq);
 719                 rcu_assign_pointer(vq->private_data, sock);
 720                 vhost_net_enable_vq(n, vq);
 721
 722                 r = vhost_init_used(vq);
 723                 if (r)
 724                         goto err_vq;
 725         }
 726
 727         mutex_unlock(&vq->mutex);
 728
 729         if (oldubufs) {
 730                 vhost_ubuf_put_and_wait(oldubufs);
 731                 mutex_lock(&vq->mutex);
 732                 vhost_zerocopy_signal_used(vq);
 733                 mutex_unlock(&vq->mutex);
 734         }
 735
 736         if (oldsock) {
 737                 vhost_net_flush_vq(n, index);
 738                 fput(oldsock->file);
 739         }
 740
 741         mutex_unlock(&n->dev.mutex);
 742         return 0;
 743
 744 err_ubufs:
 745         fput(sock->file);
 746 err_vq:
 747         mutex_unlock(&vq->mutex);
 748 err:
 749         mutex_unlock(&n->dev.mutex);
 750         return r;
 751 }
 752
 753 static long vhost_net_reset_owner(struct vhost_net *n)
 754 {
 755         struct socket *tx_sock = NULL;
 756         struct socket *rx_sock = NULL;
 757         long err;
 758
 759         mutex_lock(&n->dev.mutex);
 760         err = vhost_dev_check_owner(&n->dev);
 761         if (err)
 762                 goto done;
 763         vhost_net_stop(n, &tx_sock, &rx_sock);
 764         vhost_net_flush(n);
 765         err = vhost_dev_reset_owner(&n->dev);
 766 done:
 767         mutex_unlock(&n->dev.mutex);
 768         if (tx_sock)
 769                 fput(tx_sock->file);
 770         if (rx_sock)
 771                 fput(rx_sock->file);
 772         return err;
 773 }
 774
 775 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 776 {
 777         size_t vhost_hlen, sock_hlen, hdr_len;
 778         int i;
 779
 780         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 781                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 782                         sizeof(struct virtio_net_hdr);
 783         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 784                 /* vhost provides vnet_hdr */
 785                 vhost_hlen = hdr_len;
 786                 sock_hlen = 0;
 787         } else {
 788                 /* socket provides vnet_hdr */
 789                 vhost_hlen = 0;
 790                 sock_hlen = hdr_len;
 791         }
 792         mutex_lock(&n->dev.mutex);
 793         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 794             !vhost_log_access_ok(&n->dev)) {
 795                 mutex_unlock(&n->dev.mutex);
 796                 return -EFAULT;
 797         }
 798         n->dev.acked_features = features;
 799         smp_wmb();
 800         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 801                 mutex_lock(&n->vqs[i].mutex);
 802                 n->vqs[i].vhost_hlen = vhost_hlen;
 803                 n->vqs[i].sock_hlen = sock_hlen;
 804                 mutex_unlock(&n->vqs[i].mutex);
 805         }
 806         vhost_net_flush(n);
 807         mutex_unlock(&n->dev.mutex);
 808         return 0;
 809 }
 810
 811 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 812                             unsigned long arg)
 813 {
 814         struct vhost_net *n = f->private_data;
 815         void __user *argp = (void __user *)arg;
 816         u64 __user *featurep = argp;
 817         struct vhost_vring_file backend;
 818         u64 features;
 819         int r;
 820
 821         switch (ioctl) {
 822         case VHOST_NET_SET_BACKEND:
 823                 if (copy_from_user(&backend, argp, sizeof backend))
 824                         return -EFAULT;
 825                 return vhost_net_set_backend(n, backend.index, backend.fd);
 826         case VHOST_GET_FEATURES:
 827                 features = VHOST_NET_FEATURES;
 828                 if (copy_to_user(featurep, &features, sizeof features))
 829                         return -EFAULT;
 830                 return 0;
 831         case VHOST_SET_FEATURES:
 832                 if (copy_from_user(&features, featurep, sizeof features))
 833                         return -EFAULT;
 834                 if (features & ~VHOST_NET_FEATURES)
 835                         return -EOPNOTSUPP;
 836                 return vhost_net_set_features(n, features);
 837         case VHOST_RESET_OWNER:
 838                 return vhost_net_reset_owner(n);
 839         default:
 840                 mutex_lock(&n->dev.mutex);
 841                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 842                 vhost_net_flush(n);
 843                 mutex_unlock(&n->dev.mutex);
 844                 return r;
 845         }
 846 }
 847
 848 #ifdef CONFIG_COMPAT
 849 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 850                                    unsigned long arg)
 851 {
 852         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 853 }
 854 #endif
 855
 856 static const struct file_operations vhost_net_fops = {
 857         .owner          = THIS_MODULE,
 858         .release        = vhost_net_release,
 859         .unlocked_ioctl = vhost_net_ioctl,
 860 #ifdef CONFIG_COMPAT
 861         .compat_ioctl   = vhost_net_compat_ioctl,
 862 #endif
 863         .open           = vhost_net_open,
 864         .llseek         = noop_llseek,
 865 };
 866
 867 static struct miscdevice vhost_net_misc = {
 868         .minor = VHOST_NET_MINOR,
 869         .name = "vhost-net",
 870         .fops = &vhost_net_fops,
 871 };
 872
 873 static int vhost_net_init(void)
 874 {
 875         if (experimental_zcopytx)
 876                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
 877         return misc_register(&vhost_net_misc);
 878 }
 879 module_init(vhost_net_init);
 880
 881 static void vhost_net_exit(void)
 882 {
 883         misc_deregister(&vhost_net_misc);
 884 }
 885 module_exit(vhost_net_exit);
 886
 887 MODULE_VERSION("0.0.1");
 888 MODULE_LICENSE("GPL v2");
 889 MODULE_AUTHOR("Michael S. Tsirkin");
 890 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
 891 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
 892 MODULE_ALIAS("devname:vhost-net");