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
  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
  84         while (len && seg < iovcount) {
  85                 size = min(from->iov_len, len);
  86                 to->iov_base = from->iov_base;
  87                 to->iov_len = size;
  88                 len -= size;
  89                 ++from;
  90                 ++to;
  91                 ++seg;
  92         }
  93 }
  94
  95 /* Caller must have TX VQ lock */
  96 static void tx_poll_stop(struct vhost_net *net)
  97 {
  98         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
  99                 return;
 100         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 101         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 102 }
 103
 104 /* Caller must have TX VQ lock */
 105 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 106 {
 107         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 108                 return;
 109         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 110         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 111 }
 112
 113 /* Expects to be always run from workqueue - which acts as
 114  * read-size critical section for our kind of RCU. */
 115 static void handle_tx(struct vhost_net *net)
 116 {
 117         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 118         unsigned out, in, s;
 119         int head;
 120         struct msghdr msg = {
 121                 .msg_name = NULL,
 122                 .msg_namelen = 0,
 123                 .msg_control = NULL,
 124                 .msg_controllen = 0,
 125                 .msg_iov = vq->iov,
 126                 .msg_flags = MSG_DONTWAIT,
 127         };
 128         size_t len, total_len = 0;
 129         int err, wmem;
 130         size_t hdr_size;
 131         struct socket *sock;
 132
 133         /* TODO: check that we are running from vhost_worker? */
 134         sock = rcu_dereference_check(vq->private_data, 1);
 135         if (!sock)
 136                 return;
 137
 138         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 139         if (wmem >= sock->sk->sk_sndbuf) {
 140                 mutex_lock(&vq->mutex);
 141                 tx_poll_start(net, sock);
 142                 mutex_unlock(&vq->mutex);
 143                 return;
 144         }
 145
 146         mutex_lock(&vq->mutex);
 147         vhost_disable_notify(vq);
 148
 149         if (wmem < sock->sk->sk_sndbuf / 2)
 150                 tx_poll_stop(net);
 151         hdr_size = vq->vhost_hlen;
 152
 153         for (;;) {
 154                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 155                                          ARRAY_SIZE(vq->iov),
 156                                          &out, &in,
 157                                          NULL, NULL);
 158                 /* On error, stop handling until the next kick. */
 159                 if (unlikely(head < 0))
 160                         break;
 161                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 162                 if (head == vq->num) {
 163                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 164                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 165                                 tx_poll_start(net, sock);
 166                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 167                                 break;
 168                         }
 169                         if (unlikely(vhost_enable_notify(vq))) {
 170                                 vhost_disable_notify(vq);
 171                                 continue;
 172                         }
 173                         break;
 174                 }
 175                 if (in) {
 176                         vq_err(vq, "Unexpected descriptor format for TX: "
 177                                "out %d, int %d\n", out, in);
 178                         break;
 179                 }
 180                 /* Skip header. TODO: support TSO. */
 181                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 182                 msg.msg_iovlen = out;
 183                 len = iov_length(vq->iov, out);
 184                 /* Sanity check */
 185                 if (!len) {
 186                         vq_err(vq, "Unexpected header len for TX: "
 187                                "%zd expected %zd\n",
 188                                iov_length(vq->hdr, s), hdr_size);
 189                         break;
 190                 }
 191                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 192                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 193                 if (unlikely(err < 0)) {
 194                         vhost_discard_vq_desc(vq, 1);
 195                         tx_poll_start(net, sock);
 196                         break;
 197                 }
 198                 if (err != len)
 199                         pr_debug("Truncated TX packet: "
 200                                  " len %d != %zd\n", err, len);
 201                 vhost_add_used_and_signal(&net->dev, vq, head, 0);
 202                 total_len += len;
 203                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 204                         vhost_poll_queue(&vq->poll);
 205                         break;
 206                 }
 207         }
 208
 209         mutex_unlock(&vq->mutex);
 210 }
 211
 212 static int peek_head_len(struct sock *sk)
 213 {
 214         struct sk_buff *head;
 215         int len = 0;
 216
 217         lock_sock(sk);
 218         head = skb_peek(&sk->sk_receive_queue);
 219         if (head)
 220                 len = head->len;
 221         release_sock(sk);
 222         return len;
 223 }
 224
 225 /* This is a multi-buffer version of vhost_get_desc, that works if
 226  *      vq has read descriptors only.
 227  * @vq          - the relevant virtqueue
 228  * @datalen     - data length we'll be reading
 229  * @iovcount    - returned count of io vectors we fill
 230  * @log         - vhost log
 231  * @log_num     - log offset
 232  *      returns number of buffer heads allocated, negative on error
 233  */
 234 static int get_rx_bufs(struct vhost_virtqueue *vq,
 235                        struct vring_used_elem *heads,
 236                        int datalen,
 237                        unsigned *iovcount,
 238                        struct vhost_log *log,
 239                        unsigned *log_num)
 240 {
 241         unsigned int out, in;
 242         int seg = 0;
 243         int headcount = 0;
 244         unsigned d;
 245         int r, nlogs = 0;
 246
 247         while (datalen > 0) {
 248                 if (unlikely(seg >= UIO_MAXIOV)) {
 249                         r = -ENOBUFS;
 250                         goto err;
 251                 }
 252                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 253                                       ARRAY_SIZE(vq->iov) - seg, &out,
 254                                       &in, log, log_num);
 255                 if (d == vq->num) {
 256                         r = 0;
 257                         goto err;
 258                 }
 259                 if (unlikely(out || in <= 0)) {
 260                         vq_err(vq, "unexpected descriptor format for RX: "
 261                                 "out %d, in %d\n", out, in);
 262                         r = -EINVAL;
 263                         goto err;
 264                 }
 265                 if (unlikely(log)) {
 266                         nlogs += *log_num;
 267                         log += *log_num;
 268                 }
 269                 heads[headcount].id = d;
 270                 heads[headcount].len = iov_length(vq->iov + seg, in);
 271                 datalen -= heads[headcount].len;
 272                 ++headcount;
 273                 seg += in;
 274         }
 275         heads[headcount - 1].len += datalen;
 276         *iovcount = seg;
 277         if (unlikely(log))
 278                 *log_num = nlogs;
 279         return headcount;
 280 err:
 281         vhost_discard_vq_desc(vq, headcount);
 282         return r;
 283 }
 284
 285 /* Expects to be always run from workqueue - which acts as
 286  * read-size critical section for our kind of RCU. */
 287 static void handle_rx_big(struct vhost_net *net)
 288 {
 289         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 290         unsigned out, in, log, s;
 291         int head;
 292         struct vhost_log *vq_log;
 293         struct msghdr msg = {
 294                 .msg_name = NULL,
 295                 .msg_namelen = 0,
 296                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 297                 .msg_controllen = 0,
 298                 .msg_iov = vq->iov,
 299                 .msg_flags = MSG_DONTWAIT,
 300         };
 301         struct virtio_net_hdr hdr = {
 302                 .flags = 0,
 303                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 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
 311         if (!sock || skb_queue_empty(&sock->sk->sk_receive_queue))
 312                 return;
 313
 314         mutex_lock(&vq->mutex);
 315         vhost_disable_notify(vq);
 316         hdr_size = vq->vhost_hlen;
 317
 318         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 319                 vq->log : NULL;
 320
 321         for (;;) {
 322                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 323                                          ARRAY_SIZE(vq->iov),
 324                                          &out, &in,
 325                                          vq_log, &log);
 326                 /* On error, stop handling until the next kick. */
 327                 if (unlikely(head < 0))
 328                         break;
 329                 /* OK, now we need to know about added descriptors. */
 330                 if (head == vq->num) {
 331                         if (unlikely(vhost_enable_notify(vq))) {
 332                                 /* They have slipped one in as we were
 333                                  * doing that: check again. */
 334                                 vhost_disable_notify(vq);
 335                                 continue;
 336                         }
 337                         /* Nothing new?  Wait for eventfd to tell us
 338                          * they refilled. */
 339                         break;
 340                 }
 341                 /* We don't need to be notified again. */
 342                 if (out) {
 343                         vq_err(vq, "Unexpected descriptor format for RX: "
 344                                "out %d, int %d\n",
 345                                out, in);
 346                         break;
 347                 }
 348                 /* Skip header. TODO: support TSO/mergeable rx buffers. */
 349                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, in);
 350                 msg.msg_iovlen = in;
 351                 len = iov_length(vq->iov, in);
 352                 /* Sanity check */
 353                 if (!len) {
 354                         vq_err(vq, "Unexpected header len for RX: "
 355                                "%zd expected %zd\n",
 356                                iov_length(vq->hdr, s), hdr_size);
 357                         break;
 358                 }
 359                 err = sock->ops->recvmsg(NULL, sock, &msg,
 360                                          len, MSG_DONTWAIT | MSG_TRUNC);
 361                 /* TODO: Check specific error and bomb out unless EAGAIN? */
 362                 if (err < 0) {
 363                         vhost_discard_vq_desc(vq, 1);
 364                         break;
 365                 }
 366                 /* TODO: Should check and handle checksum. */
 367                 if (err > len) {
 368                         pr_debug("Discarded truncated rx packet: "
 369                                  " len %d > %zd\n", err, len);
 370                         vhost_discard_vq_desc(vq, 1);
 371                         continue;
 372                 }
 373                 len = err;
 374                 err = memcpy_toiovec(vq->hdr, (unsigned char *)&hdr, hdr_size);
 375                 if (err) {
 376                         vq_err(vq, "Unable to write vnet_hdr at addr %p: %d\n",
 377                                vq->iov->iov_base, err);
 378                         break;
 379                 }
 380                 len += hdr_size;
 381                 vhost_add_used_and_signal(&net->dev, vq, head, len);
 382                 if (unlikely(vq_log))
 383                         vhost_log_write(vq, vq_log, log, len);
 384                 total_len += len;
 385                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 386                         vhost_poll_queue(&vq->poll);
 387                         break;
 388                 }
 389         }
 390
 391         mutex_unlock(&vq->mutex);
 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         struct virtio_net_hdr_mrg_rxbuf hdr = {
 410                 .hdr.flags = 0,
 411                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 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         /* TODO: check that we are running from vhost_worker? */
 418         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 419
 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
 658         if (!sock)
 659                 return ERR_PTR(-ENOTSOCK);
 660
 661         /* Parameter checking */
 662         if (sock->sk->sk_type != SOCK_RAW) {
 663                 r = -ESOCKTNOSUPPORT;
 664                 goto err;
 665         }
 666
 667         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 668                                &uaddr_len, 0);
 669         if (r)
 670                 goto err;
 671
 672         if (uaddr.sa.sll_family != AF_PACKET) {
 673                 r = -EPFNOSUPPORT;
 674                 goto err;
 675         }
 676         return sock;
 677 err:
 678         fput(sock->file);
 679         return ERR_PTR(r);
 680 }
 681
 682 static struct socket *get_tap_socket(int fd)
 683 {
 684         struct file *file = fget(fd);
 685         struct socket *sock;
 686
 687         if (!file)
 688                 return ERR_PTR(-EBADF);
 689         sock = tun_get_socket(file);
 690         if (!IS_ERR(sock))
 691                 return sock;
 692         sock = macvtap_get_socket(file);
 693         if (IS_ERR(sock))
 694                 fput(file);
 695         return sock;
 696 }
 697
 698 static struct socket *get_socket(int fd)
 699 {
 700         struct socket *sock;
 701
 702         /* special case to disable backend */
 703         if (fd == -1)
 704                 return NULL;
 705         sock = get_raw_socket(fd);
 706         if (!IS_ERR(sock))
 707                 return sock;
 708         sock = get_tap_socket(fd);
 709         if (!IS_ERR(sock))
 710                 return sock;
 711         return ERR_PTR(-ENOTSOCK);
 712 }
 713
 714 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 715 {
 716         struct socket *sock, *oldsock;
 717         struct vhost_virtqueue *vq;
 718         int r;
 719
 720         mutex_lock(&n->dev.mutex);
 721         r = vhost_dev_check_owner(&n->dev);
 722         if (r)
 723                 goto err;
 724
 725         if (index >= VHOST_NET_VQ_MAX) {
 726                 r = -ENOBUFS;
 727                 goto err;
 728         }
 729         vq = n->vqs + index;
 730         mutex_lock(&vq->mutex);
 731
 732         /* Verify that ring has been setup correctly. */
 733         if (!vhost_vq_access_ok(vq)) {
 734                 r = -EFAULT;
 735                 goto err_vq;
 736         }
 737         sock = get_socket(fd);
 738         if (IS_ERR(sock)) {
 739                 r = PTR_ERR(sock);
 740                 goto err_vq;
 741         }
 742
 743         /* start polling new socket */
 744         oldsock = rcu_dereference_protected(vq->private_data,
 745                                             lockdep_is_held(&vq->mutex));
 746         if (sock != oldsock) {
 747                 vhost_net_disable_vq(n, vq);
 748                 rcu_assign_pointer(vq->private_data, sock);
 749                 vhost_net_enable_vq(n, vq);
 750         }
 751
 752         mutex_unlock(&vq->mutex);
 753
 754         if (oldsock) {
 755                 vhost_net_flush_vq(n, index);
 756                 fput(oldsock->file);
 757         }
 758
 759         mutex_unlock(&n->dev.mutex);
 760         return 0;
 761
 762 err_vq:
 763         mutex_unlock(&vq->mutex);
 764 err:
 765         mutex_unlock(&n->dev.mutex);
 766         return r;
 767 }
 768
 769 static long vhost_net_reset_owner(struct vhost_net *n)
 770 {
 771         struct socket *tx_sock = NULL;
 772         struct socket *rx_sock = NULL;
 773         long err;
 774
 775         mutex_lock(&n->dev.mutex);
 776         err = vhost_dev_check_owner(&n->dev);
 777         if (err)
 778                 goto done;
 779         vhost_net_stop(n, &tx_sock, &rx_sock);
 780         vhost_net_flush(n);
 781         err = vhost_dev_reset_owner(&n->dev);
 782 done:
 783         mutex_unlock(&n->dev.mutex);
 784         if (tx_sock)
 785                 fput(tx_sock->file);
 786         if (rx_sock)
 787                 fput(rx_sock->file);
 788         return err;
 789 }
 790
 791 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 792 {
 793         size_t vhost_hlen, sock_hlen, hdr_len;
 794         int i;
 795
 796         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 797                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 798                         sizeof(struct virtio_net_hdr);
 799         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 800                 /* vhost provides vnet_hdr */
 801                 vhost_hlen = hdr_len;
 802                 sock_hlen = 0;
 803         } else {
 804                 /* socket provides vnet_hdr */
 805                 vhost_hlen = 0;
 806                 sock_hlen = hdr_len;
 807         }
 808         mutex_lock(&n->dev.mutex);
 809         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 810             !vhost_log_access_ok(&n->dev)) {
 811                 mutex_unlock(&n->dev.mutex);
 812                 return -EFAULT;
 813         }
 814         n->dev.acked_features = features;
 815         smp_wmb();
 816         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 817                 mutex_lock(&n->vqs[i].mutex);
 818                 n->vqs[i].vhost_hlen = vhost_hlen;
 819                 n->vqs[i].sock_hlen = sock_hlen;
 820                 mutex_unlock(&n->vqs[i].mutex);
 821         }
 822         vhost_net_flush(n);
 823         mutex_unlock(&n->dev.mutex);
 824         return 0;
 825 }
 826
 827 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 828                             unsigned long arg)
 829 {
 830         struct vhost_net *n = f->private_data;
 831         void __user *argp = (void __user *)arg;
 832         u64 __user *featurep = argp;
 833         struct vhost_vring_file backend;
 834         u64 features;
 835         int r;
 836
 837         switch (ioctl) {
 838         case VHOST_NET_SET_BACKEND:
 839                 if (copy_from_user(&backend, argp, sizeof backend))
 840                         return -EFAULT;
 841                 return vhost_net_set_backend(n, backend.index, backend.fd);
 842         case VHOST_GET_FEATURES:
 843                 features = VHOST_FEATURES;
 844                 if (copy_to_user(featurep, &features, sizeof features))
 845                         return -EFAULT;
 846                 return 0;
 847         case VHOST_SET_FEATURES:
 848                 if (copy_from_user(&features, featurep, sizeof features))
 849                         return -EFAULT;
 850                 if (features & ~VHOST_FEATURES)
 851                         return -EOPNOTSUPP;
 852                 return vhost_net_set_features(n, features);
 853         case VHOST_RESET_OWNER:
 854                 return vhost_net_reset_owner(n);
 855         default:
 856                 mutex_lock(&n->dev.mutex);
 857                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 858                 vhost_net_flush(n);
 859                 mutex_unlock(&n->dev.mutex);
 860                 return r;
 861         }
 862 }
 863
 864 #ifdef CONFIG_COMPAT
 865 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 866                                    unsigned long arg)
 867 {
 868         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 869 }
 870 #endif
 871
 872 static const struct file_operations vhost_net_fops = {
 873         .owner          = THIS_MODULE,
 874         .release        = vhost_net_release,
 875         .unlocked_ioctl = vhost_net_ioctl,
 876 #ifdef CONFIG_COMPAT
 877         .compat_ioctl   = vhost_net_compat_ioctl,
 878 #endif
 879         .open           = vhost_net_open,
 880         .llseek         = noop_llseek,
 881 };
 882
 883 static struct miscdevice vhost_net_misc = {
 884         MISC_DYNAMIC_MINOR,
 885         "vhost-net",
 886         &vhost_net_fops,
 887 };
 888
 889 static int vhost_net_init(void)
 890 {
 891         return misc_register(&vhost_net_misc);
 892 }
 893 module_init(vhost_net_init);
 894
 895 static void vhost_net_exit(void)
 896 {
 897         misc_deregister(&vhost_net_misc);
 898 }
 899 module_exit(vhost_net_exit);
 900
 901 MODULE_VERSION("0.0.1");
 902 MODULE_LICENSE("GPL v2");
 903 MODULE_AUTHOR("Michael S. Tsirkin");
 904 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");