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