drivers/vhost/net.c

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