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/mutex.h>
  16 #include <linux/workqueue.h>
  17 #include <linux/rcupdate.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20
  21 #include <linux/net.h>
  22 #include <linux/if_packet.h>
  23 #include <linux/if_arp.h>
  24 #include <linux/if_tun.h>
  25 #include <linux/if_macvlan.h>
  26
  27 #include <net/sock.h>
  28
  29 #include "vhost.h"
  30
  31 /* Max number of bytes transferred before requeueing the job.
  32  * Using this limit prevents one virtqueue from starving others. */
  33 #define VHOST_NET_WEIGHT 0x80000
  34
  35 enum {
  36         VHOST_NET_VQ_RX = 0,
  37         VHOST_NET_VQ_TX = 1,
  38         VHOST_NET_VQ_MAX = 2,
  39 };
  40
  41 enum vhost_net_poll_state {
  42         VHOST_NET_POLL_DISABLED = 0,
  43         VHOST_NET_POLL_STARTED = 1,
  44         VHOST_NET_POLL_STOPPED = 2,
  45 };
  46
  47 struct vhost_net {
  48         struct vhost_dev dev;
  49         struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  50         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  51         /* Tells us whether we are polling a socket for TX.
  52          * We only do this when socket buffer fills up.
  53          * Protected by tx vq lock. */
  54         enum vhost_net_poll_state tx_poll_state;
  55 };
  56
  57 /* Pop first len bytes from iovec. Return number of segments used. */
  58 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
  59                           size_t len, int iov_count)
  60 {
  61         int seg = 0;
  62         size_t size;
  63         while (len && seg < iov_count) {
  64                 size = min(from->iov_len, len);
  65                 to->iov_base = from->iov_base;
  66                 to->iov_len = size;
  67                 from->iov_len -= size;
  68                 from->iov_base += size;
  69                 len -= size;
  70                 ++from;
  71                 ++to;
  72                 ++seg;
  73         }
  74         return seg;
  75 }
  76 /* Copy iovec entries for len bytes from iovec. */
  77 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
  78                            size_t len, int iovcount)
  79 {
  80         int seg = 0;
  81         size_t size;
  82         while (len && seg < iovcount) {
  83                 size = min(from->iov_len, len);
  84                 to->iov_base = from->iov_base;
  85                 to->iov_len = size;
  86                 len -= size;
  87                 ++from;
  88                 ++to;
  89                 ++seg;
  90         }
  91 }
  92
  93 /* Caller must have TX VQ lock */
  94 static void tx_poll_stop(struct vhost_net *net)
  95 {
  96         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
  97                 return;
  98         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
  99         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 100 }
 101
 102 /* Caller must have TX VQ lock */
 103 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 104 {
 105         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 106                 return;
 107         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 108         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 109 }
 110
 111 /* Expects to be always run from workqueue - which acts as
 112  * read-size critical section for our kind of RCU. */
 113 static void handle_tx(struct vhost_net *net)
 114 {
 115         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 116         unsigned out, in, s;
 117         int head;
 118         struct msghdr msg = {
 119                 .msg_name = NULL,
 120                 .msg_namelen = 0,
 121                 .msg_control = NULL,
 122                 .msg_controllen = 0,
 123                 .msg_iov = vq->iov,
 124                 .msg_flags = MSG_DONTWAIT,
 125         };
 126         size_t len, total_len = 0;
 127         int err, wmem;
 128         size_t hdr_size;
 129         struct socket *sock;
 130
 131         /* TODO: check that we are running from vhost_worker?
 132          * Not sure it's worth it, it's straight-forward enough. */
 133         sock = rcu_dereference_check(vq->private_data, 1);
 134         if (!sock)
 135                 return;
 136
 137         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 138         if (wmem >= sock->sk->sk_sndbuf) {
 139                 mutex_lock(&vq->mutex);
 140                 tx_poll_start(net, sock);
 141                 mutex_unlock(&vq->mutex);
 142                 return;
 143         }
 144
 145         mutex_lock(&vq->mutex);
 146         vhost_disable_notify(vq);
 147
 148         if (wmem < sock->sk->sk_sndbuf / 2)
 149                 tx_poll_stop(net);
 150         hdr_size = vq->vhost_hlen;
 151
 152         for (;;) {
 153                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 154                                          ARRAY_SIZE(vq->iov),
 155                                          &out, &in,
 156                                          NULL, NULL);
 157                 /* On error, stop handling until the next kick. */
 158                 if (unlikely(head < 0))
 159                         break;
 160                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 161                 if (head == vq->num) {
 162                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 163                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 164                                 tx_poll_start(net, sock);
 165                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 166                                 break;
 167                         }
 168                         if (unlikely(vhost_enable_notify(vq))) {
 169                                 vhost_disable_notify(vq);
 170                                 continue;
 171                         }
 172                         break;
 173                 }
 174                 if (in) {
 175                         vq_err(vq, "Unexpected descriptor format for TX: "
 176                                "out %d, int %d\n", out, in);
 177                         break;
 178                 }
 179                 /* Skip header. TODO: support TSO. */
 180                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 181                 msg.msg_iovlen = out;
 182                 len = iov_length(vq->iov, out);
 183                 /* Sanity check */
 184                 if (!len) {
 185                         vq_err(vq, "Unexpected header len for TX: "
 186                                "%zd expected %zd\n",
 187                                iov_length(vq->hdr, s), hdr_size);
 188                         break;
 189                 }
 190                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 191                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 192                 if (unlikely(err < 0)) {
 193                         vhost_discard_vq_desc(vq, 1);
 194                         tx_poll_start(net, sock);
 195                         break;
 196                 }
 197                 if (err != len)
 198                         pr_debug("Truncated TX packet: "
 199                                  " len %d != %zd\n", err, len);
 200                 vhost_add_used_and_signal(&net->dev, vq, head, 0);
 201                 total_len += len;
 202                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 203                         vhost_poll_queue(&vq->poll);
 204                         break;
 205                 }
 206         }
 207
 208         mutex_unlock(&vq->mutex);
 209 }
 210
 211 static int peek_head_len(struct sock *sk)
 212 {
 213         struct sk_buff *head;
 214         int len = 0;
 215
 216         lock_sock(sk);
 217         head = skb_peek(&sk->sk_receive_queue);
 218         if (head)
 219                 len = head->len;
 220         release_sock(sk);
 221         return len;
 222 }
 223
 224 /* This is a multi-buffer version of vhost_get_desc, that works if
 225  *      vq has read descriptors only.
 226  * @vq          - the relevant virtqueue
 227  * @datalen     - data length we'll be reading
 228  * @iovcount    - returned count of io vectors we fill
 229  * @log         - vhost log
 230  * @log_num     - log offset
 231  *      returns number of buffer heads allocated, negative on error
 232  */
 233 static int get_rx_bufs(struct vhost_virtqueue *vq,
 234                        struct vring_used_elem *heads,
 235                        int datalen,
 236                        unsigned *iovcount,
 237                        struct vhost_log *log,
 238                        unsigned *log_num)
 239 {
 240         unsigned int out, in;
 241         int seg = 0;
 242         int headcount = 0;
 243         unsigned d;
 244         int r, nlogs = 0;
 245
 246         while (datalen > 0) {
 247                 if (unlikely(seg >= UIO_MAXIOV)) {
 248                         r = -ENOBUFS;
 249                         goto err;
 250                 }
 251                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 252                                       ARRAY_SIZE(vq->iov) - seg, &out,
 253                                       &in, log, log_num);
 254                 if (d == vq->num) {
 255                         r = 0;
 256                         goto err;
 257                 }
 258                 if (unlikely(out || in <= 0)) {
 259                         vq_err(vq, "unexpected descriptor format for RX: "
 260                                 "out %d, in %d\n", out, in);
 261                         r = -EINVAL;
 262                         goto err;
 263                 }
 264                 if (unlikely(log)) {
 265                         nlogs += *log_num;
 266                         log += *log_num;
 267                 }
 268                 heads[headcount].id = d;
 269                 heads[headcount].len = iov_length(vq->iov + seg, in);
 270                 datalen -= heads[headcount].len;
 271                 ++headcount;
 272                 seg += in;
 273         }
 274         heads[headcount - 1].len += datalen;
 275         *iovcount = seg;
 276         if (unlikely(log))
 277                 *log_num = nlogs;
 278         return headcount;
 279 err:
 280         vhost_discard_vq_desc(vq, headcount);
 281         return r;
 282 }
 283
 284 /* Expects to be always run from workqueue - which acts as
 285  * read-size critical section for our kind of RCU. */
 286 static void handle_rx_big(struct vhost_net *net)
 287 {
 288         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 289         unsigned out, in, log, s;
 290         int head;
 291         struct vhost_log *vq_log;
 292         struct msghdr msg = {
 293                 .msg_name = NULL,
 294                 .msg_namelen = 0,
 295                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 296                 .msg_controllen = 0,
 297                 .msg_iov = vq->iov,
 298                 .msg_flags = MSG_DONTWAIT,
 299         };
 300
 301         struct virtio_net_hdr hdr = {
 302                 .flags = 0,
 303                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 304         };
 305
 306         size_t len, total_len = 0;
 307         int err;
 308         size_t hdr_size;
 309         struct socket *sock = rcu_dereference(vq->private_data);
 310         if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
 311                 return;
 312
 313         mutex_lock(&vq->mutex);
 314         vhost_disable_notify(vq);
 315         hdr_size = vq->vhost_hlen;
 316
 317         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 318                 vq->log : NULL;
 319
 320         for (;;) {
 321                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 322                                          ARRAY_SIZE(vq->iov),
 323                                          &out, &in,
 324                                          vq_log, &log);
 325                 /* On error, stop handling until the next kick. */
 326                 if (unlikely(head < 0))
 327                         break;
 328                 /* OK, now we need to know about added descriptors. */
 329                 if (head == vq->num) {
 330                         if (unlikely(vhost_enable_notify(vq))) {
 331                                 /* They have slipped one in as we were
 332                                  * doing that: check again. */
 333                                 vhost_disable_notify(vq);
 334                                 continue;
 335                         }
 336                         /* Nothing new?  Wait for eventfd to tell us
 337                          * they refilled. */
 338                         break;
 339                 }
 340                 /* We don't need to be notified again. */
 341                 if (out) {
 342                         vq_err(vq, "Unexpected descriptor format for RX: "
 343                                "out %d, int %d\n",
 344                                out, in);
 345                         break;
 346                 }
 347                 /* Skip header. TODO: support TSO/mergeable rx buffers. */
 348                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, in);
 349                 msg.msg_iovlen = in;
 350                 len = iov_length(vq->iov, in);
 351                 /* Sanity check */
 352                 if (!len) {
 353                         vq_err(vq, "Unexpected header len for RX: "
 354                                "%zd expected %zd\n",
 355                                iov_length(vq->hdr, s), hdr_size);
 356                         break;
 357                 }
 358                 err = sock->ops->recvmsg(NULL, sock, &msg,
 359                                          len, MSG_DONTWAIT | MSG_TRUNC);
 360                 /* TODO: Check specific error and bomb out unless EAGAIN? */
 361                 if (err < 0) {
 362                         vhost_discard_vq_desc(vq, 1);
 363                         break;
 364                 }
 365                 /* TODO: Should check and handle checksum. */
 366                 if (err > len) {
 367                         pr_debug("Discarded truncated rx packet: "
 368                                  " len %d > %zd\n", err, len);
 369                         vhost_discard_vq_desc(vq, 1);
 370                         continue;
 371                 }
 372                 len = err;
 373                 err = memcpy_toiovec(vq->hdr, (unsigned char *)&hdr, hdr_size);
 374                 if (err) {
 375                         vq_err(vq, "Unable to write vnet_hdr at addr %p: %d\n",
 376                                vq->iov->iov_base, err);
 377                         break;
 378                 }
 379                 len += hdr_size;
 380                 vhost_add_used_and_signal(&net->dev, vq, head, len);
 381                 if (unlikely(vq_log))
 382                         vhost_log_write(vq, vq_log, log, len);
 383                 total_len += len;
 384                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 385                         vhost_poll_queue(&vq->poll);
 386                         break;
 387                 }
 388         }
 389
 390         mutex_unlock(&vq->mutex);
 391 }
 392
 393 /* Expects to be always run from workqueue - which acts as
 394  * read-size critical section for our kind of RCU. */
 395 static void handle_rx_mergeable(struct vhost_net *net)
 396 {
 397         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 398         unsigned uninitialized_var(in), log;
 399         struct vhost_log *vq_log;
 400         struct msghdr msg = {
 401                 .msg_name = NULL,
 402                 .msg_namelen = 0,
 403                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 404                 .msg_controllen = 0,
 405                 .msg_iov = vq->iov,
 406                 .msg_flags = MSG_DONTWAIT,
 407         };
 408
 409         struct virtio_net_hdr_mrg_rxbuf hdr = {
 410                 .hdr.flags = 0,
 411                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 412         };
 413
 414         size_t total_len = 0;
 415         int err, headcount;
 416         size_t vhost_hlen, sock_hlen;
 417         size_t vhost_len, sock_len;
 418         struct socket *sock = rcu_dereference(vq->private_data);
 419         if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
 420                 return;
 421
 422         mutex_lock(&vq->mutex);
 423         vhost_disable_notify(vq);
 424         vhost_hlen = vq->vhost_hlen;
 425         sock_hlen = vq->sock_hlen;
 426
 427         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 428                 vq->log : NULL;
 429
 430         while ((sock_len = peek_head_len(sock->sk))) {
 431                 sock_len += sock_hlen;
 432                 vhost_len = sock_len + vhost_hlen;
 433                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 434                                         &in, vq_log, &log);
 435                 /* On error, stop handling until the next kick. */
 436                 if (unlikely(headcount < 0))
 437                         break;
 438                 /* OK, now we need to know about added descriptors. */
 439                 if (!headcount) {
 440                         if (unlikely(vhost_enable_notify(vq))) {
 441                                 /* They have slipped one in as we were
 442                                  * doing that: check again. */
 443                                 vhost_disable_notify(vq);
 444                                 continue;
 445                         }
 446                         /* Nothing new?  Wait for eventfd to tell us
 447                          * they refilled. */
 448                         break;
 449                 }
 450                 /* We don't need to be notified again. */
 451                 if (unlikely((vhost_hlen)))
 452                         /* Skip header. TODO: support TSO. */
 453                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 454                 else
 455                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 456                          * needed because recvmsg can modify msg_iov. */
 457                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 458                 msg.msg_iovlen = in;
 459                 err = sock->ops->recvmsg(NULL, sock, &msg,
 460                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 461                 /* Userspace might have consumed the packet meanwhile:
 462                  * it's not supposed to do this usually, but might be hard
 463                  * to prevent. Discard data we got (if any) and keep going. */
 464                 if (unlikely(err != sock_len)) {
 465                         pr_debug("Discarded rx packet: "
 466                                  " len %d, expected %zd\n", err, sock_len);
 467                         vhost_discard_vq_desc(vq, headcount);
 468                         continue;
 469                 }
 470                 if (unlikely(vhost_hlen) &&
 471                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 472                                       vhost_hlen)) {
 473                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 474                                vq->iov->iov_base);
 475                         break;
 476                 }
 477                 /* TODO: Should check and handle checksum. */
 478                 if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF) &&
 479                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 480                                       offsetof(typeof(hdr), num_buffers),
 481                                       sizeof hdr.num_buffers)) {
 482                         vq_err(vq, "Failed num_buffers write");
 483                         vhost_discard_vq_desc(vq, headcount);
 484                         break;
 485                 }
 486                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 487                                             headcount);
 488                 if (unlikely(vq_log))
 489                         vhost_log_write(vq, vq_log, log, vhost_len);
 490                 total_len += vhost_len;
 491                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 492                         vhost_poll_queue(&vq->poll);
 493                         break;
 494                 }
 495         }
 496
 497         mutex_unlock(&vq->mutex);
 498 }
 499
 500 static void handle_rx(struct vhost_net *net)
 501 {
 502         if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF))
 503                 handle_rx_mergeable(net);
 504         else
 505                 handle_rx_big(net);
 506 }
 507
 508 static void handle_tx_kick(struct vhost_work *work)
 509 {
 510         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 511                                                   poll.work);
 512         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 513
 514         handle_tx(net);
 515 }
 516
 517 static void handle_rx_kick(struct vhost_work *work)
 518 {
 519         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 520                                                   poll.work);
 521         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 522
 523         handle_rx(net);
 524 }
 525
 526 static void handle_tx_net(struct vhost_work *work)
 527 {
 528         struct vhost_net *net = container_of(work, struct vhost_net,
 529                                              poll[VHOST_NET_VQ_TX].work);
 530         handle_tx(net);
 531 }
 532
 533 static void handle_rx_net(struct vhost_work *work)
 534 {
 535         struct vhost_net *net = container_of(work, struct vhost_net,
 536                                              poll[VHOST_NET_VQ_RX].work);
 537         handle_rx(net);
 538 }
 539
 540 static int vhost_net_open(struct inode *inode, struct file *f)
 541 {
 542         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 543         struct vhost_dev *dev;
 544         int r;
 545
 546         if (!n)
 547                 return -ENOMEM;
 548
 549         dev = &n->dev;
 550         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 551         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 552         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 553         if (r < 0) {
 554                 kfree(n);
 555                 return r;
 556         }
 557
 558         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 559         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 560         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 561
 562         f->private_data = n;
 563
 564         return 0;
 565 }
 566
 567 static void vhost_net_disable_vq(struct vhost_net *n,
 568                                  struct vhost_virtqueue *vq)
 569 {
 570         if (!vq->private_data)
 571                 return;
 572         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 573                 tx_poll_stop(n);
 574                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 575         } else
 576                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 577 }
 578
 579 static void vhost_net_enable_vq(struct vhost_net *n,
 580                                 struct vhost_virtqueue *vq)
 581 {
 582         struct socket *sock;
 583
 584         sock = rcu_dereference_protected(vq->private_data,
 585                                          lockdep_is_held(&vq->mutex));
 586         if (!sock)
 587                 return;
 588         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 589                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 590                 tx_poll_start(n, sock);
 591         } else
 592                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 593 }
 594
 595 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 596                                         struct vhost_virtqueue *vq)
 597 {
 598         struct socket *sock;
 599
 600         mutex_lock(&vq->mutex);
 601         sock = rcu_dereference_protected(vq->private_data,
 602                                          lockdep_is_held(&vq->mutex));
 603         vhost_net_disable_vq(n, vq);
 604         rcu_assign_pointer(vq->private_data, NULL);
 605         mutex_unlock(&vq->mutex);
 606         return sock;
 607 }
 608
 609 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 610                            struct socket **rx_sock)
 611 {
 612         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 613         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 614 }
 615
 616 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 617 {
 618         vhost_poll_flush(n->poll + index);
 619         vhost_poll_flush(&n->dev.vqs[index].poll);
 620 }
 621
 622 static void vhost_net_flush(struct vhost_net *n)
 623 {
 624         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 625         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 626 }
 627
 628 static int vhost_net_release(struct inode *inode, struct file *f)
 629 {
 630         struct vhost_net *n = f->private_data;
 631         struct socket *tx_sock;
 632         struct socket *rx_sock;
 633
 634         vhost_net_stop(n, &tx_sock, &rx_sock);
 635         vhost_net_flush(n);
 636         vhost_dev_cleanup(&n->dev);
 637         if (tx_sock)
 638                 fput(tx_sock->file);
 639         if (rx_sock)
 640                 fput(rx_sock->file);
 641         /* We do an extra flush before freeing memory,
 642          * since jobs can re-queue themselves. */
 643         vhost_net_flush(n);
 644         kfree(n);
 645         return 0;
 646 }
 647
 648 static struct socket *get_raw_socket(int fd)
 649 {
 650         struct {
 651                 struct sockaddr_ll sa;
 652                 char  buf[MAX_ADDR_LEN];
 653         } uaddr;
 654         int uaddr_len = sizeof uaddr, r;
 655         struct socket *sock = sockfd_lookup(fd, &r);
 656         if (!sock)
 657                 return ERR_PTR(-ENOTSOCK);
 658
 659         /* Parameter checking */
 660         if (sock->sk->sk_type != SOCK_RAW) {
 661                 r = -ESOCKTNOSUPPORT;
 662                 goto err;
 663         }
 664
 665         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 666                                &uaddr_len, 0);
 667         if (r)
 668                 goto err;
 669
 670         if (uaddr.sa.sll_family != AF_PACKET) {
 671                 r = -EPFNOSUPPORT;
 672                 goto err;
 673         }
 674         return sock;
 675 err:
 676         fput(sock->file);
 677         return ERR_PTR(r);
 678 }
 679
 680 static struct socket *get_tap_socket(int fd)
 681 {
 682         struct file *file = fget(fd);
 683         struct socket *sock;
 684         if (!file)
 685                 return ERR_PTR(-EBADF);
 686         sock = tun_get_socket(file);
 687         if (!IS_ERR(sock))
 688                 return sock;
 689         sock = macvtap_get_socket(file);
 690         if (IS_ERR(sock))
 691                 fput(file);
 692         return sock;
 693 }
 694
 695 static struct socket *get_socket(int fd)
 696 {
 697         struct socket *sock;
 698         /* special case to disable backend */
 699         if (fd == -1)
 700                 return NULL;
 701         sock = get_raw_socket(fd);
 702         if (!IS_ERR(sock))
 703                 return sock;
 704         sock = get_tap_socket(fd);
 705         if (!IS_ERR(sock))
 706                 return sock;
 707         return ERR_PTR(-ENOTSOCK);
 708 }
 709
 710 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 711 {
 712         struct socket *sock, *oldsock;
 713         struct vhost_virtqueue *vq;
 714         int r;
 715
 716         mutex_lock(&n->dev.mutex);
 717         r = vhost_dev_check_owner(&n->dev);
 718         if (r)
 719                 goto err;
 720
 721         if (index >= VHOST_NET_VQ_MAX) {
 722                 r = -ENOBUFS;
 723                 goto err;
 724         }
 725         vq = n->vqs + index;
 726         mutex_lock(&vq->mutex);
 727
 728         /* Verify that ring has been setup correctly. */
 729         if (!vhost_vq_access_ok(vq)) {
 730                 r = -EFAULT;
 731                 goto err_vq;
 732         }
 733         sock = get_socket(fd);
 734         if (IS_ERR(sock)) {
 735                 r = PTR_ERR(sock);
 736                 goto err_vq;
 737         }
 738
 739         /* start polling new socket */
 740         oldsock = rcu_dereference_protected(vq->private_data,
 741                                             lockdep_is_held(&vq->mutex));
 742         if (sock != oldsock) {
 743                 vhost_net_disable_vq(n, vq);
 744                 rcu_assign_pointer(vq->private_data, sock);
 745                 vhost_net_enable_vq(n, vq);
 746         }
 747
 748         mutex_unlock(&vq->mutex);
 749
 750         if (oldsock) {
 751                 vhost_net_flush_vq(n, index);
 752                 fput(oldsock->file);
 753         }
 754
 755         mutex_unlock(&n->dev.mutex);
 756         return 0;
 757
 758 err_vq:
 759         mutex_unlock(&vq->mutex);
 760 err:
 761         mutex_unlock(&n->dev.mutex);
 762         return r;
 763 }
 764
 765 static long vhost_net_reset_owner(struct vhost_net *n)
 766 {
 767         struct socket *tx_sock = NULL;
 768         struct socket *rx_sock = NULL;
 769         long err;
 770         mutex_lock(&n->dev.mutex);
 771         err = vhost_dev_check_owner(&n->dev);
 772         if (err)
 773                 goto done;
 774         vhost_net_stop(n, &tx_sock, &rx_sock);
 775         vhost_net_flush(n);
 776         err = vhost_dev_reset_owner(&n->dev);
 777 done:
 778         mutex_unlock(&n->dev.mutex);
 779         if (tx_sock)
 780                 fput(tx_sock->file);
 781         if (rx_sock)
 782                 fput(rx_sock->file);
 783         return err;
 784 }
 785
 786 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 787 {
 788         size_t vhost_hlen, sock_hlen, hdr_len;
 789         int i;
 790
 791         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 792                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 793                         sizeof(struct virtio_net_hdr);
 794         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 795                 /* vhost provides vnet_hdr */
 796                 vhost_hlen = hdr_len;
 797                 sock_hlen = 0;
 798         } else {
 799                 /* socket provides vnet_hdr */
 800                 vhost_hlen = 0;
 801                 sock_hlen = hdr_len;
 802         }
 803         mutex_lock(&n->dev.mutex);
 804         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 805             !vhost_log_access_ok(&n->dev)) {
 806                 mutex_unlock(&n->dev.mutex);
 807                 return -EFAULT;
 808         }
 809         n->dev.acked_features = features;
 810         smp_wmb();
 811         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 812                 mutex_lock(&n->vqs[i].mutex);
 813                 n->vqs[i].vhost_hlen = vhost_hlen;
 814                 n->vqs[i].sock_hlen = sock_hlen;
 815                 mutex_unlock(&n->vqs[i].mutex);
 816         }
 817         vhost_net_flush(n);
 818         mutex_unlock(&n->dev.mutex);
 819         return 0;
 820 }
 821
 822 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 823                             unsigned long arg)
 824 {
 825         struct vhost_net *n = f->private_data;
 826         void __user *argp = (void __user *)arg;
 827         u64 __user *featurep = argp;
 828         struct vhost_vring_file backend;
 829         u64 features;
 830         int r;
 831         switch (ioctl) {
 832         case VHOST_NET_SET_BACKEND:
 833                 if (copy_from_user(&backend, argp, sizeof backend))
 834                         return -EFAULT;
 835                 return vhost_net_set_backend(n, backend.index, backend.fd);
 836         case VHOST_GET_FEATURES:
 837                 features = VHOST_FEATURES;
 838                 if (copy_to_user(featurep, &features, sizeof features))
 839                         return -EFAULT;
 840                 return 0;
 841         case VHOST_SET_FEATURES:
 842                 if (copy_from_user(&features, featurep, sizeof features))
 843                         return -EFAULT;
 844                 if (features & ~VHOST_FEATURES)
 845                         return -EOPNOTSUPP;
 846                 return vhost_net_set_features(n, features);
 847         case VHOST_RESET_OWNER:
 848                 return vhost_net_reset_owner(n);
 849         default:
 850                 mutex_lock(&n->dev.mutex);
 851                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 852                 vhost_net_flush(n);
 853                 mutex_unlock(&n->dev.mutex);
 854                 return r;
 855         }
 856 }
 857
 858 #ifdef CONFIG_COMPAT
 859 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 860                                    unsigned long arg)
 861 {
 862         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 863 }
 864 #endif
 865
 866 static const struct file_operations vhost_net_fops = {
 867         .owner          = THIS_MODULE,
 868         .release        = vhost_net_release,
 869         .unlocked_ioctl = vhost_net_ioctl,
 870 #ifdef CONFIG_COMPAT
 871         .compat_ioctl   = vhost_net_compat_ioctl,
 872 #endif
 873         .open           = vhost_net_open,
 874         .llseek         = noop_llseek,
 875 };
 876
 877 static struct miscdevice vhost_net_misc = {
 878         MISC_DYNAMIC_MINOR,
 879         "vhost-net",
 880         &vhost_net_fops,
 881 };
 882
 883 static int vhost_net_init(void)
 884 {
 885         return misc_register(&vhost_net_misc);
 886 }
 887 module_init(vhost_net_init);
 888
 889 static void vhost_net_exit(void)
 890 {
 891         misc_deregister(&vhost_net_misc);
 892 }
 893 module_exit(vhost_net_exit);
 894
 895 MODULE_VERSION("0.0.1");
 896 MODULE_LICENSE("GPL v2");
 897 MODULE_AUTHOR("Michael S. Tsirkin");
 898 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");