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 = rcu_dereference(vq->private_data);
 131         if (!sock)
 132                 return;
 133
 134         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 135         if (wmem >= sock->sk->sk_sndbuf) {
 136                 mutex_lock(&vq->mutex);
 137                 tx_poll_start(net, sock);
 138                 mutex_unlock(&vq->mutex);
 139                 return;
 140         }
 141
 142         use_mm(net->dev.mm);
 143         mutex_lock(&vq->mutex);
 144         vhost_disable_notify(vq);
 145
 146         if (wmem < sock->sk->sk_sndbuf / 2)
 147                 tx_poll_stop(net);
 148         hdr_size = vq->vhost_hlen;
 149
 150         for (;;) {
 151                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 152                                          ARRAY_SIZE(vq->iov),
 153                                          &out, &in,
 154                                          NULL, NULL);
 155                 /* On error, stop handling until the next kick. */
 156                 if (unlikely(head < 0))
 157                         break;
 158                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 159                 if (head == vq->num) {
 160                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 161                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 162                                 tx_poll_start(net, sock);
 163                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 164                                 break;
 165                         }
 166                         if (unlikely(vhost_enable_notify(vq))) {
 167                                 vhost_disable_notify(vq);
 168                                 continue;
 169                         }
 170                         break;
 171                 }
 172                 if (in) {
 173                         vq_err(vq, "Unexpected descriptor format for TX: "
 174                                "out %d, int %d\n", out, in);
 175                         break;
 176                 }
 177                 /* Skip header. TODO: support TSO. */
 178                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 179                 msg.msg_iovlen = out;
 180                 len = iov_length(vq->iov, out);
 181                 /* Sanity check */
 182                 if (!len) {
 183                         vq_err(vq, "Unexpected header len for TX: "
 184                                "%zd expected %zd\n",
 185                                iov_length(vq->hdr, s), hdr_size);
 186                         break;
 187                 }
 188                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 189                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 190                 if (unlikely(err < 0)) {
 191                         vhost_discard_vq_desc(vq, 1);
 192                         tx_poll_start(net, sock);
 193                         break;
 194                 }
 195                 if (err != len)
 196                         pr_debug("Truncated TX packet: "
 197                                  " len %d != %zd\n", err, len);
 198                 vhost_add_used_and_signal(&net->dev, vq, head, 0);
 199                 total_len += len;
 200                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 201                         vhost_poll_queue(&vq->poll);
 202                         break;
 203                 }
 204         }
 205
 206         mutex_unlock(&vq->mutex);
 207         unuse_mm(net->dev.mm);
 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 >= VHOST_NET_MAX_SG)) {
 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         use_mm(net->dev.mm);
 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         unuse_mm(net->dev.mm);
 392 }
 393
 394 /* Expects to be always run from workqueue - which acts as
 395  * read-size critical section for our kind of RCU. */
 396 static void handle_rx_mergeable(struct vhost_net *net)
 397 {
 398         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 399         unsigned uninitialized_var(in), log;
 400         struct vhost_log *vq_log;
 401         struct msghdr msg = {
 402                 .msg_name = NULL,
 403                 .msg_namelen = 0,
 404                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 405                 .msg_controllen = 0,
 406                 .msg_iov = vq->iov,
 407                 .msg_flags = MSG_DONTWAIT,
 408         };
 409
 410         struct virtio_net_hdr_mrg_rxbuf hdr = {
 411                 .hdr.flags = 0,
 412                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 413         };
 414
 415         size_t total_len = 0;
 416         int err, headcount;
 417         size_t vhost_hlen, sock_hlen;
 418         size_t vhost_len, sock_len;
 419         struct socket *sock = rcu_dereference(vq->private_data);
 420         if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
 421                 return;
 422
 423         use_mm(net->dev.mm);
 424         mutex_lock(&vq->mutex);
 425         vhost_disable_notify(vq);
 426         vhost_hlen = vq->vhost_hlen;
 427         sock_hlen = vq->sock_hlen;
 428
 429         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 430                 vq->log : NULL;
 431
 432         while ((sock_len = peek_head_len(sock->sk))) {
 433                 sock_len += sock_hlen;
 434                 vhost_len = sock_len + vhost_hlen;
 435                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 436                                         &in, vq_log, &log);
 437                 /* On error, stop handling until the next kick. */
 438                 if (unlikely(headcount < 0))
 439                         break;
 440                 /* OK, now we need to know about added descriptors. */
 441                 if (!headcount) {
 442                         if (unlikely(vhost_enable_notify(vq))) {
 443                                 /* They have slipped one in as we were
 444                                  * doing that: check again. */
 445                                 vhost_disable_notify(vq);
 446                                 continue;
 447                         }
 448                         /* Nothing new?  Wait for eventfd to tell us
 449                          * they refilled. */
 450                         break;
 451                 }
 452                 /* We don't need to be notified again. */
 453                 if (unlikely((vhost_hlen)))
 454                         /* Skip header. TODO: support TSO. */
 455                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 456                 else
 457                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 458                          * needed because sendmsg can modify msg_iov. */
 459                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 460                 msg.msg_iovlen = in;
 461                 err = sock->ops->recvmsg(NULL, sock, &msg,
 462                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 463                 /* Userspace might have consumed the packet meanwhile:
 464                  * it's not supposed to do this usually, but might be hard
 465                  * to prevent. Discard data we got (if any) and keep going. */
 466                 if (unlikely(err != sock_len)) {
 467                         pr_debug("Discarded rx packet: "
 468                                  " len %d, expected %zd\n", err, sock_len);
 469                         vhost_discard_vq_desc(vq, headcount);
 470                         continue;
 471                 }
 472                 if (unlikely(vhost_hlen) &&
 473                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 474                                       vhost_hlen)) {
 475                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 476                                vq->iov->iov_base);
 477                         break;
 478                 }
 479                 /* TODO: Should check and handle checksum. */
 480                 if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF) &&
 481                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 482                                       offsetof(typeof(hdr), num_buffers),
 483                                       sizeof hdr.num_buffers)) {
 484                         vq_err(vq, "Failed num_buffers write");
 485                         vhost_discard_vq_desc(vq, headcount);
 486                         break;
 487                 }
 488                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 489                                             headcount);
 490                 if (unlikely(vq_log))
 491                         vhost_log_write(vq, vq_log, log, vhost_len);
 492                 total_len += vhost_len;
 493                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 494                         vhost_poll_queue(&vq->poll);
 495                         break;
 496                 }
 497         }
 498
 499         mutex_unlock(&vq->mutex);
 500         unuse_mm(net->dev.mm);
 501 }
 502
 503 static void handle_rx(struct vhost_net *net)
 504 {
 505         if (vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF))
 506                 handle_rx_mergeable(net);
 507         else
 508                 handle_rx_big(net);
 509 }
 510
 511 static void handle_tx_kick(struct vhost_work *work)
 512 {
 513         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 514                                                   poll.work);
 515         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 516
 517         handle_tx(net);
 518 }
 519
 520 static void handle_rx_kick(struct vhost_work *work)
 521 {
 522         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 523                                                   poll.work);
 524         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 525
 526         handle_rx(net);
 527 }
 528
 529 static void handle_tx_net(struct vhost_work *work)
 530 {
 531         struct vhost_net *net = container_of(work, struct vhost_net,
 532                                              poll[VHOST_NET_VQ_TX].work);
 533         handle_tx(net);
 534 }
 535
 536 static void handle_rx_net(struct vhost_work *work)
 537 {
 538         struct vhost_net *net = container_of(work, struct vhost_net,
 539                                              poll[VHOST_NET_VQ_RX].work);
 540         handle_rx(net);
 541 }
 542
 543 static int vhost_net_open(struct inode *inode, struct file *f)
 544 {
 545         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 546         struct vhost_dev *dev;
 547         int r;
 548
 549         if (!n)
 550                 return -ENOMEM;
 551
 552         dev = &n->dev;
 553         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 554         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 555         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 556         if (r < 0) {
 557                 kfree(n);
 558                 return r;
 559         }
 560
 561         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 562         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 563         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 564
 565         f->private_data = n;
 566
 567         return 0;
 568 }
 569
 570 static void vhost_net_disable_vq(struct vhost_net *n,
 571                                  struct vhost_virtqueue *vq)
 572 {
 573         if (!vq->private_data)
 574                 return;
 575         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 576                 tx_poll_stop(n);
 577                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 578         } else
 579                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 580 }
 581
 582 static void vhost_net_enable_vq(struct vhost_net *n,
 583                                 struct vhost_virtqueue *vq)
 584 {
 585         struct socket *sock = vq->private_data;
 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 = vq->private_data;
 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 = vq->private_data;
 740         if (sock != oldsock) {
 741                 vhost_net_disable_vq(n, vq);
 742                 rcu_assign_pointer(vq->private_data, sock);
 743                 vhost_net_enable_vq(n, vq);
 744         }
 745
 746         mutex_unlock(&vq->mutex);
 747
 748         if (oldsock) {
 749                 vhost_net_flush_vq(n, index);
 750                 fput(oldsock->file);
 751         }
 752
 753         mutex_unlock(&n->dev.mutex);
 754         return 0;
 755
 756 err_vq:
 757         mutex_unlock(&vq->mutex);
 758 err:
 759         mutex_unlock(&n->dev.mutex);
 760         return r;
 761 }
 762
 763 static long vhost_net_reset_owner(struct vhost_net *n)
 764 {
 765         struct socket *tx_sock = NULL;
 766         struct socket *rx_sock = NULL;
 767         long err;
 768         mutex_lock(&n->dev.mutex);
 769         err = vhost_dev_check_owner(&n->dev);
 770         if (err)
 771                 goto done;
 772         vhost_net_stop(n, &tx_sock, &rx_sock);
 773         vhost_net_flush(n);
 774         err = vhost_dev_reset_owner(&n->dev);
 775 done:
 776         mutex_unlock(&n->dev.mutex);
 777         if (tx_sock)
 778                 fput(tx_sock->file);
 779         if (rx_sock)
 780                 fput(rx_sock->file);
 781         return err;
 782 }
 783
 784 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 785 {
 786         size_t vhost_hlen, sock_hlen, hdr_len;
 787         int i;
 788
 789         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 790                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 791                         sizeof(struct virtio_net_hdr);
 792         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 793                 /* vhost provides vnet_hdr */
 794                 vhost_hlen = hdr_len;
 795                 sock_hlen = 0;
 796         } else {
 797                 /* socket provides vnet_hdr */
 798                 vhost_hlen = 0;
 799                 sock_hlen = hdr_len;
 800         }
 801         mutex_lock(&n->dev.mutex);
 802         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 803             !vhost_log_access_ok(&n->dev)) {
 804                 mutex_unlock(&n->dev.mutex);
 805                 return -EFAULT;
 806         }
 807         n->dev.acked_features = features;
 808         smp_wmb();
 809         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 810                 mutex_lock(&n->vqs[i].mutex);
 811                 n->vqs[i].vhost_hlen = vhost_hlen;
 812                 n->vqs[i].sock_hlen = sock_hlen;
 813                 mutex_unlock(&n->vqs[i].mutex);
 814         }
 815         vhost_net_flush(n);
 816         mutex_unlock(&n->dev.mutex);
 817         return 0;
 818 }
 819
 820 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 821                             unsigned long arg)
 822 {
 823         struct vhost_net *n = f->private_data;
 824         void __user *argp = (void __user *)arg;
 825         u64 __user *featurep = argp;
 826         struct vhost_vring_file backend;
 827         u64 features;
 828         int r;
 829         switch (ioctl) {
 830         case VHOST_NET_SET_BACKEND:
 831                 if (copy_from_user(&backend, argp, sizeof backend))
 832                         return -EFAULT;
 833                 return vhost_net_set_backend(n, backend.index, backend.fd);
 834         case VHOST_GET_FEATURES:
 835                 features = VHOST_FEATURES;
 836                 if (copy_to_user(featurep, &features, sizeof features))
 837                         return -EFAULT;
 838                 return 0;
 839         case VHOST_SET_FEATURES:
 840                 if (copy_from_user(&features, featurep, sizeof features))
 841                         return -EFAULT;
 842                 if (features & ~VHOST_FEATURES)
 843                         return -EOPNOTSUPP;
 844                 return vhost_net_set_features(n, features);
 845         case VHOST_RESET_OWNER:
 846                 return vhost_net_reset_owner(n);
 847         default:
 848                 mutex_lock(&n->dev.mutex);
 849                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 850                 vhost_net_flush(n);
 851                 mutex_unlock(&n->dev.mutex);
 852                 return r;
 853         }
 854 }
 855
 856 #ifdef CONFIG_COMPAT
 857 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 858                                    unsigned long arg)
 859 {
 860         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 861 }
 862 #endif
 863
 864 static const struct file_operations vhost_net_fops = {
 865         .owner          = THIS_MODULE,
 866         .release        = vhost_net_release,
 867         .unlocked_ioctl = vhost_net_ioctl,
 868 #ifdef CONFIG_COMPAT
 869         .compat_ioctl   = vhost_net_compat_ioctl,
 870 #endif
 871         .open           = vhost_net_open,
 872 };
 873
 874 static struct miscdevice vhost_net_misc = {
 875         MISC_DYNAMIC_MINOR,
 876         "vhost-net",
 877         &vhost_net_fops,
 878 };
 879
 880 static int vhost_net_init(void)
 881 {
 882         return misc_register(&vhost_net_misc);
 883 }
 884 module_init(vhost_net_init);
 885
 886 static void vhost_net_exit(void)
 887 {
 888         misc_deregister(&vhost_net_misc);
 889 }
 890 module_exit(vhost_net_exit);
 891
 892 MODULE_VERSION("0.0.1");
 893 MODULE_LICENSE("GPL v2");
 894 MODULE_AUTHOR("Michael S. Tsirkin");
 895 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");