drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/rcupdate.h>
  19 #include <linux/file.h>
  20 #include <linux/slab.h>
  21
  22 #include <linux/net.h>
  23 #include <linux/if_packet.h>
  24 #include <linux/if_arp.h>
  25 #include <linux/if_tun.h>
  26 #include <linux/if_macvlan.h>
  27
  28 #include <net/sock.h>
  29
  30 #include "vhost.h"
  31
  32 static int experimental_zcopytx;
  33 module_param(experimental_zcopytx, int, 0444);
  34 MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
  35
  36 /* Max number of bytes transferred before requeueing the job.
  37  * Using this limit prevents one virtqueue from starving others. */
  38 #define VHOST_NET_WEIGHT 0x80000
  39
  40 /* MAX number of TX used buffers for outstanding zerocopy */
  41 #define VHOST_MAX_PEND 128
  42 #define VHOST_GOODCOPY_LEN 256
  43
  44 enum {
  45         VHOST_NET_VQ_RX = 0,
  46         VHOST_NET_VQ_TX = 1,
  47         VHOST_NET_VQ_MAX = 2,
  48 };
  49
  50 enum vhost_net_poll_state {
  51         VHOST_NET_POLL_DISABLED = 0,
  52         VHOST_NET_POLL_STARTED = 1,
  53         VHOST_NET_POLL_STOPPED = 2,
  54 };
  55
  56 struct vhost_net {
  57         struct vhost_dev dev;
  58         struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  59         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  60         /* Tells us whether we are polling a socket for TX.
  61          * We only do this when socket buffer fills up.
  62          * Protected by tx vq lock. */
  63         enum vhost_net_poll_state tx_poll_state;
  64 };
  65
  66 static bool vhost_sock_zcopy(struct socket *sock)
  67 {
  68         return unlikely(experimental_zcopytx) &&
  69                 sock_flag(sock->sk, SOCK_ZEROCOPY);
  70 }
  71
  72 /* Pop first len bytes from iovec. Return number of segments used. */
  73 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
  74                           size_t len, int iov_count)
  75 {
  76         int seg = 0;
  77         size_t size;
  78
  79         while (len && seg < iov_count) {
  80                 size = min(from->iov_len, len);
  81                 to->iov_base = from->iov_base;
  82                 to->iov_len = size;
  83                 from->iov_len -= size;
  84                 from->iov_base += size;
  85                 len -= size;
  86                 ++from;
  87                 ++to;
  88                 ++seg;
  89         }
  90         return seg;
  91 }
  92 /* Copy iovec entries for len bytes from iovec. */
  93 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
  94                            size_t len, int iovcount)
  95 {
  96         int seg = 0;
  97         size_t size;
  98
  99         while (len && seg < iovcount) {
 100                 size = min(from->iov_len, len);
 101                 to->iov_base = from->iov_base;
 102                 to->iov_len = size;
 103                 len -= size;
 104                 ++from;
 105                 ++to;
 106                 ++seg;
 107         }
 108 }
 109
 110 /* Caller must have TX VQ lock */
 111 static void tx_poll_stop(struct vhost_net *net)
 112 {
 113         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
 114                 return;
 115         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 116         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 117 }
 118
 119 /* Caller must have TX VQ lock */
 120 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 121 {
 122         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 123                 return;
 124         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 125         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 126 }
 127
 128 /* Expects to be always run from workqueue - which acts as
 129  * read-size critical section for our kind of RCU. */
 130 static void handle_tx(struct vhost_net *net)
 131 {
 132         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 133         unsigned out, in, s;
 134         int head;
 135         struct msghdr msg = {
 136                 .msg_name = NULL,
 137                 .msg_namelen = 0,
 138                 .msg_control = NULL,
 139                 .msg_controllen = 0,
 140                 .msg_iov = vq->iov,
 141                 .msg_flags = MSG_DONTWAIT,
 142         };
 143         size_t len, total_len = 0;
 144         int err, wmem;
 145         size_t hdr_size;
 146         struct socket *sock;
 147         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 148         bool zcopy;
 149
 150         /* TODO: check that we are running from vhost_worker? */
 151         sock = rcu_dereference_check(vq->private_data, 1);
 152         if (!sock)
 153                 return;
 154
 155         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 156         if (wmem >= sock->sk->sk_sndbuf) {
 157                 mutex_lock(&vq->mutex);
 158                 tx_poll_start(net, sock);
 159                 mutex_unlock(&vq->mutex);
 160                 return;
 161         }
 162
 163         mutex_lock(&vq->mutex);
 164         vhost_disable_notify(&net->dev, vq);
 165
 166         if (wmem < sock->sk->sk_sndbuf / 2)
 167                 tx_poll_stop(net);
 168         hdr_size = vq->vhost_hlen;
 169         zcopy = vhost_sock_zcopy(sock);
 170
 171         for (;;) {
 172                 /* Release DMAs done buffers first */
 173                 if (zcopy)
 174                         vhost_zerocopy_signal_used(vq);
 175
 176                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 177                                          ARRAY_SIZE(vq->iov),
 178                                          &out, &in,
 179                                          NULL, NULL);
 180                 /* On error, stop handling until the next kick. */
 181                 if (unlikely(head < 0))
 182                         break;
 183                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 184                 if (head == vq->num) {
 185                         int num_pends;
 186
 187                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 188                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 189                                 tx_poll_start(net, sock);
 190                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 191                                 break;
 192                         }
 193                         /* If more outstanding DMAs, queue the work.
 194                          * Handle upend_idx wrap around
 195                          */
 196                         num_pends = likely(vq->upend_idx >= vq->done_idx) ?
 197                                     (vq->upend_idx - vq->done_idx) :
 198                                     (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
 199                         if (unlikely(num_pends > VHOST_MAX_PEND)) {
 200                                 tx_poll_start(net, sock);
 201                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 202                                 break;
 203                         }
 204                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 205                                 vhost_disable_notify(&net->dev, vq);
 206                                 continue;
 207                         }
 208                         break;
 209                 }
 210                 if (in) {
 211                         vq_err(vq, "Unexpected descriptor format for TX: "
 212                                "out %d, int %d\n", out, in);
 213                         break;
 214                 }
 215                 /* Skip header. TODO: support TSO. */
 216                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 217                 msg.msg_iovlen = out;
 218                 len = iov_length(vq->iov, out);
 219                 /* Sanity check */
 220                 if (!len) {
 221                         vq_err(vq, "Unexpected header len for TX: "
 222                                "%zd expected %zd\n",
 223                                iov_length(vq->hdr, s), hdr_size);
 224                         break;
 225                 }
 226                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 227                 if (zcopy) {
 228                         vq->heads[vq->upend_idx].id = head;
 229                         if (len < VHOST_GOODCOPY_LEN) {
 230                                 /* copy don't need to wait for DMA done */
 231                                 vq->heads[vq->upend_idx].len =
 232                                                         VHOST_DMA_DONE_LEN;
 233                                 msg.msg_control = NULL;
 234                                 msg.msg_controllen = 0;
 235                                 ubufs = NULL;
 236                         } else {
 237                                 struct ubuf_info *ubuf;
 238                                 ubuf = vq->ubuf_info + vq->upend_idx;
 239
 240                                 vq->heads[vq->upend_idx].len = len;
 241                                 ubuf->callback = vhost_zerocopy_callback;
 242                                 ubuf->arg = 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         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 290         return len;
 291 }
 292
 293 /* This is a multi-buffer version of vhost_get_desc, that works if
 294  *      vq has read descriptors only.
 295  * @vq          - the relevant virtqueue
 296  * @datalen     - data length we'll be reading
 297  * @iovcount    - returned count of io vectors we fill
 298  * @log         - vhost log
 299  * @log_num     - log offset
 300  * @quota       - headcount quota, 1 for big buffer
 301  *      returns number of buffer heads allocated, negative on error
 302  */
 303 static int get_rx_bufs(struct vhost_virtqueue *vq,
 304                        struct vring_used_elem *heads,
 305                        int datalen,
 306                        unsigned *iovcount,
 307                        struct vhost_log *log,
 308                        unsigned *log_num,
 309                        unsigned int quota)
 310 {
 311         unsigned int out, in;
 312         int seg = 0;
 313         int headcount = 0;
 314         unsigned d;
 315         int r, nlogs = 0;
 316
 317         while (datalen > 0 && headcount < quota) {
 318                 if (unlikely(seg >= UIO_MAXIOV)) {
 319                         r = -ENOBUFS;
 320                         goto err;
 321                 }
 322                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 323                                       ARRAY_SIZE(vq->iov) - seg, &out,
 324                                       &in, log, log_num);
 325                 if (d == vq->num) {
 326                         r = 0;
 327                         goto err;
 328                 }
 329                 if (unlikely(out || in <= 0)) {
 330                         vq_err(vq, "unexpected descriptor format for RX: "
 331                                 "out %d, in %d\n", out, in);
 332                         r = -EINVAL;
 333                         goto err;
 334                 }
 335                 if (unlikely(log)) {
 336                         nlogs += *log_num;
 337                         log += *log_num;
 338                 }
 339                 heads[headcount].id = d;
 340                 heads[headcount].len = iov_length(vq->iov + seg, in);
 341                 datalen -= heads[headcount].len;
 342                 ++headcount;
 343                 seg += in;
 344         }
 345         heads[headcount - 1].len += datalen;
 346         *iovcount = seg;
 347         if (unlikely(log))
 348                 *log_num = nlogs;
 349
 350         /* Detect overrun */
 351         if (unlikely(datalen > 0)) {
 352                 r = UIO_MAXIOV + 1;
 353                 goto err;
 354         }
 355         return headcount;
 356 err:
 357         vhost_discard_vq_desc(vq, headcount);
 358         return r;
 359 }
 360
 361 /* Expects to be always run from workqueue - which acts as
 362  * read-size critical section for our kind of RCU. */
 363 static void handle_rx(struct vhost_net *net)
 364 {
 365         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 366         unsigned uninitialized_var(in), log;
 367         struct vhost_log *vq_log;
 368         struct msghdr msg = {
 369                 .msg_name = NULL,
 370                 .msg_namelen = 0,
 371                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 372                 .msg_controllen = 0,
 373                 .msg_iov = vq->iov,
 374                 .msg_flags = MSG_DONTWAIT,
 375         };
 376         struct virtio_net_hdr_mrg_rxbuf hdr = {
 377                 .hdr.flags = 0,
 378                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 379         };
 380         size_t total_len = 0;
 381         int err, mergeable;
 382         s16 headcount;
 383         size_t vhost_hlen, sock_hlen;
 384         size_t vhost_len, sock_len;
 385         /* TODO: check that we are running from vhost_worker? */
 386         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 387
 388         if (!sock)
 389                 return;
 390
 391         mutex_lock(&vq->mutex);
 392         vhost_disable_notify(&net->dev, vq);
 393         vhost_hlen = vq->vhost_hlen;
 394         sock_hlen = vq->sock_hlen;
 395
 396         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 397                 vq->log : NULL;
 398         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 399
 400         while ((sock_len = peek_head_len(sock->sk))) {
 401                 sock_len += sock_hlen;
 402                 vhost_len = sock_len + vhost_hlen;
 403                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 404                                         &in, vq_log, &log,
 405                                         likely(mergeable) ? UIO_MAXIOV : 1);
 406                 /* On error, stop handling until the next kick. */
 407                 if (unlikely(headcount < 0))
 408                         break;
 409                 /* On overrun, truncate and discard */
 410                 if (unlikely(headcount > UIO_MAXIOV)) {
 411                         msg.msg_iovlen = 1;
 412                         err = sock->ops->recvmsg(NULL, sock, &msg,
 413                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 414                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 415                         continue;
 416                 }
 417                 /* OK, now we need to know about added descriptors. */
 418                 if (!headcount) {
 419                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 420                                 /* They have slipped one in as we were
 421                                  * doing that: check again. */
 422                                 vhost_disable_notify(&net->dev, vq);
 423                                 continue;
 424                         }
 425                         /* Nothing new?  Wait for eventfd to tell us
 426                          * they refilled. */
 427                         break;
 428                 }
 429                 /* We don't need to be notified again. */
 430                 if (unlikely((vhost_hlen)))
 431                         /* Skip header. TODO: support TSO. */
 432                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 433                 else
 434                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 435                          * needed because recvmsg can modify msg_iov. */
 436                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 437                 msg.msg_iovlen = in;
 438                 err = sock->ops->recvmsg(NULL, sock, &msg,
 439                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 440                 /* Userspace might have consumed the packet meanwhile:
 441                  * it's not supposed to do this usually, but might be hard
 442                  * to prevent. Discard data we got (if any) and keep going. */
 443                 if (unlikely(err != sock_len)) {
 444                         pr_debug("Discarded rx packet: "
 445                                  " len %d, expected %zd\n", err, sock_len);
 446                         vhost_discard_vq_desc(vq, headcount);
 447                         continue;
 448                 }
 449                 if (unlikely(vhost_hlen) &&
 450                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 451                                       vhost_hlen)) {
 452                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 453                                vq->iov->iov_base);
 454                         break;
 455                 }
 456                 /* TODO: Should check and handle checksum. */
 457                 if (likely(mergeable) &&
 458                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 459                                       offsetof(typeof(hdr), num_buffers),
 460                                       sizeof hdr.num_buffers)) {
 461                         vq_err(vq, "Failed num_buffers write");
 462                         vhost_discard_vq_desc(vq, headcount);
 463                         break;
 464                 }
 465                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 466                                             headcount);
 467                 if (unlikely(vq_log))
 468                         vhost_log_write(vq, vq_log, log, vhost_len);
 469                 total_len += vhost_len;
 470                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 471                         vhost_poll_queue(&vq->poll);
 472                         break;
 473                 }
 474         }
 475
 476         mutex_unlock(&vq->mutex);
 477 }
 478
 479 static void handle_tx_kick(struct vhost_work *work)
 480 {
 481         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 482                                                   poll.work);
 483         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 484
 485         handle_tx(net);
 486 }
 487
 488 static void handle_rx_kick(struct vhost_work *work)
 489 {
 490         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 491                                                   poll.work);
 492         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 493
 494         handle_rx(net);
 495 }
 496
 497 static void handle_tx_net(struct vhost_work *work)
 498 {
 499         struct vhost_net *net = container_of(work, struct vhost_net,
 500                                              poll[VHOST_NET_VQ_TX].work);
 501         handle_tx(net);
 502 }
 503
 504 static void handle_rx_net(struct vhost_work *work)
 505 {
 506         struct vhost_net *net = container_of(work, struct vhost_net,
 507                                              poll[VHOST_NET_VQ_RX].work);
 508         handle_rx(net);
 509 }
 510
 511 static int vhost_net_open(struct inode *inode, struct file *f)
 512 {
 513         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 514         struct vhost_dev *dev;
 515         int r;
 516
 517         if (!n)
 518                 return -ENOMEM;
 519
 520         dev = &n->dev;
 521         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 522         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 523         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 524         if (r < 0) {
 525                 kfree(n);
 526                 return r;
 527         }
 528
 529         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 530         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 531         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 532
 533         f->private_data = n;
 534
 535         return 0;
 536 }
 537
 538 static void vhost_net_disable_vq(struct vhost_net *n,
 539                                  struct vhost_virtqueue *vq)
 540 {
 541         if (!vq->private_data)
 542                 return;
 543         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 544                 tx_poll_stop(n);
 545                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 546         } else
 547                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 548 }
 549
 550 static void vhost_net_enable_vq(struct vhost_net *n,
 551                                 struct vhost_virtqueue *vq)
 552 {
 553         struct socket *sock;
 554
 555         sock = rcu_dereference_protected(vq->private_data,
 556                                          lockdep_is_held(&vq->mutex));
 557         if (!sock)
 558                 return;
 559         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 560                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 561                 tx_poll_start(n, sock);
 562         } else
 563                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 564 }
 565
 566 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 567                                         struct vhost_virtqueue *vq)
 568 {
 569         struct socket *sock;
 570
 571         mutex_lock(&vq->mutex);
 572         sock = rcu_dereference_protected(vq->private_data,
 573                                          lockdep_is_held(&vq->mutex));
 574         vhost_net_disable_vq(n, vq);
 575         rcu_assign_pointer(vq->private_data, NULL);
 576         mutex_unlock(&vq->mutex);
 577         return sock;
 578 }
 579
 580 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 581                            struct socket **rx_sock)
 582 {
 583         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 584         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 585 }
 586
 587 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 588 {
 589         vhost_poll_flush(n->poll + index);
 590         vhost_poll_flush(&n->dev.vqs[index].poll);
 591 }
 592
 593 static void vhost_net_flush(struct vhost_net *n)
 594 {
 595         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 596         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 597 }
 598
 599 static int vhost_net_release(struct inode *inode, struct file *f)
 600 {
 601         struct vhost_net *n = f->private_data;
 602         struct socket *tx_sock;
 603         struct socket *rx_sock;
 604
 605         vhost_net_stop(n, &tx_sock, &rx_sock);
 606         vhost_net_flush(n);
 607         vhost_dev_cleanup(&n->dev);
 608         if (tx_sock)
 609                 fput(tx_sock->file);
 610         if (rx_sock)
 611                 fput(rx_sock->file);
 612         /* We do an extra flush before freeing memory,
 613          * since jobs can re-queue themselves. */
 614         vhost_net_flush(n);
 615         kfree(n);
 616         return 0;
 617 }
 618
 619 static struct socket *get_raw_socket(int fd)
 620 {
 621         struct {
 622                 struct sockaddr_ll sa;
 623                 char  buf[MAX_ADDR_LEN];
 624         } uaddr;
 625         int uaddr_len = sizeof uaddr, r;
 626         struct socket *sock = sockfd_lookup(fd, &r);
 627
 628         if (!sock)
 629                 return ERR_PTR(-ENOTSOCK);
 630
 631         /* Parameter checking */
 632         if (sock->sk->sk_type != SOCK_RAW) {
 633                 r = -ESOCKTNOSUPPORT;
 634                 goto err;
 635         }
 636
 637         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 638                                &uaddr_len, 0);
 639         if (r)
 640                 goto err;
 641
 642         if (uaddr.sa.sll_family != AF_PACKET) {
 643                 r = -EPFNOSUPPORT;
 644                 goto err;
 645         }
 646         return sock;
 647 err:
 648         fput(sock->file);
 649         return ERR_PTR(r);
 650 }
 651
 652 static struct socket *get_tap_socket(int fd)
 653 {
 654         struct file *file = fget(fd);
 655         struct socket *sock;
 656
 657         if (!file)
 658                 return ERR_PTR(-EBADF);
 659         sock = tun_get_socket(file);
 660         if (!IS_ERR(sock))
 661                 return sock;
 662         sock = macvtap_get_socket(file);
 663         if (IS_ERR(sock))
 664                 fput(file);
 665         return sock;
 666 }
 667
 668 static struct socket *get_socket(int fd)
 669 {
 670         struct socket *sock;
 671
 672         /* special case to disable backend */
 673         if (fd == -1)
 674                 return NULL;
 675         sock = get_raw_socket(fd);
 676         if (!IS_ERR(sock))
 677                 return sock;
 678         sock = get_tap_socket(fd);
 679         if (!IS_ERR(sock))
 680                 return sock;
 681         return ERR_PTR(-ENOTSOCK);
 682 }
 683
 684 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 685 {
 686         struct socket *sock, *oldsock;
 687         struct vhost_virtqueue *vq;
 688         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 689         int r;
 690
 691         mutex_lock(&n->dev.mutex);
 692         r = vhost_dev_check_owner(&n->dev);
 693         if (r)
 694                 goto err;
 695
 696         if (index >= VHOST_NET_VQ_MAX) {
 697                 r = -ENOBUFS;
 698                 goto err;
 699         }
 700         vq = n->vqs + index;
 701         mutex_lock(&vq->mutex);
 702
 703         /* Verify that ring has been setup correctly. */
 704         if (!vhost_vq_access_ok(vq)) {
 705                 r = -EFAULT;
 706                 goto err_vq;
 707         }
 708         sock = get_socket(fd);
 709         if (IS_ERR(sock)) {
 710                 r = PTR_ERR(sock);
 711                 goto err_vq;
 712         }
 713
 714         /* start polling new socket */
 715         oldsock = rcu_dereference_protected(vq->private_data,
 716                                             lockdep_is_held(&vq->mutex));
 717         if (sock != oldsock) {
 718                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 719                 if (IS_ERR(ubufs)) {
 720                         r = PTR_ERR(ubufs);
 721                         goto err_ubufs;
 722                 }
 723                 oldubufs = vq->ubufs;
 724                 vq->ubufs = ubufs;
 725                 vhost_net_disable_vq(n, vq);
 726                 rcu_assign_pointer(vq->private_data, sock);
 727                 vhost_net_enable_vq(n, vq);
 728
 729                 r = vhost_init_used(vq);
 730                 if (r)
 731                         goto err_vq;
 732         }
 733
 734         mutex_unlock(&vq->mutex);
 735
 736         if (oldubufs) {
 737                 vhost_ubuf_put_and_wait(oldubufs);
 738                 mutex_lock(&vq->mutex);
 739                 vhost_zerocopy_signal_used(vq);
 740                 mutex_unlock(&vq->mutex);
 741         }
 742
 743         if (oldsock) {
 744                 vhost_net_flush_vq(n, index);
 745                 fput(oldsock->file);
 746         }
 747
 748         mutex_unlock(&n->dev.mutex);
 749         return 0;
 750
 751 err_ubufs:
 752         fput(sock->file);
 753 err_vq:
 754         mutex_unlock(&vq->mutex);
 755 err:
 756         mutex_unlock(&n->dev.mutex);
 757         return r;
 758 }
 759
 760 static long vhost_net_reset_owner(struct vhost_net *n)
 761 {
 762         struct socket *tx_sock = NULL;
 763         struct socket *rx_sock = NULL;
 764         long err;
 765
 766         mutex_lock(&n->dev.mutex);
 767         err = vhost_dev_check_owner(&n->dev);
 768         if (err)
 769                 goto done;
 770         vhost_net_stop(n, &tx_sock, &rx_sock);
 771         vhost_net_flush(n);
 772         err = vhost_dev_reset_owner(&n->dev);
 773 done:
 774         mutex_unlock(&n->dev.mutex);
 775         if (tx_sock)
 776                 fput(tx_sock->file);
 777         if (rx_sock)
 778                 fput(rx_sock->file);
 779         return err;
 780 }
 781
 782 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 783 {
 784         size_t vhost_hlen, sock_hlen, hdr_len;
 785         int i;
 786
 787         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 788                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 789                         sizeof(struct virtio_net_hdr);
 790         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 791                 /* vhost provides vnet_hdr */
 792                 vhost_hlen = hdr_len;
 793                 sock_hlen = 0;
 794         } else {
 795                 /* socket provides vnet_hdr */
 796                 vhost_hlen = 0;
 797                 sock_hlen = hdr_len;
 798         }
 799         mutex_lock(&n->dev.mutex);
 800         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 801             !vhost_log_access_ok(&n->dev)) {
 802                 mutex_unlock(&n->dev.mutex);
 803                 return -EFAULT;
 804         }
 805         n->dev.acked_features = features;
 806         smp_wmb();
 807         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 808                 mutex_lock(&n->vqs[i].mutex);
 809                 n->vqs[i].vhost_hlen = vhost_hlen;
 810                 n->vqs[i].sock_hlen = sock_hlen;
 811                 mutex_unlock(&n->vqs[i].mutex);
 812         }
 813         vhost_net_flush(n);
 814         mutex_unlock(&n->dev.mutex);
 815         return 0;
 816 }
 817
 818 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 819                             unsigned long arg)
 820 {
 821         struct vhost_net *n = f->private_data;
 822         void __user *argp = (void __user *)arg;
 823         u64 __user *featurep = argp;
 824         struct vhost_vring_file backend;
 825         u64 features;
 826         int r;
 827
 828         switch (ioctl) {
 829         case VHOST_NET_SET_BACKEND:
 830                 if (copy_from_user(&backend, argp, sizeof backend))
 831                         return -EFAULT;
 832                 return vhost_net_set_backend(n, backend.index, backend.fd);
 833         case VHOST_GET_FEATURES:
 834                 features = VHOST_FEATURES;
 835                 if (copy_to_user(featurep, &features, sizeof features))
 836                         return -EFAULT;
 837                 return 0;
 838         case VHOST_SET_FEATURES:
 839                 if (copy_from_user(&features, featurep, sizeof features))
 840                         return -EFAULT;
 841                 if (features & ~VHOST_FEATURES)
 842                         return -EOPNOTSUPP;
 843                 return vhost_net_set_features(n, features);
 844         case VHOST_RESET_OWNER:
 845                 return vhost_net_reset_owner(n);
 846         default:
 847                 mutex_lock(&n->dev.mutex);
 848                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 849                 vhost_net_flush(n);
 850                 mutex_unlock(&n->dev.mutex);
 851                 return r;
 852         }
 853 }
 854
 855 #ifdef CONFIG_COMPAT
 856 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 857                                    unsigned long arg)
 858 {
 859         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 860 }
 861 #endif
 862
 863 static const struct file_operations vhost_net_fops = {
 864         .owner          = THIS_MODULE,
 865         .release        = vhost_net_release,
 866         .unlocked_ioctl = vhost_net_ioctl,
 867 #ifdef CONFIG_COMPAT
 868         .compat_ioctl   = vhost_net_compat_ioctl,
 869 #endif
 870         .open           = vhost_net_open,
 871         .llseek         = noop_llseek,
 872 };
 873
 874 static struct miscdevice vhost_net_misc = {
 875         MISC_DYNAMIC_MINOR,
 876         "vhost-net",
 877         &vhost_net_fops,
 878 };
 879
 880 static int vhost_net_init(void)
 881 {
 882         if (experimental_zcopytx)
 883                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
 884         return misc_register(&vhost_net_misc);
 885 }
 886 module_init(vhost_net_init);
 887
 888 static void vhost_net_exit(void)
 889 {
 890         misc_deregister(&vhost_net_misc);
 891 }
 892 module_exit(vhost_net_exit);
 893
 894 MODULE_VERSION("0.0.1");
 895 MODULE_LICENSE("GPL v2");
 896 MODULE_AUTHOR("Michael S. Tsirkin");
 897 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");