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/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/vmalloc.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
  33 static int experimental_zcopytx = 0;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  36                                        " 1 -Enable; 0 - Disable");
  37
  38 /* Max number of bytes transferred before requeueing the job.
  39  * Using this limit prevents one virtqueue from starving others. */
  40 #define VHOST_NET_WEIGHT 0x80000
  41
  42 /* Max number of packets transferred before requeueing the job.
  43  * Using this limit prevents one virtqueue from starving others with small
  44  * pkts.
  45  */
  46 #define VHOST_NET_PKT_WEIGHT 256
  47
  48 /* MAX number of TX used buffers for outstanding zerocopy */
  49 #define VHOST_MAX_PEND 128
  50 #define VHOST_GOODCOPY_LEN 256
  51
  52 /*
  53  * For transmit, used buffer len is unused; we override it to track buffer
  54  * status internally; used for zerocopy tx only.
  55  */
  56 /* Lower device DMA failed */
  57 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  58 /* Lower device DMA done */
  59 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  60 /* Lower device DMA in progress */
  61 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  62 /* Buffer unused */
  63 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  64
  65 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  66
  67 enum {
  68         VHOST_NET_FEATURES = VHOST_FEATURES |
  69                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  70                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  71                          (1ULL << VIRTIO_F_IOMMU_PLATFORM)
  72 };
  73
  74 enum {
  75         VHOST_NET_VQ_RX = 0,
  76         VHOST_NET_VQ_TX = 1,
  77         VHOST_NET_VQ_MAX = 2,
  78 };
  79
  80 struct vhost_net_ubuf_ref {
  81         /* refcount follows semantics similar to kref:
  82          *  0: object is released
  83          *  1: no outstanding ubufs
  84          * >1: outstanding ubufs
  85          */
  86         atomic_t refcount;
  87         wait_queue_head_t wait;
  88         struct vhost_virtqueue *vq;
  89 };
  90
  91 struct vhost_net_virtqueue {
  92         struct vhost_virtqueue vq;
  93         size_t vhost_hlen;
  94         size_t sock_hlen;
  95         /* vhost zerocopy support fields below: */
  96         /* last used idx for outstanding DMA zerocopy buffers */
  97         int upend_idx;
  98         /* first used idx for DMA done zerocopy buffers */
  99         int done_idx;
 100         /* an array of userspace buffers info */
 101         struct ubuf_info *ubuf_info;
 102         /* Reference counting for outstanding ubufs.
 103          * Protected by vq mutex. Writers must also take device mutex. */
 104         struct vhost_net_ubuf_ref *ubufs;
 105 };
 106
 107 struct vhost_net {
 108         struct vhost_dev dev;
 109         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 110         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 111         /* Number of TX recently submitted.
 112          * Protected by tx vq lock. */
 113         unsigned tx_packets;
 114         /* Number of times zerocopy TX recently failed.
 115          * Protected by tx vq lock. */
 116         unsigned tx_zcopy_err;
 117         /* Flush in progress. Protected by tx vq lock. */
 118         bool tx_flush;
 119 };
 120
 121 static unsigned vhost_net_zcopy_mask __read_mostly;
 122
 123 static void vhost_net_enable_zcopy(int vq)
 124 {
 125         vhost_net_zcopy_mask |= 0x1 << vq;
 126 }
 127
 128 static struct vhost_net_ubuf_ref *
 129 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 130 {
 131         struct vhost_net_ubuf_ref *ubufs;
 132         /* No zero copy backend? Nothing to count. */
 133         if (!zcopy)
 134                 return NULL;
 135         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 136         if (!ubufs)
 137                 return ERR_PTR(-ENOMEM);
 138         atomic_set(&ubufs->refcount, 1);
 139         init_waitqueue_head(&ubufs->wait);
 140         ubufs->vq = vq;
 141         return ubufs;
 142 }
 143
 144 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 145 {
 146         int r = atomic_sub_return(1, &ubufs->refcount);
 147         if (unlikely(!r))
 148                 wake_up(&ubufs->wait);
 149         return r;
 150 }
 151
 152 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 153 {
 154         vhost_net_ubuf_put(ubufs);
 155         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 156 }
 157
 158 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 159 {
 160         vhost_net_ubuf_put_and_wait(ubufs);
 161         kfree(ubufs);
 162 }
 163
 164 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 165 {
 166         int i;
 167
 168         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 169                 kfree(n->vqs[i].ubuf_info);
 170                 n->vqs[i].ubuf_info = NULL;
 171         }
 172 }
 173
 174 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 175 {
 176         bool zcopy;
 177         int i;
 178
 179         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 180                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 181                 if (!zcopy)
 182                         continue;
 183                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 184                                               UIO_MAXIOV, GFP_KERNEL);
 185                 if  (!n->vqs[i].ubuf_info)
 186                         goto err;
 187         }
 188         return 0;
 189
 190 err:
 191         vhost_net_clear_ubuf_info(n);
 192         return -ENOMEM;
 193 }
 194
 195 static void vhost_net_vq_reset(struct vhost_net *n)
 196 {
 197         int i;
 198
 199         vhost_net_clear_ubuf_info(n);
 200
 201         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 202                 n->vqs[i].done_idx = 0;
 203                 n->vqs[i].upend_idx = 0;
 204                 n->vqs[i].ubufs = NULL;
 205                 n->vqs[i].vhost_hlen = 0;
 206                 n->vqs[i].sock_hlen = 0;
 207         }
 208
 209 }
 210
 211 static void vhost_net_tx_packet(struct vhost_net *net)
 212 {
 213         ++net->tx_packets;
 214         if (net->tx_packets < 1024)
 215                 return;
 216         net->tx_packets = 0;
 217         net->tx_zcopy_err = 0;
 218 }
 219
 220 static void vhost_net_tx_err(struct vhost_net *net)
 221 {
 222         ++net->tx_zcopy_err;
 223 }
 224
 225 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 226 {
 227         /* TX flush waits for outstanding DMAs to be done.
 228          * Don't start new DMAs.
 229          */
 230         return !net->tx_flush &&
 231                 net->tx_packets / 64 >= net->tx_zcopy_err;
 232 }
 233
 234 static bool vhost_sock_zcopy(struct socket *sock)
 235 {
 236         return unlikely(experimental_zcopytx) &&
 237                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 238 }
 239
 240 /* In case of DMA done not in order in lower device driver for some reason.
 241  * upend_idx is used to track end of used idx, done_idx is used to track head
 242  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 243  * guest used idx.
 244  */
 245 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 246                                        struct vhost_virtqueue *vq)
 247 {
 248         struct vhost_net_virtqueue *nvq =
 249                 container_of(vq, struct vhost_net_virtqueue, vq);
 250         int i, add;
 251         int j = 0;
 252
 253         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 254                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 255                         vhost_net_tx_err(net);
 256                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 257                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 258                         ++j;
 259                 } else
 260                         break;
 261         }
 262         while (j) {
 263                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 264                 vhost_add_used_and_signal_n(vq->dev, vq,
 265                                             &vq->heads[nvq->done_idx], add);
 266                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 267                 j -= add;
 268         }
 269 }
 270
 271 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 272 {
 273         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 274         struct vhost_virtqueue *vq = ubufs->vq;
 275         int cnt;
 276
 277         rcu_read_lock_bh();
 278
 279         /* set len to mark this desc buffers done DMA */
 280         vq->heads[ubuf->desc].len = success ?
 281                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 282         cnt = vhost_net_ubuf_put(ubufs);
 283
 284         /*
 285          * Trigger polling thread if guest stopped submitting new buffers:
 286          * in this case, the refcount after decrement will eventually reach 1.
 287          * We also trigger polling periodically after each 16 packets
 288          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 289          * less than 10% of times).
 290          */
 291         if (cnt <= 1 || !(cnt % 16))
 292                 vhost_poll_queue(&vq->poll);
 293
 294         rcu_read_unlock_bh();
 295 }
 296
 297 static inline unsigned long busy_clock(void)
 298 {
 299         return local_clock() >> 10;
 300 }
 301
 302 static bool vhost_can_busy_poll(struct vhost_dev *dev,
 303                                 unsigned long endtime)
 304 {
 305         return likely(!need_resched()) &&
 306                likely(!time_after(busy_clock(), endtime)) &&
 307                likely(!signal_pending(current)) &&
 308                !vhost_has_work(dev);
 309 }
 310
 311 static void vhost_net_disable_vq(struct vhost_net *n,
 312                                  struct vhost_virtqueue *vq)
 313 {
 314         struct vhost_net_virtqueue *nvq =
 315                 container_of(vq, struct vhost_net_virtqueue, vq);
 316         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 317         if (!vq->private_data)
 318                 return;
 319         vhost_poll_stop(poll);
 320 }
 321
 322 static int vhost_net_enable_vq(struct vhost_net *n,
 323                                 struct vhost_virtqueue *vq)
 324 {
 325         struct vhost_net_virtqueue *nvq =
 326                 container_of(vq, struct vhost_net_virtqueue, vq);
 327         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 328         struct socket *sock;
 329
 330         sock = vq->private_data;
 331         if (!sock)
 332                 return 0;
 333
 334         return vhost_poll_start(poll, sock->file);
 335 }
 336
 337 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 338                                     struct vhost_virtqueue *vq,
 339                                     struct iovec iov[], unsigned int iov_size,
 340                                     unsigned int *out_num, unsigned int *in_num)
 341 {
 342         unsigned long uninitialized_var(endtime);
 343         int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 344                                   out_num, in_num, NULL, NULL);
 345
 346         if (r == vq->num && vq->busyloop_timeout) {
 347                 preempt_disable();
 348                 endtime = busy_clock() + vq->busyloop_timeout;
 349                 while (vhost_can_busy_poll(vq->dev, endtime) &&
 350                        vhost_vq_avail_empty(vq->dev, vq))
 351                         cpu_relax_lowlatency();
 352                 preempt_enable();
 353                 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 354                                       out_num, in_num, NULL, NULL);
 355         }
 356
 357         return r;
 358 }
 359
 360 /* Expects to be always run from workqueue - which acts as
 361  * read-size critical section for our kind of RCU. */
 362 static void handle_tx(struct vhost_net *net)
 363 {
 364         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 365         struct vhost_virtqueue *vq = &nvq->vq;
 366         unsigned out, in;
 367         int head;
 368         struct msghdr msg = {
 369                 .msg_name = NULL,
 370                 .msg_namelen = 0,
 371                 .msg_control = NULL,
 372                 .msg_controllen = 0,
 373                 .msg_flags = MSG_DONTWAIT,
 374         };
 375         size_t len, total_len = 0;
 376         int err;
 377         size_t hdr_size;
 378         struct socket *sock;
 379         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 380         struct ubuf_info *ubuf;
 381         bool zcopy, zcopy_used;
 382         int sent_pkts = 0;
 383
 384         mutex_lock(&vq->mutex);
 385         sock = vq->private_data;
 386         if (!sock)
 387                 goto out;
 388
 389         if (!vq_iotlb_prefetch(vq))
 390                 goto out;
 391
 392         vhost_disable_notify(&net->dev, vq);
 393
 394         hdr_size = nvq->vhost_hlen;
 395         zcopy = nvq->ubufs;
 396
 397         do {
 398                 /* Release DMAs done buffers first */
 399                 if (zcopy)
 400                         vhost_zerocopy_signal_used(net, vq);
 401
 402                 /* If more outstanding DMAs, queue the work.
 403                  * Handle upend_idx wrap around
 404                  */
 405                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
 406                               % UIO_MAXIOV == nvq->done_idx))
 407                         break;
 408
 409                 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
 410                                                 ARRAY_SIZE(vq->iov),
 411                                                 &out, &in);
 412                 /* On error, stop handling until the next kick. */
 413                 if (unlikely(head < 0))
 414                         break;
 415                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 416                 if (head == vq->num) {
 417                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 418                                 vhost_disable_notify(&net->dev, vq);
 419                                 continue;
 420                         }
 421                         break;
 422                 }
 423                 if (in) {
 424                         vq_err(vq, "Unexpected descriptor format for TX: "
 425                                "out %d, int %d\n", out, in);
 426                         break;
 427                 }
 428                 /* Skip header. TODO: support TSO. */
 429                 len = iov_length(vq->iov, out);
 430                 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
 431                 iov_iter_advance(&msg.msg_iter, hdr_size);
 432                 /* Sanity check */
 433                 if (!msg_data_left(&msg)) {
 434                         vq_err(vq, "Unexpected header len for TX: "
 435                                "%zd expected %zd\n",
 436                                len, hdr_size);
 437                         break;
 438                 }
 439                 len = msg_data_left(&msg);
 440
 441                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 442                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 443                                       nvq->done_idx
 444                                    && vhost_net_tx_select_zcopy(net);
 445
 446                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 447                 if (zcopy_used) {
 448                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 449                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 450                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 451                         ubuf->callback = vhost_zerocopy_callback;
 452                         ubuf->ctx = nvq->ubufs;
 453                         ubuf->desc = nvq->upend_idx;
 454                         msg.msg_control = ubuf;
 455                         msg.msg_controllen = sizeof(ubuf);
 456                         ubufs = nvq->ubufs;
 457                         atomic_inc(&ubufs->refcount);
 458                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 459                 } else {
 460                         msg.msg_control = NULL;
 461                         ubufs = NULL;
 462                 }
 463                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 464                 err = sock->ops->sendmsg(sock, &msg, len);
 465                 if (unlikely(err < 0)) {
 466                         if (zcopy_used) {
 467                                 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
 468                                         vhost_net_ubuf_put(ubufs);
 469                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 470                                         % UIO_MAXIOV;
 471                         }
 472                         vhost_discard_vq_desc(vq, 1);
 473                         break;
 474                 }
 475                 if (err != len)
 476                         pr_debug("Truncated TX packet: "
 477                                  " len %d != %zd\n", err, len);
 478                 if (!zcopy_used)
 479                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 480                 else
 481                         vhost_zerocopy_signal_used(net, vq);
 482                 total_len += len;
 483                 vhost_net_tx_packet(net);
 484         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 485 out:
 486         mutex_unlock(&vq->mutex);
 487 }
 488
 489 static int peek_head_len(struct sock *sk)
 490 {
 491         struct socket *sock = sk->sk_socket;
 492         struct sk_buff *head;
 493         int len = 0;
 494         unsigned long flags;
 495
 496         if (sock->ops->peek_len)
 497                 return sock->ops->peek_len(sock);
 498
 499         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 500         head = skb_peek(&sk->sk_receive_queue);
 501         if (likely(head)) {
 502                 len = head->len;
 503                 if (skb_vlan_tag_present(head))
 504                         len += VLAN_HLEN;
 505         }
 506
 507         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 508         return len;
 509 }
 510
 511 static int sk_has_rx_data(struct sock *sk)
 512 {
 513         struct socket *sock = sk->sk_socket;
 514
 515         if (sock->ops->peek_len)
 516                 return sock->ops->peek_len(sock);
 517
 518         return skb_queue_empty(&sk->sk_receive_queue);
 519 }
 520
 521 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
 522 {
 523         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 524         struct vhost_virtqueue *vq = &nvq->vq;
 525         unsigned long uninitialized_var(endtime);
 526         int len = peek_head_len(sk);
 527
 528         if (!len && vq->busyloop_timeout) {
 529                 /* Both tx vq and rx socket were polled here */
 530                 mutex_lock_nested(&vq->mutex, 1);
 531                 vhost_disable_notify(&net->dev, vq);
 532
 533                 preempt_disable();
 534                 endtime = busy_clock() + vq->busyloop_timeout;
 535
 536                 while (vhost_can_busy_poll(&net->dev, endtime) &&
 537                        !sk_has_rx_data(sk) &&
 538                        vhost_vq_avail_empty(&net->dev, vq))
 539                         cpu_relax_lowlatency();
 540
 541                 preempt_enable();
 542
 543                 if (!vhost_vq_avail_empty(&net->dev, vq))
 544                         vhost_poll_queue(&vq->poll);
 545                 else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 546                         vhost_disable_notify(&net->dev, vq);
 547                         vhost_poll_queue(&vq->poll);
 548                 }
 549
 550                 mutex_unlock(&vq->mutex);
 551
 552                 len = peek_head_len(sk);
 553         }
 554
 555         return len;
 556 }
 557
 558 /* This is a multi-buffer version of vhost_get_desc, that works if
 559  *      vq has read descriptors only.
 560  * @vq          - the relevant virtqueue
 561  * @datalen     - data length we'll be reading
 562  * @iovcount    - returned count of io vectors we fill
 563  * @log         - vhost log
 564  * @log_num     - log offset
 565  * @quota       - headcount quota, 1 for big buffer
 566  *      returns number of buffer heads allocated, negative on error
 567  */
 568 static int get_rx_bufs(struct vhost_virtqueue *vq,
 569                        struct vring_used_elem *heads,
 570                        int datalen,
 571                        unsigned *iovcount,
 572                        struct vhost_log *log,
 573                        unsigned *log_num,
 574                        unsigned int quota)
 575 {
 576         unsigned int out, in;
 577         int seg = 0;
 578         int headcount = 0;
 579         unsigned d;
 580         int r, nlogs = 0;
 581         /* len is always initialized before use since we are always called with
 582          * datalen > 0.
 583          */
 584         u32 uninitialized_var(len);
 585
 586         while (datalen > 0 && headcount < quota) {
 587                 if (unlikely(seg >= UIO_MAXIOV)) {
 588                         r = -ENOBUFS;
 589                         goto err;
 590                 }
 591                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 592                                       ARRAY_SIZE(vq->iov) - seg, &out,
 593                                       &in, log, log_num);
 594                 if (unlikely(r < 0))
 595                         goto err;
 596
 597                 d = r;
 598                 if (d == vq->num) {
 599                         r = 0;
 600                         goto err;
 601                 }
 602                 if (unlikely(out || in <= 0)) {
 603                         vq_err(vq, "unexpected descriptor format for RX: "
 604                                 "out %d, in %d\n", out, in);
 605                         r = -EINVAL;
 606                         goto err;
 607                 }
 608                 if (unlikely(log)) {
 609                         nlogs += *log_num;
 610                         log += *log_num;
 611                 }
 612                 heads[headcount].id = cpu_to_vhost32(vq, d);
 613                 len = iov_length(vq->iov + seg, in);
 614                 heads[headcount].len = cpu_to_vhost32(vq, len);
 615                 datalen -= len;
 616                 ++headcount;
 617                 seg += in;
 618         }
 619         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 620         *iovcount = seg;
 621         if (unlikely(log))
 622                 *log_num = nlogs;
 623
 624         /* Detect overrun */
 625         if (unlikely(datalen > 0)) {
 626                 r = UIO_MAXIOV + 1;
 627                 goto err;
 628         }
 629         return headcount;
 630 err:
 631         vhost_discard_vq_desc(vq, headcount);
 632         return r;
 633 }
 634
 635 /* Expects to be always run from workqueue - which acts as
 636  * read-size critical section for our kind of RCU. */
 637 static void handle_rx(struct vhost_net *net)
 638 {
 639         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 640         struct vhost_virtqueue *vq = &nvq->vq;
 641         unsigned uninitialized_var(in), log;
 642         struct vhost_log *vq_log;
 643         struct msghdr msg = {
 644                 .msg_name = NULL,
 645                 .msg_namelen = 0,
 646                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 647                 .msg_controllen = 0,
 648                 .msg_flags = MSG_DONTWAIT,
 649         };
 650         struct virtio_net_hdr hdr = {
 651                 .flags = 0,
 652                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 653         };
 654         size_t total_len = 0;
 655         int err, mergeable;
 656         s16 headcount;
 657         size_t vhost_hlen, sock_hlen;
 658         size_t vhost_len, sock_len;
 659         struct socket *sock;
 660         struct iov_iter fixup;
 661         __virtio16 num_buffers;
 662         int recv_pkts = 0;
 663
 664         mutex_lock_nested(&vq->mutex, 0);
 665         sock = vq->private_data;
 666         if (!sock)
 667                 goto out;
 668
 669         if (!vq_iotlb_prefetch(vq))
 670                 goto out;
 671
 672         vhost_disable_notify(&net->dev, vq);
 673         vhost_net_disable_vq(net, vq);
 674
 675         vhost_hlen = nvq->vhost_hlen;
 676         sock_hlen = nvq->sock_hlen;
 677
 678         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 679                 vq->log : NULL;
 680         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 681
 682         do {
 683                 sock_len = vhost_net_rx_peek_head_len(net, sock->sk);
 684                 if (!sock_len)
 685                         break;
 686                 sock_len += sock_hlen;
 687                 vhost_len = sock_len + vhost_hlen;
 688                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 689                                         &in, vq_log, &log,
 690                                         likely(mergeable) ? UIO_MAXIOV : 1);
 691                 /* On error, stop handling until the next kick. */
 692                 if (unlikely(headcount < 0))
 693                         goto out;
 694                 /* On overrun, truncate and discard */
 695                 if (unlikely(headcount > UIO_MAXIOV)) {
 696                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 697                         err = sock->ops->recvmsg(sock, &msg,
 698                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 699                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 700                         continue;
 701                 }
 702                 /* OK, now we need to know about added descriptors. */
 703                 if (!headcount) {
 704                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 705                                 /* They have slipped one in as we were
 706                                  * doing that: check again. */
 707                                 vhost_disable_notify(&net->dev, vq);
 708                                 continue;
 709                         }
 710                         /* Nothing new?  Wait for eventfd to tell us
 711                          * they refilled. */
 712                         goto out;
 713                 }
 714                 /* We don't need to be notified again. */
 715                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
 716                 fixup = msg.msg_iter;
 717                 if (unlikely((vhost_hlen))) {
 718                         /* We will supply the header ourselves
 719                          * TODO: support TSO.
 720                          */
 721                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
 722                 }
 723                 err = sock->ops->recvmsg(sock, &msg,
 724                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 725                 /* Userspace might have consumed the packet meanwhile:
 726                  * it's not supposed to do this usually, but might be hard
 727                  * to prevent. Discard data we got (if any) and keep going. */
 728                 if (unlikely(err != sock_len)) {
 729                         pr_debug("Discarded rx packet: "
 730                                  " len %d, expected %zd\n", err, sock_len);
 731                         vhost_discard_vq_desc(vq, headcount);
 732                         continue;
 733                 }
 734                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
 735                 if (unlikely(vhost_hlen)) {
 736                         if (copy_to_iter(&hdr, sizeof(hdr),
 737                                          &fixup) != sizeof(hdr)) {
 738                                 vq_err(vq, "Unable to write vnet_hdr "
 739                                        "at addr %p\n", vq->iov->iov_base);
 740                                 goto out;
 741                         }
 742                 } else {
 743                         /* Header came from socket; we'll need to patch
 744                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
 745                          */
 746                         iov_iter_advance(&fixup, sizeof(hdr));
 747                 }
 748                 /* TODO: Should check and handle checksum. */
 749
 750                 num_buffers = cpu_to_vhost16(vq, headcount);
 751                 if (likely(mergeable) &&
 752                     copy_to_iter(&num_buffers, sizeof num_buffers,
 753                                  &fixup) != sizeof num_buffers) {
 754                         vq_err(vq, "Failed num_buffers write");
 755                         vhost_discard_vq_desc(vq, headcount);
 756                         goto out;
 757                 }
 758                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 759                                             headcount);
 760                 if (unlikely(vq_log))
 761                         vhost_log_write(vq, vq_log, log, vhost_len,
 762                                         vq->iov, in);
 763                 total_len += vhost_len;
 764         } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
 765
 766         if (!sock_len)
 767                 vhost_net_enable_vq(net, vq);
 768 out:
 769         mutex_unlock(&vq->mutex);
 770 }
 771
 772 static void handle_tx_kick(struct vhost_work *work)
 773 {
 774         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 775                                                   poll.work);
 776         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 777
 778         handle_tx(net);
 779 }
 780
 781 static void handle_rx_kick(struct vhost_work *work)
 782 {
 783         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 784                                                   poll.work);
 785         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 786
 787         handle_rx(net);
 788 }
 789
 790 static void handle_tx_net(struct vhost_work *work)
 791 {
 792         struct vhost_net *net = container_of(work, struct vhost_net,
 793                                              poll[VHOST_NET_VQ_TX].work);
 794         handle_tx(net);
 795 }
 796
 797 static void handle_rx_net(struct vhost_work *work)
 798 {
 799         struct vhost_net *net = container_of(work, struct vhost_net,
 800                                              poll[VHOST_NET_VQ_RX].work);
 801         handle_rx(net);
 802 }
 803
 804 static int vhost_net_open(struct inode *inode, struct file *f)
 805 {
 806         struct vhost_net *n;
 807         struct vhost_dev *dev;
 808         struct vhost_virtqueue **vqs;
 809         int i;
 810
 811         n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 812         if (!n) {
 813                 n = vmalloc(sizeof *n);
 814                 if (!n)
 815                         return -ENOMEM;
 816         }
 817         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 818         if (!vqs) {
 819                 kvfree(n);
 820                 return -ENOMEM;
 821         }
 822
 823         dev = &n->dev;
 824         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 825         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 826         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 827         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 828         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 829                 n->vqs[i].ubufs = NULL;
 830                 n->vqs[i].ubuf_info = NULL;
 831                 n->vqs[i].upend_idx = 0;
 832                 n->vqs[i].done_idx = 0;
 833                 n->vqs[i].vhost_hlen = 0;
 834                 n->vqs[i].sock_hlen = 0;
 835         }
 836         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
 837                        VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT);
 838
 839         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 840         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 841
 842         f->private_data = n;
 843
 844         return 0;
 845 }
 846
 847 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 848                                         struct vhost_virtqueue *vq)
 849 {
 850         struct socket *sock;
 851
 852         mutex_lock(&vq->mutex);
 853         sock = vq->private_data;
 854         vhost_net_disable_vq(n, vq);
 855         vq->private_data = NULL;
 856         mutex_unlock(&vq->mutex);
 857         return sock;
 858 }
 859
 860 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 861                            struct socket **rx_sock)
 862 {
 863         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 864         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 865 }
 866
 867 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 868 {
 869         vhost_poll_flush(n->poll + index);
 870         vhost_poll_flush(&n->vqs[index].vq.poll);
 871 }
 872
 873 static void vhost_net_flush(struct vhost_net *n)
 874 {
 875         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 876         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 877         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 878                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 879                 n->tx_flush = true;
 880                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 881                 /* Wait for all lower device DMAs done. */
 882                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 883                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 884                 n->tx_flush = false;
 885                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 886                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 887         }
 888 }
 889
 890 static int vhost_net_release(struct inode *inode, struct file *f)
 891 {
 892         struct vhost_net *n = f->private_data;
 893         struct socket *tx_sock;
 894         struct socket *rx_sock;
 895
 896         vhost_net_stop(n, &tx_sock, &rx_sock);
 897         vhost_net_flush(n);
 898         vhost_dev_stop(&n->dev);
 899         vhost_dev_cleanup(&n->dev, false);
 900         vhost_net_vq_reset(n);
 901         if (tx_sock)
 902                 sockfd_put(tx_sock);
 903         if (rx_sock)
 904                 sockfd_put(rx_sock);
 905         /* Make sure no callbacks are outstanding */
 906         synchronize_rcu_bh();
 907         /* We do an extra flush before freeing memory,
 908          * since jobs can re-queue themselves. */
 909         vhost_net_flush(n);
 910         kfree(n->dev.vqs);
 911         kvfree(n);
 912         return 0;
 913 }
 914
 915 static struct socket *get_raw_socket(int fd)
 916 {
 917         int r;
 918         struct socket *sock = sockfd_lookup(fd, &r);
 919
 920         if (!sock)
 921                 return ERR_PTR(-ENOTSOCK);
 922
 923         /* Parameter checking */
 924         if (sock->sk->sk_type != SOCK_RAW) {
 925                 r = -ESOCKTNOSUPPORT;
 926                 goto err;
 927         }
 928
 929         if (sock->sk->sk_family != AF_PACKET) {
 930                 r = -EPFNOSUPPORT;
 931                 goto err;
 932         }
 933         return sock;
 934 err:
 935         sockfd_put(sock);
 936         return ERR_PTR(r);
 937 }
 938
 939 static struct socket *get_tap_socket(int fd)
 940 {
 941         struct file *file = fget(fd);
 942         struct socket *sock;
 943
 944         if (!file)
 945                 return ERR_PTR(-EBADF);
 946         sock = tun_get_socket(file);
 947         if (!IS_ERR(sock))
 948                 return sock;
 949         sock = macvtap_get_socket(file);
 950         if (IS_ERR(sock))
 951                 fput(file);
 952         return sock;
 953 }
 954
 955 static struct socket *get_socket(int fd)
 956 {
 957         struct socket *sock;
 958
 959         /* special case to disable backend */
 960         if (fd == -1)
 961                 return NULL;
 962         sock = get_raw_socket(fd);
 963         if (!IS_ERR(sock))
 964                 return sock;
 965         sock = get_tap_socket(fd);
 966         if (!IS_ERR(sock))
 967                 return sock;
 968         return ERR_PTR(-ENOTSOCK);
 969 }
 970
 971 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 972 {
 973         struct socket *sock, *oldsock;
 974         struct vhost_virtqueue *vq;
 975         struct vhost_net_virtqueue *nvq;
 976         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 977         int r;
 978
 979         mutex_lock(&n->dev.mutex);
 980         r = vhost_dev_check_owner(&n->dev);
 981         if (r)
 982                 goto err;
 983
 984         if (index >= VHOST_NET_VQ_MAX) {
 985                 r = -ENOBUFS;
 986                 goto err;
 987         }
 988         vq = &n->vqs[index].vq;
 989         nvq = &n->vqs[index];
 990         mutex_lock(&vq->mutex);
 991
 992         /* Verify that ring has been setup correctly. */
 993         if (!vhost_vq_access_ok(vq)) {
 994                 r = -EFAULT;
 995                 goto err_vq;
 996         }
 997         sock = get_socket(fd);
 998         if (IS_ERR(sock)) {
 999                 r = PTR_ERR(sock);
1000                 goto err_vq;
1001         }
1002
1003         /* start polling new socket */
1004         oldsock = vq->private_data;
1005         if (sock != oldsock) {
1006                 ubufs = vhost_net_ubuf_alloc(vq,
1007                                              sock && vhost_sock_zcopy(sock));
1008                 if (IS_ERR(ubufs)) {
1009                         r = PTR_ERR(ubufs);
1010                         goto err_ubufs;
1011                 }
1012
1013                 vhost_net_disable_vq(n, vq);
1014                 vq->private_data = sock;
1015                 r = vhost_vq_init_access(vq);
1016                 if (r)
1017                         goto err_used;
1018                 r = vhost_net_enable_vq(n, vq);
1019                 if (r)
1020                         goto err_used;
1021
1022                 oldubufs = nvq->ubufs;
1023                 nvq->ubufs = ubufs;
1024
1025                 n->tx_packets = 0;
1026                 n->tx_zcopy_err = 0;
1027                 n->tx_flush = false;
1028         }
1029
1030         mutex_unlock(&vq->mutex);
1031
1032         if (oldubufs) {
1033                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1034                 mutex_lock(&vq->mutex);
1035                 vhost_zerocopy_signal_used(n, vq);
1036                 mutex_unlock(&vq->mutex);
1037         }
1038
1039         if (oldsock) {
1040                 vhost_net_flush_vq(n, index);
1041                 sockfd_put(oldsock);
1042         }
1043
1044         mutex_unlock(&n->dev.mutex);
1045         return 0;
1046
1047 err_used:
1048         vq->private_data = oldsock;
1049         vhost_net_enable_vq(n, vq);
1050         if (ubufs)
1051                 vhost_net_ubuf_put_wait_and_free(ubufs);
1052 err_ubufs:
1053         if (sock)
1054                 sockfd_put(sock);
1055 err_vq:
1056         mutex_unlock(&vq->mutex);
1057 err:
1058         mutex_unlock(&n->dev.mutex);
1059         return r;
1060 }
1061
1062 static long vhost_net_reset_owner(struct vhost_net *n)
1063 {
1064         struct socket *tx_sock = NULL;
1065         struct socket *rx_sock = NULL;
1066         long err;
1067         struct vhost_umem *umem;
1068
1069         mutex_lock(&n->dev.mutex);
1070         err = vhost_dev_check_owner(&n->dev);
1071         if (err)
1072                 goto done;
1073         umem = vhost_dev_reset_owner_prepare();
1074         if (!umem) {
1075                 err = -ENOMEM;
1076                 goto done;
1077         }
1078         vhost_net_stop(n, &tx_sock, &rx_sock);
1079         vhost_net_flush(n);
1080         vhost_dev_stop(&n->dev);
1081         vhost_dev_reset_owner(&n->dev, umem);
1082         vhost_net_vq_reset(n);
1083 done:
1084         mutex_unlock(&n->dev.mutex);
1085         if (tx_sock)
1086                 sockfd_put(tx_sock);
1087         if (rx_sock)
1088                 sockfd_put(rx_sock);
1089         return err;
1090 }
1091
1092 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1093 {
1094         size_t vhost_hlen, sock_hlen, hdr_len;
1095         int i;
1096
1097         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1098                                (1ULL << VIRTIO_F_VERSION_1))) ?
1099                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1100                         sizeof(struct virtio_net_hdr);
1101         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1102                 /* vhost provides vnet_hdr */
1103                 vhost_hlen = hdr_len;
1104                 sock_hlen = 0;
1105         } else {
1106                 /* socket provides vnet_hdr */
1107                 vhost_hlen = 0;
1108                 sock_hlen = hdr_len;
1109         }
1110         mutex_lock(&n->dev.mutex);
1111         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1112             !vhost_log_access_ok(&n->dev))
1113                 goto out_unlock;
1114
1115         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1116                 if (vhost_init_device_iotlb(&n->dev, true))
1117                         goto out_unlock;
1118         }
1119
1120         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1121                 mutex_lock(&n->vqs[i].vq.mutex);
1122                 n->vqs[i].vq.acked_features = features;
1123                 n->vqs[i].vhost_hlen = vhost_hlen;
1124                 n->vqs[i].sock_hlen = sock_hlen;
1125                 mutex_unlock(&n->vqs[i].vq.mutex);
1126         }
1127         mutex_unlock(&n->dev.mutex);
1128         return 0;
1129
1130 out_unlock:
1131         mutex_unlock(&n->dev.mutex);
1132         return -EFAULT;
1133 }
1134
1135 static long vhost_net_set_owner(struct vhost_net *n)
1136 {
1137         int r;
1138
1139         mutex_lock(&n->dev.mutex);
1140         if (vhost_dev_has_owner(&n->dev)) {
1141                 r = -EBUSY;
1142                 goto out;
1143         }
1144         r = vhost_net_set_ubuf_info(n);
1145         if (r)
1146                 goto out;
1147         r = vhost_dev_set_owner(&n->dev);
1148         if (r)
1149                 vhost_net_clear_ubuf_info(n);
1150         vhost_net_flush(n);
1151 out:
1152         mutex_unlock(&n->dev.mutex);
1153         return r;
1154 }
1155
1156 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1157                             unsigned long arg)
1158 {
1159         struct vhost_net *n = f->private_data;
1160         void __user *argp = (void __user *)arg;
1161         u64 __user *featurep = argp;
1162         struct vhost_vring_file backend;
1163         u64 features;
1164         int r;
1165
1166         switch (ioctl) {
1167         case VHOST_NET_SET_BACKEND:
1168                 if (copy_from_user(&backend, argp, sizeof backend))
1169                         return -EFAULT;
1170                 return vhost_net_set_backend(n, backend.index, backend.fd);
1171         case VHOST_GET_FEATURES:
1172                 features = VHOST_NET_FEATURES;
1173                 if (copy_to_user(featurep, &features, sizeof features))
1174                         return -EFAULT;
1175                 return 0;
1176         case VHOST_SET_FEATURES:
1177                 if (copy_from_user(&features, featurep, sizeof features))
1178                         return -EFAULT;
1179                 if (features & ~VHOST_NET_FEATURES)
1180                         return -EOPNOTSUPP;
1181                 return vhost_net_set_features(n, features);
1182         case VHOST_RESET_OWNER:
1183                 return vhost_net_reset_owner(n);
1184         case VHOST_SET_OWNER:
1185                 return vhost_net_set_owner(n);
1186         default:
1187                 mutex_lock(&n->dev.mutex);
1188                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1189                 if (r == -ENOIOCTLCMD)
1190                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1191                 else
1192                         vhost_net_flush(n);
1193                 mutex_unlock(&n->dev.mutex);
1194                 return r;
1195         }
1196 }
1197
1198 #ifdef CONFIG_COMPAT
1199 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1200                                    unsigned long arg)
1201 {
1202         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1203 }
1204 #endif
1205
1206 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1207 {
1208         struct file *file = iocb->ki_filp;
1209         struct vhost_net *n = file->private_data;
1210         struct vhost_dev *dev = &n->dev;
1211         int noblock = file->f_flags & O_NONBLOCK;
1212
1213         return vhost_chr_read_iter(dev, to, noblock);
1214 }
1215
1216 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1217                                         struct iov_iter *from)
1218 {
1219         struct file *file = iocb->ki_filp;
1220         struct vhost_net *n = file->private_data;
1221         struct vhost_dev *dev = &n->dev;
1222
1223         return vhost_chr_write_iter(dev, from);
1224 }
1225
1226 static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait)
1227 {
1228         struct vhost_net *n = file->private_data;
1229         struct vhost_dev *dev = &n->dev;
1230
1231         return vhost_chr_poll(file, dev, wait);
1232 }
1233
1234 static const struct file_operations vhost_net_fops = {
1235         .owner          = THIS_MODULE,
1236         .release        = vhost_net_release,
1237         .read_iter      = vhost_net_chr_read_iter,
1238         .write_iter     = vhost_net_chr_write_iter,
1239         .poll           = vhost_net_chr_poll,
1240         .unlocked_ioctl = vhost_net_ioctl,
1241 #ifdef CONFIG_COMPAT
1242         .compat_ioctl   = vhost_net_compat_ioctl,
1243 #endif
1244         .open           = vhost_net_open,
1245         .llseek         = noop_llseek,
1246 };
1247
1248 static struct miscdevice vhost_net_misc = {
1249         .minor = VHOST_NET_MINOR,
1250         .name = "vhost-net",
1251         .fops = &vhost_net_fops,
1252 };
1253
1254 static int vhost_net_init(void)
1255 {
1256         if (experimental_zcopytx)
1257                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1258         return misc_register(&vhost_net_misc);
1259 }
1260 module_init(vhost_net_init);
1261
1262 static void vhost_net_exit(void)
1263 {
1264         misc_deregister(&vhost_net_misc);
1265 }
1266 module_exit(vhost_net_exit);
1267
1268 MODULE_VERSION("0.0.1");
1269 MODULE_LICENSE("GPL v2");
1270 MODULE_AUTHOR("Michael S. Tsirkin");
1271 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1272 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1273 MODULE_ALIAS("devname:vhost-net");