linux/drivers/vhost/net.c
<<
>>
Prefs
   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/moduleparam.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#include <linux/if_vlan.h>
  28
  29#include <net/sock.h>
  30
  31#include "vhost.h"
  32
  33static int experimental_zcopytx;
  34module_param(experimental_zcopytx, int, 0444);
  35MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
  36
  37/* Max number of bytes transferred before requeueing the job.
  38 * Using this limit prevents one virtqueue from starving others. */
  39#define VHOST_NET_WEIGHT 0x80000
  40
  41/* MAX number of TX used buffers for outstanding zerocopy */
  42#define VHOST_MAX_PEND 128
  43#define VHOST_GOODCOPY_LEN 256
  44
  45enum {
  46        VHOST_NET_VQ_RX = 0,
  47        VHOST_NET_VQ_TX = 1,
  48        VHOST_NET_VQ_MAX = 2,
  49};
  50
  51enum vhost_net_poll_state {
  52        VHOST_NET_POLL_DISABLED = 0,
  53        VHOST_NET_POLL_STARTED = 1,
  54        VHOST_NET_POLL_STOPPED = 2,
  55};
  56
  57struct vhost_net {
  58        struct vhost_dev dev;
  59        struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  60        struct vhost_poll poll[VHOST_NET_VQ_MAX];
  61        /* Tells us whether we are polling a socket for TX.
  62         * We only do this when socket buffer fills up.
  63         * Protected by tx vq lock. */
  64        enum vhost_net_poll_state tx_poll_state;
  65};
  66
  67static bool vhost_sock_zcopy(struct socket *sock)
  68{
  69        return unlikely(experimental_zcopytx) &&
  70                sock_flag(sock->sk, SOCK_ZEROCOPY);
  71}
  72
  73/* Pop first len bytes from iovec. Return number of segments used. */
  74static int move_iovec_hdr(struct iovec *from, struct iovec *to,
  75                          size_t len, int iov_count)
  76{
  77        int seg = 0;
  78        size_t size;
  79
  80        while (len && seg < iov_count) {
  81                size = min(from->iov_len, len);
  82                to->iov_base = from->iov_base;
  83                to->iov_len = size;
  84                from->iov_len -= size;
  85                from->iov_base += size;
  86                len -= size;
  87                ++from;
  88                ++to;
  89                ++seg;
  90        }
  91        return seg;
  92}
  93/* Copy iovec entries for len bytes from iovec. */
  94static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
  95                           size_t len, int iovcount)
  96{
  97        int seg = 0;
  98        size_t size;
  99
 100        while (len && seg < iovcount) {
 101                size = min(from->iov_len, len);
 102                to->iov_base = from->iov_base;
 103                to->iov_len = size;
 104                len -= size;
 105                ++from;
 106                ++to;
 107                ++seg;
 108        }
 109}
 110
 111/* Caller must have TX VQ lock */
 112static void tx_poll_stop(struct vhost_net *net)
 113{
 114        if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
 115                return;
 116        vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 117        net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 118}
 119
 120/* Caller must have TX VQ lock */
 121static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 122{
 123        if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 124                return;
 125        vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 126        net->tx_poll_state = VHOST_NET_POLL_STARTED;
 127}
 128
 129/* Expects to be always run from workqueue - which acts as
 130 * read-size critical section for our kind of RCU. */
 131static void handle_tx(struct vhost_net *net)
 132{
 133        struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 134        unsigned out, in, s;
 135        int head;
 136        struct msghdr msg = {
 137                .msg_name = NULL,
 138                .msg_namelen = 0,
 139                .msg_control = NULL,
 140                .msg_controllen = 0,
 141                .msg_iov = vq->iov,
 142                .msg_flags = MSG_DONTWAIT,
 143        };
 144        size_t len, total_len = 0;
 145        int err, wmem;
 146        size_t hdr_size;
 147        struct socket *sock;
 148        struct vhost_ubuf_ref *uninitialized_var(ubufs);
 149        bool zcopy;
 150
 151        /* TODO: check that we are running from vhost_worker? */
 152        sock = rcu_dereference_check(vq->private_data, 1);
 153        if (!sock)
 154                return;
 155
 156        wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 157        if (wmem >= sock->sk->sk_sndbuf) {
 158                mutex_lock(&vq->mutex);
 159                tx_poll_start(net, sock);
 160                mutex_unlock(&vq->mutex);
 161                return;
 162        }
 163
 164        mutex_lock(&vq->mutex);
 165        vhost_disable_notify(&net->dev, vq);
 166
 167        if (wmem < sock->sk->sk_sndbuf / 2)
 168                tx_poll_stop(net);
 169        hdr_size = vq->vhost_hlen;
 170        zcopy = vq->ubufs;
 171
 172        for (;;) {
 173                /* Release DMAs done buffers first */
 174                if (zcopy)
 175                        vhost_zerocopy_signal_used(vq);
 176
 177                head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 178                                         ARRAY_SIZE(vq->iov),
 179                                         &out, &in,
 180                                         NULL, NULL);
 181                /* On error, stop handling until the next kick. */
 182                if (unlikely(head < 0))
 183                        break;
 184                /* Nothing new?  Wait for eventfd to tell us they refilled. */
 185                if (head == vq->num) {
 186                        int num_pends;
 187
 188                        wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 189                        if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 190                                tx_poll_start(net, sock);
 191                                set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 192                                break;
 193                        }
 194                        /* If more outstanding DMAs, queue the work.
 195                         * Handle upend_idx wrap around
 196                         */
 197                        num_pends = likely(vq->upend_idx >= vq->done_idx) ?
 198                                    (vq->upend_idx - vq->done_idx) :
 199                                    (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
 200                        if (unlikely(num_pends > VHOST_MAX_PEND)) {
 201                                tx_poll_start(net, sock);
 202                                set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 203                                break;
 204                        }
 205                        if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 206                                vhost_disable_notify(&net->dev, vq);
 207                                continue;
 208                        }
 209                        break;
 210                }
 211                if (in) {
 212                        vq_err(vq, "Unexpected descriptor format for TX: "
 213                               "out %d, int %d\n", out, in);
 214                        break;
 215                }
 216                /* Skip header. TODO: support TSO. */
 217                s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 218                msg.msg_iovlen = out;
 219                len = iov_length(vq->iov, out);
 220                /* Sanity check */
 221                if (!len) {
 222                        vq_err(vq, "Unexpected header len for TX: "
 223                               "%zd expected %zd\n",
 224                               iov_length(vq->hdr, s), hdr_size);
 225                        break;
 226                }
 227                /* use msg_control to pass vhost zerocopy ubuf info to skb */
 228                if (zcopy) {
 229                        vq->heads[vq->upend_idx].id = head;
 230                        if (len < VHOST_GOODCOPY_LEN) {
 231                                /* copy don't need to wait for DMA done */
 232                                vq->heads[vq->upend_idx].len =
 233                                                        VHOST_DMA_DONE_LEN;
 234                                msg.msg_control = NULL;
 235                                msg.msg_controllen = 0;
 236                                ubufs = NULL;
 237                        } else {
 238                                struct ubuf_info *ubuf = &vq->ubuf_info[head];
 239
 240                                vq->heads[vq->upend_idx].len = len;
 241                                ubuf->callback = vhost_zerocopy_callback;
 242                                ubuf->ctx = vq->ubufs;
 243                                ubuf->desc = vq->upend_idx;
 244                                msg.msg_control = ubuf;
 245                                msg.msg_controllen = sizeof(ubuf);
 246                                ubufs = vq->ubufs;
 247                                kref_get(&ubufs->kref);
 248                        }
 249                        vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
 250                }
 251                /* TODO: Check specific error and bomb out unless ENOBUFS? */
 252                err = sock->ops->sendmsg(NULL, sock, &msg, len);
 253                if (unlikely(err < 0)) {
 254                        if (zcopy) {
 255                                if (ubufs)
 256                                        vhost_ubuf_put(ubufs);
 257                                vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
 258                                        UIO_MAXIOV;
 259                        }
 260                        vhost_discard_vq_desc(vq, 1);
 261                        if (err == -EAGAIN || err == -ENOBUFS)
 262                                tx_poll_start(net, sock);
 263                        break;
 264                }
 265                if (err != len)
 266                        pr_debug("Truncated TX packet: "
 267                                 " len %d != %zd\n", err, len);
 268                if (!zcopy)
 269                        vhost_add_used_and_signal(&net->dev, vq, head, 0);
 270                else
 271                        vhost_zerocopy_signal_used(vq);
 272                total_len += len;
 273                if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 274                        vhost_poll_queue(&vq->poll);
 275                        break;
 276                }
 277        }
 278
 279        mutex_unlock(&vq->mutex);
 280}
 281
 282static int peek_head_len(struct sock *sk)
 283{
 284        struct sk_buff *head;
 285        int len = 0;
 286        unsigned long flags;
 287
 288        spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 289        head = skb_peek(&sk->sk_receive_queue);
 290        if (likely(head)) {
 291                len = head->len;
 292                if (vlan_tx_tag_present(head))
 293                        len += VLAN_HLEN;
 294        }
 295
 296        spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 297        return len;
 298}
 299
 300/* This is a multi-buffer version of vhost_get_desc, that works if
 301 *      vq has read descriptors only.
 302 * @vq          - the relevant virtqueue
 303 * @datalen     - data length we'll be reading
 304 * @iovcount    - returned count of io vectors we fill
 305 * @log         - vhost log
 306 * @log_num     - log offset
 307 * @quota       - headcount quota, 1 for big buffer
 308 *      returns number of buffer heads allocated, negative on error
 309 */
 310static int get_rx_bufs(struct vhost_virtqueue *vq,
 311                       struct vring_used_elem *heads,
 312                       int datalen,
 313                       unsigned *iovcount,
 314                       struct vhost_log *log,
 315                       unsigned *log_num,
 316                       unsigned int quota)
 317{
 318        unsigned int out, in;
 319        int seg = 0;
 320        int headcount = 0;
 321        unsigned d;
 322        int r, nlogs = 0;
 323
 324        while (datalen > 0 && headcount < quota) {
 325                if (unlikely(seg >= UIO_MAXIOV)) {
 326                        r = -ENOBUFS;
 327                        goto err;
 328                }
 329                d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 330                                      ARRAY_SIZE(vq->iov) - seg, &out,
 331                                      &in, log, log_num);
 332                if (d == vq->num) {
 333                        r = 0;
 334                        goto err;
 335                }
 336                if (unlikely(out || in <= 0)) {
 337                        vq_err(vq, "unexpected descriptor format for RX: "
 338                                "out %d, in %d\n", out, in);
 339                        r = -EINVAL;
 340                        goto err;
 341                }
 342                if (unlikely(log)) {
 343                        nlogs += *log_num;
 344                        log += *log_num;
 345                }
 346                heads[headcount].id = d;
 347                heads[headcount].len = iov_length(vq->iov + seg, in);
 348                datalen -= heads[headcount].len;
 349                ++headcount;
 350                seg += in;
 351        }
 352        heads[headcount - 1].len += datalen;
 353        *iovcount = seg;
 354        if (unlikely(log))
 355                *log_num = nlogs;
 356        return headcount;
 357err:
 358        vhost_discard_vq_desc(vq, headcount);
 359        return r;
 360}
 361
 362/* Expects to be always run from workqueue - which acts as
 363 * read-size critical section for our kind of RCU. */
 364static void handle_rx(struct vhost_net *net)
 365{
 366        struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 367        unsigned uninitialized_var(in), log;
 368        struct vhost_log *vq_log;
 369        struct msghdr msg = {
 370                .msg_name = NULL,
 371                .msg_namelen = 0,
 372                .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 373                .msg_controllen = 0,
 374                .msg_iov = vq->iov,
 375                .msg_flags = MSG_DONTWAIT,
 376        };
 377        struct virtio_net_hdr_mrg_rxbuf hdr = {
 378                .hdr.flags = 0,
 379                .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 380        };
 381        size_t total_len = 0;
 382        int err, headcount, mergeable;
 383        size_t vhost_hlen, sock_hlen;
 384        size_t vhost_len, sock_len;
 385        /* TODO: check that we are running from vhost_worker? */
 386        struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 387
 388        if (!sock)
 389                return;
 390
 391        mutex_lock(&vq->mutex);
 392        vhost_disable_notify(&net->dev, vq);
 393        vhost_hlen = vq->vhost_hlen;
 394        sock_hlen = vq->sock_hlen;
 395
 396        vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 397                vq->log : NULL;
 398        mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 399
 400        while ((sock_len = peek_head_len(sock->sk))) {
 401                sock_len += sock_hlen;
 402                vhost_len = sock_len + vhost_hlen;
 403                headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 404                                        &in, vq_log, &log,
 405                                        likely(mergeable) ? UIO_MAXIOV : 1);
 406                /* On error, stop handling until the next kick. */
 407                if (unlikely(headcount < 0))
 408                        break;
 409                /* OK, now we need to know about added descriptors. */
 410                if (!headcount) {
 411                        if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 412                                /* They have slipped one in as we were
 413                                 * doing that: check again. */
 414                                vhost_disable_notify(&net->dev, vq);
 415                                continue;
 416                        }
 417                        /* Nothing new?  Wait for eventfd to tell us
 418                         * they refilled. */
 419                        break;
 420                }
 421                /* We don't need to be notified again. */
 422                if (unlikely((vhost_hlen)))
 423                        /* Skip header. TODO: support TSO. */
 424                        move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 425                else
 426                        /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 427                         * needed because recvmsg can modify msg_iov. */
 428                        copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 429                msg.msg_iovlen = in;
 430                err = sock->ops->recvmsg(NULL, sock, &msg,
 431                                         sock_len, MSG_DONTWAIT | MSG_TRUNC);
 432                /* Userspace might have consumed the packet meanwhile:
 433                 * it's not supposed to do this usually, but might be hard
 434                 * to prevent. Discard data we got (if any) and keep going. */
 435                if (unlikely(err != sock_len)) {
 436                        pr_debug("Discarded rx packet: "
 437                                 " len %d, expected %zd\n", err, sock_len);
 438                        vhost_discard_vq_desc(vq, headcount);
 439                        continue;
 440                }
 441                if (unlikely(vhost_hlen) &&
 442                    memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 443                                      vhost_hlen)) {
 444                        vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 445                               vq->iov->iov_base);
 446                        break;
 447                }
 448                /* TODO: Should check and handle checksum. */
 449                if (likely(mergeable) &&
 450                    memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 451                                      offsetof(typeof(hdr), num_buffers),
 452                                      sizeof hdr.num_buffers)) {
 453                        vq_err(vq, "Failed num_buffers write");
 454                        vhost_discard_vq_desc(vq, headcount);
 455                        break;
 456                }
 457                vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 458                                            headcount);
 459                if (unlikely(vq_log))
 460                        vhost_log_write(vq, vq_log, log, vhost_len);
 461                total_len += vhost_len;
 462                if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 463                        vhost_poll_queue(&vq->poll);
 464                        break;
 465                }
 466        }
 467
 468        mutex_unlock(&vq->mutex);
 469}
 470
 471static void handle_tx_kick(struct vhost_work *work)
 472{
 473        struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 474                                                  poll.work);
 475        struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 476
 477        handle_tx(net);
 478}
 479
 480static void handle_rx_kick(struct vhost_work *work)
 481{
 482        struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 483                                                  poll.work);
 484        struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 485
 486        handle_rx(net);
 487}
 488
 489static void handle_tx_net(struct vhost_work *work)
 490{
 491        struct vhost_net *net = container_of(work, struct vhost_net,
 492                                             poll[VHOST_NET_VQ_TX].work);
 493        handle_tx(net);
 494}
 495
 496static void handle_rx_net(struct vhost_work *work)
 497{
 498        struct vhost_net *net = container_of(work, struct vhost_net,
 499                                             poll[VHOST_NET_VQ_RX].work);
 500        handle_rx(net);
 501}
 502
 503static int vhost_net_open(struct inode *inode, struct file *f)
 504{
 505        struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 506        struct vhost_dev *dev;
 507        int r;
 508
 509        if (!n)
 510                return -ENOMEM;
 511
 512        dev = &n->dev;
 513        n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 514        n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 515        r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 516        if (r < 0) {
 517                kfree(n);
 518                return r;
 519        }
 520
 521        vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 522        vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 523        n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 524
 525        f->private_data = n;
 526
 527        return 0;
 528}
 529
 530static void vhost_net_disable_vq(struct vhost_net *n,
 531                                 struct vhost_virtqueue *vq)
 532{
 533        if (!vq->private_data)
 534                return;
 535        if (vq == n->vqs + VHOST_NET_VQ_TX) {
 536                tx_poll_stop(n);
 537                n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 538        } else
 539                vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 540}
 541
 542static void vhost_net_enable_vq(struct vhost_net *n,
 543                                struct vhost_virtqueue *vq)
 544{
 545        struct socket *sock;
 546
 547        sock = rcu_dereference_protected(vq->private_data,
 548                                         lockdep_is_held(&vq->mutex));
 549        if (!sock)
 550                return;
 551        if (vq == n->vqs + VHOST_NET_VQ_TX) {
 552                n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 553                tx_poll_start(n, sock);
 554        } else
 555                vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 556}
 557
 558static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 559                                        struct vhost_virtqueue *vq)
 560{
 561        struct socket *sock;
 562
 563        mutex_lock(&vq->mutex);
 564        sock = rcu_dereference_protected(vq->private_data,
 565                                         lockdep_is_held(&vq->mutex));
 566        vhost_net_disable_vq(n, vq);
 567        rcu_assign_pointer(vq->private_data, NULL);
 568        mutex_unlock(&vq->mutex);
 569        return sock;
 570}
 571
 572static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 573                           struct socket **rx_sock)
 574{
 575        *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 576        *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 577}
 578
 579static void vhost_net_flush_vq(struct vhost_net *n, int index)
 580{
 581        vhost_poll_flush(n->poll + index);
 582        vhost_poll_flush(&n->dev.vqs[index].poll);
 583}
 584
 585static void vhost_net_flush(struct vhost_net *n)
 586{
 587        vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 588        vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 589}
 590
 591static int vhost_net_release(struct inode *inode, struct file *f)
 592{
 593        struct vhost_net *n = f->private_data;
 594        struct socket *tx_sock;
 595        struct socket *rx_sock;
 596
 597        vhost_net_stop(n, &tx_sock, &rx_sock);
 598        vhost_net_flush(n);
 599        vhost_dev_cleanup(&n->dev, false);
 600        if (tx_sock)
 601                fput(tx_sock->file);
 602        if (rx_sock)
 603                fput(rx_sock->file);
 604        /* We do an extra flush before freeing memory,
 605         * since jobs can re-queue themselves. */
 606        vhost_net_flush(n);
 607        kfree(n);
 608        return 0;
 609}
 610
 611static struct socket *get_raw_socket(int fd)
 612{
 613        struct {
 614                struct sockaddr_ll sa;
 615                char  buf[MAX_ADDR_LEN];
 616        } uaddr;
 617        int uaddr_len = sizeof uaddr, r;
 618        struct socket *sock = sockfd_lookup(fd, &r);
 619
 620        if (!sock)
 621                return ERR_PTR(-ENOTSOCK);
 622
 623        /* Parameter checking */
 624        if (sock->sk->sk_type != SOCK_RAW) {
 625                r = -ESOCKTNOSUPPORT;
 626                goto err;
 627        }
 628
 629        r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 630                               &uaddr_len, 0);
 631        if (r)
 632                goto err;
 633
 634        if (uaddr.sa.sll_family != AF_PACKET) {
 635                r = -EPFNOSUPPORT;
 636                goto err;
 637        }
 638        return sock;
 639err:
 640        fput(sock->file);
 641        return ERR_PTR(r);
 642}
 643
 644static struct socket *get_tap_socket(int fd)
 645{
 646        struct file *file = fget(fd);
 647        struct socket *sock;
 648
 649        if (!file)
 650                return ERR_PTR(-EBADF);
 651        sock = tun_get_socket(file);
 652        if (!IS_ERR(sock))
 653                return sock;
 654        sock = macvtap_get_socket(file);
 655        if (IS_ERR(sock))
 656                fput(file);
 657        return sock;
 658}
 659
 660static struct socket *get_socket(int fd)
 661{
 662        struct socket *sock;
 663
 664        /* special case to disable backend */
 665        if (fd == -1)
 666                return NULL;
 667        sock = get_raw_socket(fd);
 668        if (!IS_ERR(sock))
 669                return sock;
 670        sock = get_tap_socket(fd);
 671        if (!IS_ERR(sock))
 672                return sock;
 673        return ERR_PTR(-ENOTSOCK);
 674}
 675
 676static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 677{
 678        struct socket *sock, *oldsock;
 679        struct vhost_virtqueue *vq;
 680        struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 681        int r;
 682
 683        mutex_lock(&n->dev.mutex);
 684        r = vhost_dev_check_owner(&n->dev);
 685        if (r)
 686                goto err;
 687
 688        if (index >= VHOST_NET_VQ_MAX) {
 689                r = -ENOBUFS;
 690                goto err;
 691        }
 692        vq = n->vqs + index;
 693        mutex_lock(&vq->mutex);
 694
 695        /* Verify that ring has been setup correctly. */
 696        if (!vhost_vq_access_ok(vq)) {
 697                r = -EFAULT;
 698                goto err_vq;
 699        }
 700        sock = get_socket(fd);
 701        if (IS_ERR(sock)) {
 702                r = PTR_ERR(sock);
 703                goto err_vq;
 704        }
 705
 706        /* start polling new socket */
 707        oldsock = rcu_dereference_protected(vq->private_data,
 708                                            lockdep_is_held(&vq->mutex));
 709        if (sock != oldsock) {
 710                ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 711                if (IS_ERR(ubufs)) {
 712                        r = PTR_ERR(ubufs);
 713                        goto err_ubufs;
 714                }
 715                oldubufs = vq->ubufs;
 716                vq->ubufs = ubufs;
 717                vhost_net_disable_vq(n, vq);
 718                rcu_assign_pointer(vq->private_data, sock);
 719                vhost_net_enable_vq(n, vq);
 720
 721                r = vhost_init_used(vq);
 722                if (r)
 723                        goto err_vq;
 724        }
 725
 726        mutex_unlock(&vq->mutex);
 727
 728        if (oldubufs) {
 729                vhost_ubuf_put_and_wait(oldubufs);
 730                mutex_lock(&vq->mutex);
 731                vhost_zerocopy_signal_used(vq);
 732                mutex_unlock(&vq->mutex);
 733        }
 734
 735        if (oldsock) {
 736                vhost_net_flush_vq(n, index);
 737                fput(oldsock->file);
 738        }
 739
 740        mutex_unlock(&n->dev.mutex);
 741        return 0;
 742
 743err_ubufs:
 744        fput(sock->file);
 745err_vq:
 746        mutex_unlock(&vq->mutex);
 747err:
 748        mutex_unlock(&n->dev.mutex);
 749        return r;
 750}
 751
 752static long vhost_net_reset_owner(struct vhost_net *n)
 753{
 754        struct socket *tx_sock = NULL;
 755        struct socket *rx_sock = NULL;
 756        long err;
 757
 758        mutex_lock(&n->dev.mutex);
 759        err = vhost_dev_check_owner(&n->dev);
 760        if (err)
 761                goto done;
 762        vhost_net_stop(n, &tx_sock, &rx_sock);
 763        vhost_net_flush(n);
 764        err = vhost_dev_reset_owner(&n->dev);
 765done:
 766        mutex_unlock(&n->dev.mutex);
 767        if (tx_sock)
 768                fput(tx_sock->file);
 769        if (rx_sock)
 770                fput(rx_sock->file);
 771        return err;
 772}
 773
 774static int vhost_net_set_features(struct vhost_net *n, u64 features)
 775{
 776        size_t vhost_hlen, sock_hlen, hdr_len;
 777        int i;
 778
 779        hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 780                        sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 781                        sizeof(struct virtio_net_hdr);
 782        if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 783                /* vhost provides vnet_hdr */
 784                vhost_hlen = hdr_len;
 785                sock_hlen = 0;
 786        } else {
 787                /* socket provides vnet_hdr */
 788                vhost_hlen = 0;
 789                sock_hlen = hdr_len;
 790        }
 791        mutex_lock(&n->dev.mutex);
 792        if ((features & (1 << VHOST_F_LOG_ALL)) &&
 793            !vhost_log_access_ok(&n->dev)) {
 794                mutex_unlock(&n->dev.mutex);
 795                return -EFAULT;
 796        }
 797        n->dev.acked_features = features;
 798        smp_wmb();
 799        for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 800                mutex_lock(&n->vqs[i].mutex);
 801                n->vqs[i].vhost_hlen = vhost_hlen;
 802                n->vqs[i].sock_hlen = sock_hlen;
 803                mutex_unlock(&n->vqs[i].mutex);
 804        }
 805        vhost_net_flush(n);
 806        mutex_unlock(&n->dev.mutex);
 807        return 0;
 808}
 809
 810static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 811                            unsigned long arg)
 812{
 813        struct vhost_net *n = f->private_data;
 814        void __user *argp = (void __user *)arg;
 815        u64 __user *featurep = argp;
 816        struct vhost_vring_file backend;
 817        u64 features;
 818        int r;
 819
 820        switch (ioctl) {
 821        case VHOST_NET_SET_BACKEND:
 822                if (copy_from_user(&backend, argp, sizeof backend))
 823                        return -EFAULT;
 824                return vhost_net_set_backend(n, backend.index, backend.fd);
 825        case VHOST_GET_FEATURES:
 826                features = VHOST_NET_FEATURES;
 827                if (copy_to_user(featurep, &features, sizeof features))
 828                        return -EFAULT;
 829                return 0;
 830        case VHOST_SET_FEATURES:
 831                if (copy_from_user(&features, featurep, sizeof features))
 832                        return -EFAULT;
 833                if (features & ~VHOST_NET_FEATURES)
 834                        return -EOPNOTSUPP;
 835                return vhost_net_set_features(n, features);
 836        case VHOST_RESET_OWNER:
 837                return vhost_net_reset_owner(n);
 838        default:
 839                mutex_lock(&n->dev.mutex);
 840                r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 841                vhost_net_flush(n);
 842                mutex_unlock(&n->dev.mutex);
 843                return r;
 844        }
 845}
 846
 847#ifdef CONFIG_COMPAT
 848static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 849                                   unsigned long arg)
 850{
 851        return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 852}
 853#endif
 854
 855static const struct file_operations vhost_net_fops = {
 856        .owner          = THIS_MODULE,
 857        .release        = vhost_net_release,
 858        .unlocked_ioctl = vhost_net_ioctl,
 859#ifdef CONFIG_COMPAT
 860        .compat_ioctl   = vhost_net_compat_ioctl,
 861#endif
 862        .open           = vhost_net_open,
 863        .llseek         = noop_llseek,
 864};
 865
 866static struct miscdevice vhost_net_misc = {
 867        .minor = VHOST_NET_MINOR,
 868        .name = "vhost-net",
 869        .fops = &vhost_net_fops,
 870};
 871
 872static int vhost_net_init(void)
 873{
 874        if (experimental_zcopytx)
 875                vhost_enable_zcopy(VHOST_NET_VQ_TX);
 876        return misc_register(&vhost_net_misc);
 877}
 878module_init(vhost_net_init);
 879
 880static void vhost_net_exit(void)
 881{
 882        misc_deregister(&vhost_net_misc);
 883}
 884module_exit(vhost_net_exit);
 885
 886MODULE_VERSION("0.0.1");
 887MODULE_LICENSE("GPL v2");
 888MODULE_AUTHOR("Michael S. Tsirkin");
 889MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
 890MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
 891MODULE_ALIAS("devname:vhost-net");
 892
lxr.linux.no kindly hosted by Redpill Linpro AS, provider of Linux consulting and operations services since 1995.