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/mmu_context.h>
  14 #include <linux/miscdevice.h>
  15 #include <linux/module.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 /* Max number of bytes transferred before requeueing the job.
  33  * Using this limit prevents one virtqueue from starving others. */
  34 #define VHOST_NET_WEIGHT 0x80000
  35
  36 enum {
  37         VHOST_NET_VQ_RX = 0,
  38         VHOST_NET_VQ_TX = 1,
  39         VHOST_NET_VQ_MAX = 2,
  40 };
  41
  42 enum vhost_net_poll_state {
  43         VHOST_NET_POLL_DISABLED = 0,
  44         VHOST_NET_POLL_STARTED = 1,
  45         VHOST_NET_POLL_STOPPED = 2,
  46 };
  47
  48 struct vhost_net {
  49         struct vhost_dev dev;
  50         struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  51         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  52         /* Tells us whether we are polling a socket for TX.
  53          * We only do this when socket buffer fills up.
  54          * Protected by tx vq lock. */
  55         enum vhost_net_poll_state tx_poll_state;
  56 };
  57
  58 /* Pop first len bytes from iovec. Return number of segments used. */
  59 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
  60                           size_t len, int iov_count)
  61 {
  62         int seg = 0;
  63         size_t size;
  64         while (len && seg < iov_count) {
  65                 size = min(from->iov_len, len);
  66                 to->iov_base = from->iov_base;
  67                 to->iov_len = size;
  68                 from->iov_len -= size;
  69                 from->iov_base += size;
  70                 len -= size;
  71                 ++from;
  72                 ++to;
  73                 ++seg;
  74         }
  75         return seg;
  76 }
  77 /* Copy iovec entries for len bytes from iovec. */
  78 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
  79                            size_t len, int iovcount)
  80 {
  81         int seg = 0;
  82         size_t size;
  83         while (len && seg < iovcount) {
  84                 size = min(from->iov_len, len);
  85                 to->iov_base = from->iov_base;
  86                 to->iov_len = size;
  87                 len -= size;
  88                 ++from;
  89                 ++to;
  90                 ++seg;
  91         }
  92 }
  93
  94 /* Caller must have TX VQ lock */
  95 static void tx_poll_stop(struct vhost_net *net)
  96 {
  97         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
  98                 return;
  99         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 100         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 101 }
 102
 103 /* Caller must have TX VQ lock */
 104 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 105 {
 106         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 107                 return;
 108         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 109         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 110 }
 111
 112 /* Expects to be always run from workqueue - which acts as
 113  * read-size critical section for our kind of RCU. */
 114 static void handle_tx(struct vhost_net *net)
 115 {
 116         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 117         unsigned out, in, s;
 118         int head;
 119         struct msghdr msg = {
 120                 .msg_name = NULL,
 121                 .msg_namelen = 0,
 122                 .msg_control = NULL,
 123                 .msg_controllen = 0,
 124                 .msg_iov = vq->iov,
 125                 .msg_flags = MSG_DONTWAIT,
 126         };
 127         size_t len, total_len = 0;
 128         int err, wmem;
 129         size_t hdr_size;
 130         struct socket *sock;
 131
 132         /* TODO: check that we are running from vhost_worker?
 133          * Not sure it's worth it, it's straight-forward enough. */
 134         sock = rcu_dereference_check(vq->private_data, 1);
 135         if (!sock)
 136                 return;
 137
 138         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 139         if (wmem >= sock->sk->sk_sndbuf) {
 140                 mutex_lock(&vq->mutex);
 141                 tx_poll_start(net, sock);
 142                 mutex_unlock(&vq->mutex);
 143                 return;
 144         }
 145
 146         use_mm(net->dev.mm);
 147         mutex_lock(&vq->mutex);
 148         vhost_disable_notify(vq);
 149
 150         if (wmem < sock->sk->sk_sndbuf / 2)
 151                 tx_poll_stop(net);
 152         hdr_size = vq->vhost_hlen;
 153
 154         for (;;) {
 155                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 156                                          ARRAY_SIZE(vq->iov),
 157                                          &out, &in,
 158                                          NULL, NULL);
 159                 /* On error, stop handling until the next kick. */
 160                 if (unlikely(head < 0))
 161                         break;
 162                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 163                 if (head == vq->num) {
 164                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 165                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 166                                 tx_poll_start(net, sock);
 167                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 168                                 break;
 169                         }
 170                         if (unlikely(vhost_enable_notify(vq))) {
 171                                 vhost_disable_notify(vq);
 172                                 continue;
 173                         }
 174                         break;
 175                 }
 176                 if (in) {
 177                         vq_err(vq, "Unexpected descriptor format for TX: "
 178                                "out %d, int %d\n", out, in);
 179                         break;
 180                 }
 181                 /* Skip header. TODO: support TSO. */
 182                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 183                 msg.msg_iovlen = out;
 184                 len = iov_length(vq->iov, out);
 185                 /* Sanity check */
 186                 if (!len) {
 187                         vq_err(vq, "Unexpected header len for TX: "
 188                                "%zd expected %zd\n",
 189                                iov_length(vq->hdr, s), hdr_size);
 190                         break;
 191                 }
 192                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 193                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 194                 if (unlikely(err < 0)) {
 195                         vhost_discard_vq_desc(vq, 1);
 196                         tx_poll_start(net, sock);
 197                         break;
 198                 }
 199                 if (err != len)
 200                         pr_debug("Truncated TX packet: "
 201                                  " len %d != %zd\n", err, len);
 202                 vhost_add_used_and_signal(&net->dev, vq, head, 0);
 203                 total_len += len;
 204                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 205                         vhost_poll_queue(&vq->poll);
 206                         break;
 207                 }
 208         }
 209
 210         mutex_unlock(&vq->mutex);
 211         unuse_mm(net->dev.mm);
 212 }
 213
 214 static int peek_head_len(struct sock *sk)
 215 {
 216         struct sk_buff *head;
 217         int len = 0;
 218
 219         lock_sock(sk);
 220         head = skb_peek(&sk->sk_receive_queue);
 221         if (head)
 222                 len = head->len;
 223         release_sock(sk);
 224         return len;
 225 }
 226
 227 /* This is a multi-buffer version of vhost_get_desc, that works if
 228  *      vq has read descriptors only.
 229  * @vq          - the relevant virtqueue
 230  * @datalen     - data length we'll be reading
 231  * @iovcount    - returned count of io vectors we fill
 232  * @log         - vhost log
 233  * @log_num     - log offset
 234  *      returns number of buffer heads allocated, negative on error
 235  */
 236 static int get_rx_bufs(struct vhost_virtqueue *vq,
 237                        struct vring_used_elem *heads,
 238                        int datalen,
 239                        unsigned *iovcount,
 240                        struct vhost_log *log,
 241                        unsigned *log_num)
 242 {
 243         unsigned int out, in;
 244         int seg = 0;
 245         int headcount = 0;
 246         unsigned d;
 247         int r, nlogs = 0;
 248
 249         while (datalen > 0) {
 250                 if (unlikely(seg >= UIO_MAXIOV)) {
 251                         r = -ENOBUFS;
 252                         goto err;
 253                 }
 254                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 255                                       ARRAY_SIZE(vq->iov) - seg, &out,
 256                                       &in, log, log_num);
 257                 if (d == vq->num) {
 258                         r = 0;
 259                         goto err;
 260                 }
 261                 if (unlikely(out || in <= 0)) {
 262                         vq_err(vq, "unexpected descriptor format for RX: "
 263                                 "out %d, in %d\n", out, in);
 264                         r = -EINVAL;
 265                         goto err;
 266                 }
 267                 if (unlikely(log)) {
 268                         nlogs += *log_num;
 269                         log += *log_num;
 270                 }
 271                 heads[headcount].id = d;
 272                 heads[headcount].len = iov_length(vq->iov + seg, in);
 273                 datalen -= heads[headcount].len;
 274                 ++headcount;
 275                 seg += in;
 276         }
 277         heads[headcount - 1].len += datalen;
 278         *iovcount = seg;
 279         if (unlikely(log))
 280                 *log_num = nlogs;
 281         return headcount;
 282 err:
 283         vhost_discard_vq_desc(vq, headcount);
 284         return r;
 285 }
 286
 287 /* Expects to be always run from workqueue - which acts as
 288  * read-size critical section for our kind of RCU. */
 289 static void handle_rx_big(struct vhost_net *net)
 290 {
 291         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 292         unsigned out, in, log, s;
 293         int head;
 294         struct vhost_log *vq_log;
 295         struct msghdr msg = {
 296                 .msg_name = NULL,
 297                 .msg_namelen = 0,
 298                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 299                 .msg_controllen = 0,
 300                 .msg_iov = vq->iov,
 301                 .msg_flags = MSG_DONTWAIT,
 302         };
 303
 304         struct virtio_net_hdr hdr = {
 305                 .flags = 0,
 306                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 307         };
 308
 309         size_t len, total_len = 0;
 310         int err;
 311         size_t hdr_size;
 312         struct socket *sock = rcu_dereference(vq->private_data);
 313         if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
 314                 return;
 315
 316         use_mm(net->dev.mm);
 317         mutex_lock(&vq->mutex);
 318         vhost_disable_notify(vq);
 319         hdr_size = vq->vhost_hlen;
 320
 321         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 322                 vq->log : NULL;
 323
 324         for (;;) {
 325                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 326                                          ARRAY_SIZE(vq->iov),
 327                                          &out, &in,
 328                                          vq_log, &log);
 329                 /* On error, stop handling until the next kick. */
 330                 if (unlikely(head < 0))
 331                         break;
 332                 /* OK, now we need to know about added descriptors. */
 333                 if (head == vq->num) {
 334                         if (unlikely(vhost_enable_notify(vq))) {
 335                                 /* They have slipped one in as we were
 336                                  * doing that: check again. */
 337                                 vhost_disable_notify(vq);
 338                                 continue;
 339                         }
 340                         /* Nothing new?  Wait for eventfd to tell us
 341                          * they refilled. */
 342                         break;
 343                 }
 344                 /* We don't need to be notified again. */
 345                 if (out) {
 346                         vq_err(vq, "Unexpected descriptor format for RX: "
 347                                "out %d, int %d\n",
 348                                out, in);
 349                         break;
 350                 }
 351                 /* Skip header. TODO: support TSO/mergeable rx buffers. */
 352                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, in);
 353                 msg.msg_iovlen = in;
 354                 len = iov_length(vq->iov, in);
 355                 /* Sanity check */
 356                 if (!len) {
 357                         vq_err(vq, "Unexpected header len for RX: "
 358                                "%zd expected %zd\n",
 359                                iov_length(vq->hdr, s), hdr_size);
 360                         break;
 361                 }
 362                 err = sock->ops->recvmsg(NULL, sock, &msg,
 363                                          len, MSG_DONTWAIT | MSG_TRUNC);
 364                 /* TODO: Check specific error and bomb out unless EAGAIN? */
 365                 if (err < 0) {
 366                         vhost_discard_vq_desc(vq, 1);
 367                         break;
 368                 }
 369                 /* TODO: Should check and handle checksum. */
 370                 if (err > len) {
 371                         pr_debug("Discarded truncated rx packet: "
 372                                  " len %d > %zd\n", err, len);
 373                         vhost_discard_vq_desc(vq, 1);
 374                         continue;
 375                 }
 376                 len = err;
 377                 err = memcpy_toiovec(vq->hdr, (unsigned char *)&hdr, hdr_size);
 378                 if (err) {
 379                         vq_err(vq, "Unable to write vnet_hdr at addr %p: %d\n",
 380                                vq->iov->iov_base, err);
 381                         break;
 382                 }
 383                 len += hdr_size;
 384                 vhost_add_used_and_signal(&net->dev, vq, head, len);
 385                 if (unlikely(vq_log))
 386                         vhost_log_write(vq, vq_log, log, len);
 387                 total_len += len;
 388                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 389                         vhost_poll_queue(&vq->poll);
 390                         break;
 391                 }
 392         }
 393
 394         mutex_unlock(&vq->mutex);
 395         unuse_mm(net->dev.mm);
 396 }
 397
 398 /* Expects to be always run from workqueue - which acts as
 399  * read-size critical section for our kind of RCU. */
 400 static void handle_rx_mergeable(struct vhost_net *net)
 401 {
 402         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 403         unsigned uninitialized_var(in), log;
 404         struct vhost_log *vq_log;
 405         struct msghdr msg = {
 406                 .msg_name = NULL,
 407                 .msg_namelen = 0,
 408                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 409                 .msg_controllen = 0,
 410                 .msg_iov = vq->iov,
 411                 .msg_flags = MSG_DONTWAIT,
 412         };
 413
 414         struct virtio_net_hdr_mrg_rxbuf hdr = {
 415                 .hdr.flags = 0,
 416                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 417         };
 418
 419         size_t total_len = 0;
 420         int err, headcount;
 421         size_t vhost_hlen, sock_hlen;
 422         size_t vhost_len, sock_len;
 423         struct socket *sock = rcu_dereference(vq->private_data);
 424         if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
 425                 return;
 426
 427         use_mm(net->dev.mm);
 428         mutex_lock(&vq->mutex);
 429         vhost_disable_notify(vq);
 430         vhost_hlen = vq->vhost_hlen;
 431         sock_hlen = vq->sock_hlen;
 432
 433         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 434                 vq->log : NULL;
 435
 436         while ((sock_len = peek_head_len(sock->sk))) {
 437                 sock_len += sock_hlen;
 438                 vhost_len = sock_len + vhost_hlen;
 439                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 440                                         &in, vq_log, &log);
 441                 /* On error, stop handling until the next kick. */
 442                 if (unlikely(headcount < 0))
 443                         break;
 444                 /* OK, now we need to know about added descriptors. */
 445                 if (!headcount) {
 446                         if (unlikely(vhost_enable_notify(vq))) {
 447                                 /* They have slipped one in as we were
 448                                  * doing that: check again. */
 449                                 vhost_disable_notify(vq);
 450                                 continue;
 451                         }
 452                         /* Nothing new?  Wait for eventfd to tell us
 453                          * they refilled. */
 454                         break;
 455                 }
 456                 /* We don't need to be notified again. */
 457                 if (unlikely((vhost_hlen)))
 458                         /* Skip header. TODO: support TSO. */
 459                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 460                 else
 461                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 462                          * needed because sendmsg can modify msg_iov. */
 463                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 464                 msg.msg_iovlen = in;
 465                 err = sock->ops->recvmsg(NULL, sock, &msg,
 466                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 467                 /* Userspace might have consumed the packet meanwhile:
 468                  * it's not supposed to do this usually, but might be hard
 469                  * to prevent. Discard data we got (if any) and keep going. */
 470                 if (unlikely(err != sock_len)) {
 471                         pr_debug("Discarded rx packet: "
 472                                  " len %d, expected %zd\n", err, sock_len);
 473                         vhost_discard_vq_desc(vq, headcount);
 474                         continue;
 475                 }
 476                 if (unlikely(vhost_hlen) &&
 477                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 478                                       vhost_hlen)) {
 479                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 480                                vq->iov->iov_base);
 481                         break;
 482                 }
 483                 /* TODO: Should check and handle checksum. */
 484                 if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF) &&
 485                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 486                                       offsetof(typeof(hdr), num_buffers),
 487                                       sizeof hdr.num_buffers)) {
 488                         vq_err(vq, "Failed num_buffers write");
 489                         vhost_discard_vq_desc(vq, headcount);
 490                         break;
 491                 }
 492                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 493                                             headcount);
 494                 if (unlikely(vq_log))
 495                         vhost_log_write(vq, vq_log, log, vhost_len);
 496                 total_len += vhost_len;
 497                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 498                         vhost_poll_queue(&vq->poll);
 499                         break;
 500                 }
 501         }
 502
 503         mutex_unlock(&vq->mutex);
 504         unuse_mm(net->dev.mm);
 505 }
 506
 507 static void handle_rx(struct vhost_net *net)
 508 {
 509         if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF))
 510                 handle_rx_mergeable(net);
 511         else
 512                 handle_rx_big(net);
 513 }
 514
 515 static void handle_tx_kick(struct vhost_work *work)
 516 {
 517         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 518                                                   poll.work);
 519         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 520
 521         handle_tx(net);
 522 }
 523
 524 static void handle_rx_kick(struct vhost_work *work)
 525 {
 526         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 527                                                   poll.work);
 528         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 529
 530         handle_rx(net);
 531 }
 532
 533 static void handle_tx_net(struct vhost_work *work)
 534 {
 535         struct vhost_net *net = container_of(work, struct vhost_net,
 536                                              poll[VHOST_NET_VQ_TX].work);
 537         handle_tx(net);
 538 }
 539
 540 static void handle_rx_net(struct vhost_work *work)
 541 {
 542         struct vhost_net *net = container_of(work, struct vhost_net,
 543                                              poll[VHOST_NET_VQ_RX].work);
 544         handle_rx(net);
 545 }
 546
 547 static int vhost_net_open(struct inode *inode, struct file *f)
 548 {
 549         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 550         struct vhost_dev *dev;
 551         int r;
 552
 553         if (!n)
 554                 return -ENOMEM;
 555
 556         dev = &n->dev;
 557         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 558         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 559         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 560         if (r < 0) {
 561                 kfree(n);
 562                 return r;
 563         }
 564
 565         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 566         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 567         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 568
 569         f->private_data = n;
 570
 571         return 0;
 572 }
 573
 574 static void vhost_net_disable_vq(struct vhost_net *n,
 575                                  struct vhost_virtqueue *vq)
 576 {
 577         if (!vq->private_data)
 578                 return;
 579         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 580                 tx_poll_stop(n);
 581                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 582         } else
 583                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 584 }
 585
 586 static void vhost_net_enable_vq(struct vhost_net *n,
 587                                 struct vhost_virtqueue *vq)
 588 {
 589         struct socket *sock;
 590
 591         sock = rcu_dereference_protected(vq->private_data,
 592                                          lockdep_is_held(&vq->mutex));
 593         if (!sock)
 594                 return;
 595         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 596                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 597                 tx_poll_start(n, sock);
 598         } else
 599                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 600 }
 601
 602 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 603                                         struct vhost_virtqueue *vq)
 604 {
 605         struct socket *sock;
 606
 607         mutex_lock(&vq->mutex);
 608         sock = rcu_dereference_protected(vq->private_data,
 609                                          lockdep_is_held(&vq->mutex));
 610         vhost_net_disable_vq(n, vq);
 611         rcu_assign_pointer(vq->private_data, NULL);
 612         mutex_unlock(&vq->mutex);
 613         return sock;
 614 }
 615
 616 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 617                            struct socket **rx_sock)
 618 {
 619         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 620         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 621 }
 622
 623 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 624 {
 625         vhost_poll_flush(n->poll + index);
 626         vhost_poll_flush(&n->dev.vqs[index].poll);
 627 }
 628
 629 static void vhost_net_flush(struct vhost_net *n)
 630 {
 631         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 632         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 633 }
 634
 635 static int vhost_net_release(struct inode *inode, struct file *f)
 636 {
 637         struct vhost_net *n = f->private_data;
 638         struct socket *tx_sock;
 639         struct socket *rx_sock;
 640
 641         vhost_net_stop(n, &tx_sock, &rx_sock);
 642         vhost_net_flush(n);
 643         vhost_dev_cleanup(&n->dev);
 644         if (tx_sock)
 645                 fput(tx_sock->file);
 646         if (rx_sock)
 647                 fput(rx_sock->file);
 648         /* We do an extra flush before freeing memory,
 649          * since jobs can re-queue themselves. */
 650         vhost_net_flush(n);
 651         kfree(n);
 652         return 0;
 653 }
 654
 655 static struct socket *get_raw_socket(int fd)
 656 {
 657         struct {
 658                 struct sockaddr_ll sa;
 659                 char  buf[MAX_ADDR_LEN];
 660         } uaddr;
 661         int uaddr_len = sizeof uaddr, r;
 662         struct socket *sock = sockfd_lookup(fd, &r);
 663         if (!sock)
 664                 return ERR_PTR(-ENOTSOCK);
 665
 666         /* Parameter checking */
 667         if (sock->sk->sk_type != SOCK_RAW) {
 668                 r = -ESOCKTNOSUPPORT;
 669                 goto err;
 670         }
 671
 672         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 673                                &uaddr_len, 0);
 674         if (r)
 675                 goto err;
 676
 677         if (uaddr.sa.sll_family != AF_PACKET) {
 678                 r = -EPFNOSUPPORT;
 679                 goto err;
 680         }
 681         return sock;
 682 err:
 683         fput(sock->file);
 684         return ERR_PTR(r);
 685 }
 686
 687 static struct socket *get_tap_socket(int fd)
 688 {
 689         struct file *file = fget(fd);
 690         struct socket *sock;
 691         if (!file)
 692                 return ERR_PTR(-EBADF);
 693         sock = tun_get_socket(file);
 694         if (!IS_ERR(sock))
 695                 return sock;
 696         sock = macvtap_get_socket(file);
 697         if (IS_ERR(sock))
 698                 fput(file);
 699         return sock;
 700 }
 701
 702 static struct socket *get_socket(int fd)
 703 {
 704         struct socket *sock;
 705         /* special case to disable backend */
 706         if (fd == -1)
 707                 return NULL;
 708         sock = get_raw_socket(fd);
 709         if (!IS_ERR(sock))
 710                 return sock;
 711         sock = get_tap_socket(fd);
 712         if (!IS_ERR(sock))
 713                 return sock;
 714         return ERR_PTR(-ENOTSOCK);
 715 }
 716
 717 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 718 {
 719         struct socket *sock, *oldsock;
 720         struct vhost_virtqueue *vq;
 721         int r;
 722
 723         mutex_lock(&n->dev.mutex);
 724         r = vhost_dev_check_owner(&n->dev);
 725         if (r)
 726                 goto err;
 727
 728         if (index >= VHOST_NET_VQ_MAX) {
 729                 r = -ENOBUFS;
 730                 goto err;
 731         }
 732         vq = n->vqs + index;
 733         mutex_lock(&vq->mutex);
 734
 735         /* Verify that ring has been setup correctly. */
 736         if (!vhost_vq_access_ok(vq)) {
 737                 r = -EFAULT;
 738                 goto err_vq;
 739         }
 740         sock = get_socket(fd);
 741         if (IS_ERR(sock)) {
 742                 r = PTR_ERR(sock);
 743                 goto err_vq;
 744         }
 745
 746         /* start polling new socket */
 747         oldsock = rcu_dereference_protected(vq->private_data,
 748                                             lockdep_is_held(&vq->mutex));
 749         if (sock != oldsock) {
 750                 vhost_net_disable_vq(n, vq);
 751                 rcu_assign_pointer(vq->private_data, sock);
 752                 vhost_net_enable_vq(n, vq);
 753         }
 754
 755         mutex_unlock(&vq->mutex);
 756
 757         if (oldsock) {
 758                 vhost_net_flush_vq(n, index);
 759                 fput(oldsock->file);
 760         }
 761
 762         mutex_unlock(&n->dev.mutex);
 763         return 0;
 764
 765 err_vq:
 766         mutex_unlock(&vq->mutex);
 767 err:
 768         mutex_unlock(&n->dev.mutex);
 769         return r;
 770 }
 771
 772 static long vhost_net_reset_owner(struct vhost_net *n)
 773 {
 774         struct socket *tx_sock = NULL;
 775         struct socket *rx_sock = NULL;
 776         long err;
 777         mutex_lock(&n->dev.mutex);
 778         err = vhost_dev_check_owner(&n->dev);
 779         if (err)
 780                 goto done;
 781         vhost_net_stop(n, &tx_sock, &rx_sock);
 782         vhost_net_flush(n);
 783         err = vhost_dev_reset_owner(&n->dev);
 784 done:
 785         mutex_unlock(&n->dev.mutex);
 786         if (tx_sock)
 787                 fput(tx_sock->file);
 788         if (rx_sock)
 789                 fput(rx_sock->file);
 790         return err;
 791 }
 792
 793 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 794 {
 795         size_t vhost_hlen, sock_hlen, hdr_len;
 796         int i;
 797
 798         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 799                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 800                         sizeof(struct virtio_net_hdr);
 801         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 802                 /* vhost provides vnet_hdr */
 803                 vhost_hlen = hdr_len;
 804                 sock_hlen = 0;
 805         } else {
 806                 /* socket provides vnet_hdr */
 807                 vhost_hlen = 0;
 808                 sock_hlen = hdr_len;
 809         }
 810         mutex_lock(&n->dev.mutex);
 811         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 812             !vhost_log_access_ok(&n->dev)) {
 813                 mutex_unlock(&n->dev.mutex);
 814                 return -EFAULT;
 815         }
 816         n->dev.acked_features = features;
 817         smp_wmb();
 818         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 819                 mutex_lock(&n->vqs[i].mutex);
 820                 n->vqs[i].vhost_hlen = vhost_hlen;
 821                 n->vqs[i].sock_hlen = sock_hlen;
 822                 mutex_unlock(&n->vqs[i].mutex);
 823         }
 824         vhost_net_flush(n);
 825         mutex_unlock(&n->dev.mutex);
 826         return 0;
 827 }
 828
 829 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 830                             unsigned long arg)
 831 {
 832         struct vhost_net *n = f->private_data;
 833         void __user *argp = (void __user *)arg;
 834         u64 __user *featurep = argp;
 835         struct vhost_vring_file backend;
 836         u64 features;
 837         int r;
 838         switch (ioctl) {
 839         case VHOST_NET_SET_BACKEND:
 840                 if (copy_from_user(&backend, argp, sizeof backend))
 841                         return -EFAULT;
 842                 return vhost_net_set_backend(n, backend.index, backend.fd);
 843         case VHOST_GET_FEATURES:
 844                 features = VHOST_FEATURES;
 845                 if (copy_to_user(featurep, &features, sizeof features))
 846                         return -EFAULT;
 847                 return 0;
 848         case VHOST_SET_FEATURES:
 849                 if (copy_from_user(&features, featurep, sizeof features))
 850                         return -EFAULT;
 851                 if (features & ~VHOST_FEATURES)
 852                         return -EOPNOTSUPP;
 853                 return vhost_net_set_features(n, features);
 854         case VHOST_RESET_OWNER:
 855                 return vhost_net_reset_owner(n);
 856         default:
 857                 mutex_lock(&n->dev.mutex);
 858                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 859                 vhost_net_flush(n);
 860                 mutex_unlock(&n->dev.mutex);
 861                 return r;
 862         }
 863 }
 864
 865 #ifdef CONFIG_COMPAT
 866 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 867                                    unsigned long arg)
 868 {
 869         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 870 }
 871 #endif
 872
 873 static const struct file_operations vhost_net_fops = {
 874         .owner          = THIS_MODULE,
 875         .release        = vhost_net_release,
 876         .unlocked_ioctl = vhost_net_ioctl,
 877 #ifdef CONFIG_COMPAT
 878         .compat_ioctl   = vhost_net_compat_ioctl,
 879 #endif
 880         .open           = vhost_net_open,
 881         .llseek         = noop_llseek,
 882 };
 883
 884 static struct miscdevice vhost_net_misc = {
 885         MISC_DYNAMIC_MINOR,
 886         "vhost-net",
 887         &vhost_net_fops,
 888 };
 889
 890 static int vhost_net_init(void)
 891 {
 892         return misc_register(&vhost_net_misc);
 893 }
 894 module_init(vhost_net_init);
 895
 896 static void vhost_net_exit(void)
 897 {
 898         misc_deregister(&vhost_net_misc);
 899 }
 900 module_exit(vhost_net_exit);
 901
 902 MODULE_VERSION("0.0.1");
 903 MODULE_LICENSE("GPL v2");
 904 MODULE_AUTHOR("Michael S. Tsirkin");
 905 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");