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/sched/clock.h>
  21 #include <linux/sched/signal.h>
  22 #include <linux/vmalloc.h>
  23
  24 #include <linux/net.h>
  25 #include <linux/if_packet.h>
  26 #include <linux/if_arp.h>
  27 #include <linux/if_tun.h>
  28 #include <linux/if_macvlan.h>
  29 #include <linux/if_tap.h>
  30 #include <linux/if_vlan.h>
  31 #include <linux/skb_array.h>
  32 #include <linux/skbuff.h>
  33
  34 #include <net/sock.h>
  35
  36 #include "vhost.h"
  37
  38 static int experimental_zcopytx = 1;
  39 module_param(experimental_zcopytx, int, 0444);
  40 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  41                                        " 1 -Enable; 0 - Disable");
  42
  43 /* Max number of bytes transferred before requeueing the job.
  44  * Using this limit prevents one virtqueue from starving others. */
  45 #define VHOST_NET_WEIGHT 0x80000
  46
  47 /* MAX number of TX used buffers for outstanding zerocopy */
  48 #define VHOST_MAX_PEND 128
  49 #define VHOST_GOODCOPY_LEN 256
  50
  51 /*
  52  * For transmit, used buffer len is unused; we override it to track buffer
  53  * status internally; used for zerocopy tx only.
  54  */
  55 /* Lower device DMA failed */
  56 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  57 /* Lower device DMA done */
  58 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  59 /* Lower device DMA in progress */
  60 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  61 /* Buffer unused */
  62 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  63
  64 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  65
  66 enum {
  67         VHOST_NET_FEATURES = VHOST_FEATURES |
  68                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  69                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  70                          (1ULL << VIRTIO_F_IOMMU_PLATFORM)
  71 };
  72
  73 enum {
  74         VHOST_NET_VQ_RX = 0,
  75         VHOST_NET_VQ_TX = 1,
  76         VHOST_NET_VQ_MAX = 2,
  77 };
  78
  79 struct vhost_net_ubuf_ref {
  80         /* refcount follows semantics similar to kref:
  81          *  0: object is released
  82          *  1: no outstanding ubufs
  83          * >1: outstanding ubufs
  84          */
  85         atomic_t refcount;
  86         wait_queue_head_t wait;
  87         struct vhost_virtqueue *vq;
  88 };
  89
  90 #define VHOST_RX_BATCH 64
  91 struct vhost_net_buf {
  92         struct sk_buff **queue;
  93         int tail;
  94         int head;
  95 };
  96
  97 struct vhost_net_virtqueue {
  98         struct vhost_virtqueue vq;
  99         size_t vhost_hlen;
 100         size_t sock_hlen;
 101         /* vhost zerocopy support fields below: */
 102         /* last used idx for outstanding DMA zerocopy buffers */
 103         int upend_idx;
 104         /* first used idx for DMA done zerocopy buffers */
 105         int done_idx;
 106         /* an array of userspace buffers info */
 107         struct ubuf_info *ubuf_info;
 108         /* Reference counting for outstanding ubufs.
 109          * Protected by vq mutex. Writers must also take device mutex. */
 110         struct vhost_net_ubuf_ref *ubufs;
 111         struct skb_array *rx_array;
 112         struct vhost_net_buf rxq;
 113 };
 114
 115 struct vhost_net {
 116         struct vhost_dev dev;
 117         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 118         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 119         /* Number of TX recently submitted.
 120          * Protected by tx vq lock. */
 121         unsigned tx_packets;
 122         /* Number of times zerocopy TX recently failed.
 123          * Protected by tx vq lock. */
 124         unsigned tx_zcopy_err;
 125         /* Flush in progress. Protected by tx vq lock. */
 126         bool tx_flush;
 127 };
 128
 129 static unsigned vhost_net_zcopy_mask __read_mostly;
 130
 131 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 132 {
 133         if (rxq->tail != rxq->head)
 134                 return rxq->queue[rxq->head];
 135         else
 136                 return NULL;
 137 }
 138
 139 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 140 {
 141         return rxq->tail - rxq->head;
 142 }
 143
 144 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 145 {
 146         return rxq->tail == rxq->head;
 147 }
 148
 149 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 150 {
 151         void *ret = vhost_net_buf_get_ptr(rxq);
 152         ++rxq->head;
 153         return ret;
 154 }
 155
 156 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 157 {
 158         struct vhost_net_buf *rxq = &nvq->rxq;
 159
 160         rxq->head = 0;
 161         rxq->tail = skb_array_consume_batched(nvq->rx_array, rxq->queue,
 162                                               VHOST_RX_BATCH);
 163         return rxq->tail;
 164 }
 165
 166 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 167 {
 168         struct vhost_net_buf *rxq = &nvq->rxq;
 169
 170         if (nvq->rx_array && !vhost_net_buf_is_empty(rxq)) {
 171                 skb_array_unconsume(nvq->rx_array, rxq->queue + rxq->head,
 172                                     vhost_net_buf_get_size(rxq));
 173                 rxq->head = rxq->tail = 0;
 174         }
 175 }
 176
 177 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 178 {
 179         struct vhost_net_buf *rxq = &nvq->rxq;
 180
 181         if (!vhost_net_buf_is_empty(rxq))
 182                 goto out;
 183
 184         if (!vhost_net_buf_produce(nvq))
 185                 return 0;
 186
 187 out:
 188         return __skb_array_len_with_tag(vhost_net_buf_get_ptr(rxq));
 189 }
 190
 191 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 192 {
 193         rxq->head = rxq->tail = 0;
 194 }
 195
 196 static void vhost_net_enable_zcopy(int vq)
 197 {
 198         vhost_net_zcopy_mask |= 0x1 << vq;
 199 }
 200
 201 static struct vhost_net_ubuf_ref *
 202 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 203 {
 204         struct vhost_net_ubuf_ref *ubufs;
 205         /* No zero copy backend? Nothing to count. */
 206         if (!zcopy)
 207                 return NULL;
 208         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 209         if (!ubufs)
 210                 return ERR_PTR(-ENOMEM);
 211         atomic_set(&ubufs->refcount, 1);
 212         init_waitqueue_head(&ubufs->wait);
 213         ubufs->vq = vq;
 214         return ubufs;
 215 }
 216
 217 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 218 {
 219         int r = atomic_sub_return(1, &ubufs->refcount);
 220         if (unlikely(!r))
 221                 wake_up(&ubufs->wait);
 222         return r;
 223 }
 224
 225 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 226 {
 227         vhost_net_ubuf_put(ubufs);
 228         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 229 }
 230
 231 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 232 {
 233         vhost_net_ubuf_put_and_wait(ubufs);
 234         kfree(ubufs);
 235 }
 236
 237 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 238 {
 239         int i;
 240
 241         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 242                 kfree(n->vqs[i].ubuf_info);
 243                 n->vqs[i].ubuf_info = NULL;
 244         }
 245 }
 246
 247 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 248 {
 249         bool zcopy;
 250         int i;
 251
 252         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 253                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 254                 if (!zcopy)
 255                         continue;
 256                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 257                                               UIO_MAXIOV, GFP_KERNEL);
 258                 if  (!n->vqs[i].ubuf_info)
 259                         goto err;
 260         }
 261         return 0;
 262
 263 err:
 264         vhost_net_clear_ubuf_info(n);
 265         return -ENOMEM;
 266 }
 267
 268 static void vhost_net_vq_reset(struct vhost_net *n)
 269 {
 270         int i;
 271
 272         vhost_net_clear_ubuf_info(n);
 273
 274         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 275                 n->vqs[i].done_idx = 0;
 276                 n->vqs[i].upend_idx = 0;
 277                 n->vqs[i].ubufs = NULL;
 278                 n->vqs[i].vhost_hlen = 0;
 279                 n->vqs[i].sock_hlen = 0;
 280                 vhost_net_buf_init(&n->vqs[i].rxq);
 281         }
 282
 283 }
 284
 285 static void vhost_net_tx_packet(struct vhost_net *net)
 286 {
 287         ++net->tx_packets;
 288         if (net->tx_packets < 1024)
 289                 return;
 290         net->tx_packets = 0;
 291         net->tx_zcopy_err = 0;
 292 }
 293
 294 static void vhost_net_tx_err(struct vhost_net *net)
 295 {
 296         ++net->tx_zcopy_err;
 297 }
 298
 299 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 300 {
 301         /* TX flush waits for outstanding DMAs to be done.
 302          * Don't start new DMAs.
 303          */
 304         return !net->tx_flush &&
 305                 net->tx_packets / 64 >= net->tx_zcopy_err;
 306 }
 307
 308 static bool vhost_sock_zcopy(struct socket *sock)
 309 {
 310         return unlikely(experimental_zcopytx) &&
 311                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 312 }
 313
 314 /* In case of DMA done not in order in lower device driver for some reason.
 315  * upend_idx is used to track end of used idx, done_idx is used to track head
 316  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 317  * guest used idx.
 318  */
 319 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 320                                        struct vhost_virtqueue *vq)
 321 {
 322         struct vhost_net_virtqueue *nvq =
 323                 container_of(vq, struct vhost_net_virtqueue, vq);
 324         int i, add;
 325         int j = 0;
 326
 327         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 328                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 329                         vhost_net_tx_err(net);
 330                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 331                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 332                         ++j;
 333                 } else
 334                         break;
 335         }
 336         while (j) {
 337                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 338                 vhost_add_used_and_signal_n(vq->dev, vq,
 339                                             &vq->heads[nvq->done_idx], add);
 340                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 341                 j -= add;
 342         }
 343 }
 344
 345 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 346 {
 347         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 348         struct vhost_virtqueue *vq = ubufs->vq;
 349         int cnt;
 350
 351         rcu_read_lock_bh();
 352
 353         /* set len to mark this desc buffers done DMA */
 354         vq->heads[ubuf->desc].len = success ?
 355                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 356         cnt = vhost_net_ubuf_put(ubufs);
 357
 358         /*
 359          * Trigger polling thread if guest stopped submitting new buffers:
 360          * in this case, the refcount after decrement will eventually reach 1.
 361          * We also trigger polling periodically after each 16 packets
 362          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 363          * less than 10% of times).
 364          */
 365         if (cnt <= 1 || !(cnt % 16))
 366                 vhost_poll_queue(&vq->poll);
 367
 368         rcu_read_unlock_bh();
 369 }
 370
 371 static inline unsigned long busy_clock(void)
 372 {
 373         return local_clock() >> 10;
 374 }
 375
 376 static bool vhost_can_busy_poll(struct vhost_dev *dev,
 377                                 unsigned long endtime)
 378 {
 379         return likely(!need_resched()) &&
 380                likely(!time_after(busy_clock(), endtime)) &&
 381                likely(!signal_pending(current)) &&
 382                !vhost_has_work(dev);
 383 }
 384
 385 static void vhost_net_disable_vq(struct vhost_net *n,
 386                                  struct vhost_virtqueue *vq)
 387 {
 388         struct vhost_net_virtqueue *nvq =
 389                 container_of(vq, struct vhost_net_virtqueue, vq);
 390         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 391         if (!vq->private_data)
 392                 return;
 393         vhost_poll_stop(poll);
 394 }
 395
 396 static int vhost_net_enable_vq(struct vhost_net *n,
 397                                 struct vhost_virtqueue *vq)
 398 {
 399         struct vhost_net_virtqueue *nvq =
 400                 container_of(vq, struct vhost_net_virtqueue, vq);
 401         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 402         struct socket *sock;
 403
 404         sock = vq->private_data;
 405         if (!sock)
 406                 return 0;
 407
 408         return vhost_poll_start(poll, sock->file);
 409 }
 410
 411 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 412                                     struct vhost_virtqueue *vq,
 413                                     struct iovec iov[], unsigned int iov_size,
 414                                     unsigned int *out_num, unsigned int *in_num)
 415 {
 416         unsigned long uninitialized_var(endtime);
 417         int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 418                                   out_num, in_num, NULL, NULL);
 419
 420         if (r == vq->num && vq->busyloop_timeout) {
 421                 preempt_disable();
 422                 endtime = busy_clock() + vq->busyloop_timeout;
 423                 while (vhost_can_busy_poll(vq->dev, endtime) &&
 424                        vhost_vq_avail_empty(vq->dev, vq))
 425                         cpu_relax();
 426                 preempt_enable();
 427                 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 428                                       out_num, in_num, NULL, NULL);
 429         }
 430
 431         return r;
 432 }
 433
 434 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 435 {
 436         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 437         struct vhost_virtqueue *vq = &nvq->vq;
 438
 439         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 440                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 441 }
 442
 443 /* Expects to be always run from workqueue - which acts as
 444  * read-size critical section for our kind of RCU. */
 445 static void handle_tx(struct vhost_net *net)
 446 {
 447         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 448         struct vhost_virtqueue *vq = &nvq->vq;
 449         unsigned out, in;
 450         int head;
 451         struct msghdr msg = {
 452                 .msg_name = NULL,
 453                 .msg_namelen = 0,
 454                 .msg_control = NULL,
 455                 .msg_controllen = 0,
 456                 .msg_flags = MSG_DONTWAIT,
 457         };
 458         size_t len, total_len = 0;
 459         int err;
 460         size_t hdr_size;
 461         struct socket *sock;
 462         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 463         bool zcopy, zcopy_used;
 464
 465         mutex_lock(&vq->mutex);
 466         sock = vq->private_data;
 467         if (!sock)
 468                 goto out;
 469
 470         if (!vq_iotlb_prefetch(vq))
 471                 goto out;
 472
 473         vhost_disable_notify(&net->dev, vq);
 474         vhost_net_disable_vq(net, vq);
 475
 476         hdr_size = nvq->vhost_hlen;
 477         zcopy = nvq->ubufs;
 478
 479         for (;;) {
 480                 /* Release DMAs done buffers first */
 481                 if (zcopy)
 482                         vhost_zerocopy_signal_used(net, vq);
 483
 484
 485                 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
 486                                                 ARRAY_SIZE(vq->iov),
 487                                                 &out, &in);
 488                 /* On error, stop handling until the next kick. */
 489                 if (unlikely(head < 0))
 490                         break;
 491                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 492                 if (head == vq->num) {
 493                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 494                                 vhost_disable_notify(&net->dev, vq);
 495                                 continue;
 496                         }
 497                         break;
 498                 }
 499                 if (in) {
 500                         vq_err(vq, "Unexpected descriptor format for TX: "
 501                                "out %d, int %d\n", out, in);
 502                         break;
 503                 }
 504                 /* Skip header. TODO: support TSO. */
 505                 len = iov_length(vq->iov, out);
 506                 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
 507                 iov_iter_advance(&msg.msg_iter, hdr_size);
 508                 /* Sanity check */
 509                 if (!msg_data_left(&msg)) {
 510                         vq_err(vq, "Unexpected header len for TX: "
 511                                "%zd expected %zd\n",
 512                                len, hdr_size);
 513                         break;
 514                 }
 515                 len = msg_data_left(&msg);
 516
 517                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 518                                    && !vhost_exceeds_maxpend(net)
 519                                    && vhost_net_tx_select_zcopy(net);
 520
 521                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 522                 if (zcopy_used) {
 523                         struct ubuf_info *ubuf;
 524                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 525
 526                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 527                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 528                         ubuf->callback = vhost_zerocopy_callback;
 529                         ubuf->ctx = nvq->ubufs;
 530                         ubuf->desc = nvq->upend_idx;
 531                         refcount_set(&ubuf->refcnt, 1);
 532                         msg.msg_control = ubuf;
 533                         msg.msg_controllen = sizeof(ubuf);
 534                         ubufs = nvq->ubufs;
 535                         atomic_inc(&ubufs->refcount);
 536                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 537                 } else {
 538                         msg.msg_control = NULL;
 539                         ubufs = NULL;
 540                 }
 541
 542                 total_len += len;
 543                 if (total_len < VHOST_NET_WEIGHT &&
 544                     !vhost_vq_avail_empty(&net->dev, vq) &&
 545                     likely(!vhost_exceeds_maxpend(net))) {
 546                         msg.msg_flags |= MSG_MORE;
 547                 } else {
 548                         msg.msg_flags &= ~MSG_MORE;
 549                 }
 550
 551                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 552                 err = sock->ops->sendmsg(sock, &msg, len);
 553                 if (unlikely(err < 0)) {
 554                         if (zcopy_used) {
 555                                 vhost_net_ubuf_put(ubufs);
 556                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 557                                         % UIO_MAXIOV;
 558                         }
 559                         vhost_discard_vq_desc(vq, 1);
 560                         vhost_net_enable_vq(net, vq);
 561                         break;
 562                 }
 563                 if (err != len)
 564                         pr_debug("Truncated TX packet: "
 565                                  " len %d != %zd\n", err, len);
 566                 if (!zcopy_used)
 567                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 568                 else
 569                         vhost_zerocopy_signal_used(net, vq);
 570                 vhost_net_tx_packet(net);
 571                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 572                         vhost_poll_queue(&vq->poll);
 573                         break;
 574                 }
 575         }
 576 out:
 577         mutex_unlock(&vq->mutex);
 578 }
 579
 580 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 581 {
 582         struct sk_buff *head;
 583         int len = 0;
 584         unsigned long flags;
 585
 586         if (rvq->rx_array)
 587                 return vhost_net_buf_peek(rvq);
 588
 589         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 590         head = skb_peek(&sk->sk_receive_queue);
 591         if (likely(head)) {
 592                 len = head->len;
 593                 if (skb_vlan_tag_present(head))
 594                         len += VLAN_HLEN;
 595         }
 596
 597         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 598         return len;
 599 }
 600
 601 static int sk_has_rx_data(struct sock *sk)
 602 {
 603         struct socket *sock = sk->sk_socket;
 604
 605         if (sock->ops->peek_len)
 606                 return sock->ops->peek_len(sock);
 607
 608         return skb_queue_empty(&sk->sk_receive_queue);
 609 }
 610
 611 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
 612 {
 613         struct vhost_net_virtqueue *rvq = &net->vqs[VHOST_NET_VQ_RX];
 614         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 615         struct vhost_virtqueue *vq = &nvq->vq;
 616         unsigned long uninitialized_var(endtime);
 617         int len = peek_head_len(rvq, sk);
 618
 619         if (!len && vq->busyloop_timeout) {
 620                 /* Both tx vq and rx socket were polled here */
 621                 mutex_lock(&vq->mutex);
 622                 vhost_disable_notify(&net->dev, vq);
 623
 624                 preempt_disable();
 625                 endtime = busy_clock() + vq->busyloop_timeout;
 626
 627                 while (vhost_can_busy_poll(&net->dev, endtime) &&
 628                        !sk_has_rx_data(sk) &&
 629                        vhost_vq_avail_empty(&net->dev, vq))
 630                         cpu_relax();
 631
 632                 preempt_enable();
 633
 634                 if (!vhost_vq_avail_empty(&net->dev, vq))
 635                         vhost_poll_queue(&vq->poll);
 636                 else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 637                         vhost_disable_notify(&net->dev, vq);
 638                         vhost_poll_queue(&vq->poll);
 639                 }
 640
 641                 mutex_unlock(&vq->mutex);
 642
 643                 len = peek_head_len(rvq, sk);
 644         }
 645
 646         return len;
 647 }
 648
 649 /* This is a multi-buffer version of vhost_get_desc, that works if
 650  *      vq has read descriptors only.
 651  * @vq          - the relevant virtqueue
 652  * @datalen     - data length we'll be reading
 653  * @iovcount    - returned count of io vectors we fill
 654  * @log         - vhost log
 655  * @log_num     - log offset
 656  * @quota       - headcount quota, 1 for big buffer
 657  *      returns number of buffer heads allocated, negative on error
 658  */
 659 static int get_rx_bufs(struct vhost_virtqueue *vq,
 660                        struct vring_used_elem *heads,
 661                        int datalen,
 662                        unsigned *iovcount,
 663                        struct vhost_log *log,
 664                        unsigned *log_num,
 665                        unsigned int quota)
 666 {
 667         unsigned int out, in;
 668         int seg = 0;
 669         int headcount = 0;
 670         unsigned d;
 671         int r, nlogs = 0;
 672         /* len is always initialized before use since we are always called with
 673          * datalen > 0.
 674          */
 675         u32 uninitialized_var(len);
 676
 677         while (datalen > 0 && headcount < quota) {
 678                 if (unlikely(seg >= UIO_MAXIOV)) {
 679                         r = -ENOBUFS;
 680                         goto err;
 681                 }
 682                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 683                                       ARRAY_SIZE(vq->iov) - seg, &out,
 684                                       &in, log, log_num);
 685                 if (unlikely(r < 0))
 686                         goto err;
 687
 688                 d = r;
 689                 if (d == vq->num) {
 690                         r = 0;
 691                         goto err;
 692                 }
 693                 if (unlikely(out || in <= 0)) {
 694                         vq_err(vq, "unexpected descriptor format for RX: "
 695                                 "out %d, in %d\n", out, in);
 696                         r = -EINVAL;
 697                         goto err;
 698                 }
 699                 if (unlikely(log)) {
 700                         nlogs += *log_num;
 701                         log += *log_num;
 702                 }
 703                 heads[headcount].id = cpu_to_vhost32(vq, d);
 704                 len = iov_length(vq->iov + seg, in);
 705                 heads[headcount].len = cpu_to_vhost32(vq, len);
 706                 datalen -= len;
 707                 ++headcount;
 708                 seg += in;
 709         }
 710         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 711         *iovcount = seg;
 712         if (unlikely(log))
 713                 *log_num = nlogs;
 714
 715         /* Detect overrun */
 716         if (unlikely(datalen > 0)) {
 717                 r = UIO_MAXIOV + 1;
 718                 goto err;
 719         }
 720         return headcount;
 721 err:
 722         vhost_discard_vq_desc(vq, headcount);
 723         return r;
 724 }
 725
 726 /* Expects to be always run from workqueue - which acts as
 727  * read-size critical section for our kind of RCU. */
 728 static void handle_rx(struct vhost_net *net)
 729 {
 730         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 731         struct vhost_virtqueue *vq = &nvq->vq;
 732         unsigned uninitialized_var(in), log;
 733         struct vhost_log *vq_log;
 734         struct msghdr msg = {
 735                 .msg_name = NULL,
 736                 .msg_namelen = 0,
 737                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 738                 .msg_controllen = 0,
 739                 .msg_flags = MSG_DONTWAIT,
 740         };
 741         struct virtio_net_hdr hdr = {
 742                 .flags = 0,
 743                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 744         };
 745         size_t total_len = 0;
 746         int err, mergeable;
 747         s16 headcount;
 748         size_t vhost_hlen, sock_hlen;
 749         size_t vhost_len, sock_len;
 750         struct socket *sock;
 751         struct iov_iter fixup;
 752         __virtio16 num_buffers;
 753
 754         mutex_lock(&vq->mutex);
 755         sock = vq->private_data;
 756         if (!sock)
 757                 goto out;
 758
 759         if (!vq_iotlb_prefetch(vq))
 760                 goto out;
 761
 762         vhost_disable_notify(&net->dev, vq);
 763         vhost_net_disable_vq(net, vq);
 764
 765         vhost_hlen = nvq->vhost_hlen;
 766         sock_hlen = nvq->sock_hlen;
 767
 768         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 769                 vq->log : NULL;
 770         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 771
 772         while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
 773                 sock_len += sock_hlen;
 774                 vhost_len = sock_len + vhost_hlen;
 775                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 776                                         &in, vq_log, &log,
 777                                         likely(mergeable) ? UIO_MAXIOV : 1);
 778                 /* On error, stop handling until the next kick. */
 779                 if (unlikely(headcount < 0))
 780                         goto out;
 781                 /* OK, now we need to know about added descriptors. */
 782                 if (!headcount) {
 783                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 784                                 /* They have slipped one in as we were
 785                                  * doing that: check again. */
 786                                 vhost_disable_notify(&net->dev, vq);
 787                                 continue;
 788                         }
 789                         /* Nothing new?  Wait for eventfd to tell us
 790                          * they refilled. */
 791                         goto out;
 792                 }
 793                 if (nvq->rx_array)
 794                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
 795                 /* On overrun, truncate and discard */
 796                 if (unlikely(headcount > UIO_MAXIOV)) {
 797                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 798                         err = sock->ops->recvmsg(sock, &msg,
 799                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 800                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 801                         continue;
 802                 }
 803                 /* We don't need to be notified again. */
 804                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
 805                 fixup = msg.msg_iter;
 806                 if (unlikely((vhost_hlen))) {
 807                         /* We will supply the header ourselves
 808                          * TODO: support TSO.
 809                          */
 810                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
 811                 }
 812                 err = sock->ops->recvmsg(sock, &msg,
 813                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 814                 /* Userspace might have consumed the packet meanwhile:
 815                  * it's not supposed to do this usually, but might be hard
 816                  * to prevent. Discard data we got (if any) and keep going. */
 817                 if (unlikely(err != sock_len)) {
 818                         pr_debug("Discarded rx packet: "
 819                                  " len %d, expected %zd\n", err, sock_len);
 820                         vhost_discard_vq_desc(vq, headcount);
 821                         continue;
 822                 }
 823                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
 824                 if (unlikely(vhost_hlen)) {
 825                         if (copy_to_iter(&hdr, sizeof(hdr),
 826                                          &fixup) != sizeof(hdr)) {
 827                                 vq_err(vq, "Unable to write vnet_hdr "
 828                                        "at addr %p\n", vq->iov->iov_base);
 829                                 goto out;
 830                         }
 831                 } else {
 832                         /* Header came from socket; we'll need to patch
 833                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
 834                          */
 835                         iov_iter_advance(&fixup, sizeof(hdr));
 836                 }
 837                 /* TODO: Should check and handle checksum. */
 838
 839                 num_buffers = cpu_to_vhost16(vq, headcount);
 840                 if (likely(mergeable) &&
 841                     copy_to_iter(&num_buffers, sizeof num_buffers,
 842                                  &fixup) != sizeof num_buffers) {
 843                         vq_err(vq, "Failed num_buffers write");
 844                         vhost_discard_vq_desc(vq, headcount);
 845                         goto out;
 846                 }
 847                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 848                                             headcount);
 849                 if (unlikely(vq_log))
 850                         vhost_log_write(vq, vq_log, log, vhost_len);
 851                 total_len += vhost_len;
 852                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 853                         vhost_poll_queue(&vq->poll);
 854                         goto out;
 855                 }
 856         }
 857         vhost_net_enable_vq(net, vq);
 858 out:
 859         mutex_unlock(&vq->mutex);
 860 }
 861
 862 static void handle_tx_kick(struct vhost_work *work)
 863 {
 864         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 865                                                   poll.work);
 866         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 867
 868         handle_tx(net);
 869 }
 870
 871 static void handle_rx_kick(struct vhost_work *work)
 872 {
 873         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 874                                                   poll.work);
 875         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 876
 877         handle_rx(net);
 878 }
 879
 880 static void handle_tx_net(struct vhost_work *work)
 881 {
 882         struct vhost_net *net = container_of(work, struct vhost_net,
 883                                              poll[VHOST_NET_VQ_TX].work);
 884         handle_tx(net);
 885 }
 886
 887 static void handle_rx_net(struct vhost_work *work)
 888 {
 889         struct vhost_net *net = container_of(work, struct vhost_net,
 890                                              poll[VHOST_NET_VQ_RX].work);
 891         handle_rx(net);
 892 }
 893
 894 static int vhost_net_open(struct inode *inode, struct file *f)
 895 {
 896         struct vhost_net *n;
 897         struct vhost_dev *dev;
 898         struct vhost_virtqueue **vqs;
 899         struct sk_buff **queue;
 900         int i;
 901
 902         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
 903         if (!n)
 904                 return -ENOMEM;
 905         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 906         if (!vqs) {
 907                 kvfree(n);
 908                 return -ENOMEM;
 909         }
 910
 911         queue = kmalloc_array(VHOST_RX_BATCH, sizeof(struct sk_buff *),
 912                               GFP_KERNEL);
 913         if (!queue) {
 914                 kfree(vqs);
 915                 kvfree(n);
 916                 return -ENOMEM;
 917         }
 918         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
 919
 920         dev = &n->dev;
 921         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 922         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 923         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 924         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 925         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 926                 n->vqs[i].ubufs = NULL;
 927                 n->vqs[i].ubuf_info = NULL;
 928                 n->vqs[i].upend_idx = 0;
 929                 n->vqs[i].done_idx = 0;
 930                 n->vqs[i].vhost_hlen = 0;
 931                 n->vqs[i].sock_hlen = 0;
 932                 vhost_net_buf_init(&n->vqs[i].rxq);
 933         }
 934         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 935
 936         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 937         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 938
 939         f->private_data = n;
 940
 941         return 0;
 942 }
 943
 944 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 945                                         struct vhost_virtqueue *vq)
 946 {
 947         struct socket *sock;
 948         struct vhost_net_virtqueue *nvq =
 949                 container_of(vq, struct vhost_net_virtqueue, vq);
 950
 951         mutex_lock(&vq->mutex);
 952         sock = vq->private_data;
 953         vhost_net_disable_vq(n, vq);
 954         vq->private_data = NULL;
 955         vhost_net_buf_unproduce(nvq);
 956         mutex_unlock(&vq->mutex);
 957         return sock;
 958 }
 959
 960 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 961                            struct socket **rx_sock)
 962 {
 963         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 964         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 965 }
 966
 967 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 968 {
 969         vhost_poll_flush(n->poll + index);
 970         vhost_poll_flush(&n->vqs[index].vq.poll);
 971 }
 972
 973 static void vhost_net_flush(struct vhost_net *n)
 974 {
 975         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 976         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 977         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 978                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 979                 n->tx_flush = true;
 980                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 981                 /* Wait for all lower device DMAs done. */
 982                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 983                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 984                 n->tx_flush = false;
 985                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 986                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 987         }
 988 }
 989
 990 static int vhost_net_release(struct inode *inode, struct file *f)
 991 {
 992         struct vhost_net *n = f->private_data;
 993         struct socket *tx_sock;
 994         struct socket *rx_sock;
 995
 996         vhost_net_stop(n, &tx_sock, &rx_sock);
 997         vhost_net_flush(n);
 998         vhost_dev_stop(&n->dev);
 999         vhost_dev_cleanup(&n->dev, false);
1000         vhost_net_vq_reset(n);
1001         if (tx_sock)
1002                 sockfd_put(tx_sock);
1003         if (rx_sock)
1004                 sockfd_put(rx_sock);
1005         /* Make sure no callbacks are outstanding */
1006         synchronize_rcu_bh();
1007         /* We do an extra flush before freeing memory,
1008          * since jobs can re-queue themselves. */
1009         vhost_net_flush(n);
1010         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1011         kfree(n->dev.vqs);
1012         kvfree(n);
1013         return 0;
1014 }
1015
1016 static struct socket *get_raw_socket(int fd)
1017 {
1018         struct {
1019                 struct sockaddr_ll sa;
1020                 char  buf[MAX_ADDR_LEN];
1021         } uaddr;
1022         int uaddr_len = sizeof uaddr, r;
1023         struct socket *sock = sockfd_lookup(fd, &r);
1024
1025         if (!sock)
1026                 return ERR_PTR(-ENOTSOCK);
1027
1028         /* Parameter checking */
1029         if (sock->sk->sk_type != SOCK_RAW) {
1030                 r = -ESOCKTNOSUPPORT;
1031                 goto err;
1032         }
1033
1034         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
1035                                &uaddr_len, 0);
1036         if (r)
1037                 goto err;
1038
1039         if (uaddr.sa.sll_family != AF_PACKET) {
1040                 r = -EPFNOSUPPORT;
1041                 goto err;
1042         }
1043         return sock;
1044 err:
1045         sockfd_put(sock);
1046         return ERR_PTR(r);
1047 }
1048
1049 static struct skb_array *get_tap_skb_array(int fd)
1050 {
1051         struct skb_array *array;
1052         struct file *file = fget(fd);
1053
1054         if (!file)
1055                 return NULL;
1056         array = tun_get_skb_array(file);
1057         if (!IS_ERR(array))
1058                 goto out;
1059         array = tap_get_skb_array(file);
1060         if (!IS_ERR(array))
1061                 goto out;
1062         array = NULL;
1063 out:
1064         fput(file);
1065         return array;
1066 }
1067
1068 static struct socket *get_tap_socket(int fd)
1069 {
1070         struct file *file = fget(fd);
1071         struct socket *sock;
1072
1073         if (!file)
1074                 return ERR_PTR(-EBADF);
1075         sock = tun_get_socket(file);
1076         if (!IS_ERR(sock))
1077                 return sock;
1078         sock = tap_get_socket(file);
1079         if (IS_ERR(sock))
1080                 fput(file);
1081         return sock;
1082 }
1083
1084 static struct socket *get_socket(int fd)
1085 {
1086         struct socket *sock;
1087
1088         /* special case to disable backend */
1089         if (fd == -1)
1090                 return NULL;
1091         sock = get_raw_socket(fd);
1092         if (!IS_ERR(sock))
1093                 return sock;
1094         sock = get_tap_socket(fd);
1095         if (!IS_ERR(sock))
1096                 return sock;
1097         return ERR_PTR(-ENOTSOCK);
1098 }
1099
1100 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1101 {
1102         struct socket *sock, *oldsock;
1103         struct vhost_virtqueue *vq;
1104         struct vhost_net_virtqueue *nvq;
1105         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1106         int r;
1107
1108         mutex_lock(&n->dev.mutex);
1109         r = vhost_dev_check_owner(&n->dev);
1110         if (r)
1111                 goto err;
1112
1113         if (index >= VHOST_NET_VQ_MAX) {
1114                 r = -ENOBUFS;
1115                 goto err;
1116         }
1117         vq = &n->vqs[index].vq;
1118         nvq = &n->vqs[index];
1119         mutex_lock(&vq->mutex);
1120
1121         /* Verify that ring has been setup correctly. */
1122         if (!vhost_vq_access_ok(vq)) {
1123                 r = -EFAULT;
1124                 goto err_vq;
1125         }
1126         sock = get_socket(fd);
1127         if (IS_ERR(sock)) {
1128                 r = PTR_ERR(sock);
1129                 goto err_vq;
1130         }
1131
1132         /* start polling new socket */
1133         oldsock = vq->private_data;
1134         if (sock != oldsock) {
1135                 ubufs = vhost_net_ubuf_alloc(vq,
1136                                              sock && vhost_sock_zcopy(sock));
1137                 if (IS_ERR(ubufs)) {
1138                         r = PTR_ERR(ubufs);
1139                         goto err_ubufs;
1140                 }
1141
1142                 vhost_net_disable_vq(n, vq);
1143                 vq->private_data = sock;
1144                 vhost_net_buf_unproduce(nvq);
1145                 if (index == VHOST_NET_VQ_RX)
1146                         nvq->rx_array = get_tap_skb_array(fd);
1147                 r = vhost_vq_init_access(vq);
1148                 if (r)
1149                         goto err_used;
1150                 r = vhost_net_enable_vq(n, vq);
1151                 if (r)
1152                         goto err_used;
1153
1154                 oldubufs = nvq->ubufs;
1155                 nvq->ubufs = ubufs;
1156
1157                 n->tx_packets = 0;
1158                 n->tx_zcopy_err = 0;
1159                 n->tx_flush = false;
1160         }
1161
1162         mutex_unlock(&vq->mutex);
1163
1164         if (oldubufs) {
1165                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1166                 mutex_lock(&vq->mutex);
1167                 vhost_zerocopy_signal_used(n, vq);
1168                 mutex_unlock(&vq->mutex);
1169         }
1170
1171         if (oldsock) {
1172                 vhost_net_flush_vq(n, index);
1173                 sockfd_put(oldsock);
1174         }
1175
1176         mutex_unlock(&n->dev.mutex);
1177         return 0;
1178
1179 err_used:
1180         vq->private_data = oldsock;
1181         vhost_net_enable_vq(n, vq);
1182         if (ubufs)
1183                 vhost_net_ubuf_put_wait_and_free(ubufs);
1184 err_ubufs:
1185         sockfd_put(sock);
1186 err_vq:
1187         mutex_unlock(&vq->mutex);
1188 err:
1189         mutex_unlock(&n->dev.mutex);
1190         return r;
1191 }
1192
1193 static long vhost_net_reset_owner(struct vhost_net *n)
1194 {
1195         struct socket *tx_sock = NULL;
1196         struct socket *rx_sock = NULL;
1197         long err;
1198         struct vhost_umem *umem;
1199
1200         mutex_lock(&n->dev.mutex);
1201         err = vhost_dev_check_owner(&n->dev);
1202         if (err)
1203                 goto done;
1204         umem = vhost_dev_reset_owner_prepare();
1205         if (!umem) {
1206                 err = -ENOMEM;
1207                 goto done;
1208         }
1209         vhost_net_stop(n, &tx_sock, &rx_sock);
1210         vhost_net_flush(n);
1211         vhost_dev_reset_owner(&n->dev, umem);
1212         vhost_net_vq_reset(n);
1213 done:
1214         mutex_unlock(&n->dev.mutex);
1215         if (tx_sock)
1216                 sockfd_put(tx_sock);
1217         if (rx_sock)
1218                 sockfd_put(rx_sock);
1219         return err;
1220 }
1221
1222 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1223 {
1224         size_t vhost_hlen, sock_hlen, hdr_len;
1225         int i;
1226
1227         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1228                                (1ULL << VIRTIO_F_VERSION_1))) ?
1229                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1230                         sizeof(struct virtio_net_hdr);
1231         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1232                 /* vhost provides vnet_hdr */
1233                 vhost_hlen = hdr_len;
1234                 sock_hlen = 0;
1235         } else {
1236                 /* socket provides vnet_hdr */
1237                 vhost_hlen = 0;
1238                 sock_hlen = hdr_len;
1239         }
1240         mutex_lock(&n->dev.mutex);
1241         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1242             !vhost_log_access_ok(&n->dev))
1243                 goto out_unlock;
1244
1245         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1246                 if (vhost_init_device_iotlb(&n->dev, true))
1247                         goto out_unlock;
1248         }
1249
1250         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1251                 mutex_lock(&n->vqs[i].vq.mutex);
1252                 n->vqs[i].vq.acked_features = features;
1253                 n->vqs[i].vhost_hlen = vhost_hlen;
1254                 n->vqs[i].sock_hlen = sock_hlen;
1255                 mutex_unlock(&n->vqs[i].vq.mutex);
1256         }
1257         mutex_unlock(&n->dev.mutex);
1258         return 0;
1259
1260 out_unlock:
1261         mutex_unlock(&n->dev.mutex);
1262         return -EFAULT;
1263 }
1264
1265 static long vhost_net_set_owner(struct vhost_net *n)
1266 {
1267         int r;
1268
1269         mutex_lock(&n->dev.mutex);
1270         if (vhost_dev_has_owner(&n->dev)) {
1271                 r = -EBUSY;
1272                 goto out;
1273         }
1274         r = vhost_net_set_ubuf_info(n);
1275         if (r)
1276                 goto out;
1277         r = vhost_dev_set_owner(&n->dev);
1278         if (r)
1279                 vhost_net_clear_ubuf_info(n);
1280         vhost_net_flush(n);
1281 out:
1282         mutex_unlock(&n->dev.mutex);
1283         return r;
1284 }
1285
1286 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1287                             unsigned long arg)
1288 {
1289         struct vhost_net *n = f->private_data;
1290         void __user *argp = (void __user *)arg;
1291         u64 __user *featurep = argp;
1292         struct vhost_vring_file backend;
1293         u64 features;
1294         int r;
1295
1296         switch (ioctl) {
1297         case VHOST_NET_SET_BACKEND:
1298                 if (copy_from_user(&backend, argp, sizeof backend))
1299                         return -EFAULT;
1300                 return vhost_net_set_backend(n, backend.index, backend.fd);
1301         case VHOST_GET_FEATURES:
1302                 features = VHOST_NET_FEATURES;
1303                 if (copy_to_user(featurep, &features, sizeof features))
1304                         return -EFAULT;
1305                 return 0;
1306         case VHOST_SET_FEATURES:
1307                 if (copy_from_user(&features, featurep, sizeof features))
1308                         return -EFAULT;
1309                 if (features & ~VHOST_NET_FEATURES)
1310                         return -EOPNOTSUPP;
1311                 return vhost_net_set_features(n, features);
1312         case VHOST_RESET_OWNER:
1313                 return vhost_net_reset_owner(n);
1314         case VHOST_SET_OWNER:
1315                 return vhost_net_set_owner(n);
1316         default:
1317                 mutex_lock(&n->dev.mutex);
1318                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1319                 if (r == -ENOIOCTLCMD)
1320                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1321                 else
1322                         vhost_net_flush(n);
1323                 mutex_unlock(&n->dev.mutex);
1324                 return r;
1325         }
1326 }
1327
1328 #ifdef CONFIG_COMPAT
1329 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1330                                    unsigned long arg)
1331 {
1332         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1333 }
1334 #endif
1335
1336 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1337 {
1338         struct file *file = iocb->ki_filp;
1339         struct vhost_net *n = file->private_data;
1340         struct vhost_dev *dev = &n->dev;
1341         int noblock = file->f_flags & O_NONBLOCK;
1342
1343         return vhost_chr_read_iter(dev, to, noblock);
1344 }
1345
1346 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1347                                         struct iov_iter *from)
1348 {
1349         struct file *file = iocb->ki_filp;
1350         struct vhost_net *n = file->private_data;
1351         struct vhost_dev *dev = &n->dev;
1352
1353         return vhost_chr_write_iter(dev, from);
1354 }
1355
1356 static unsigned int vhost_net_chr_poll(struct file *file, poll_table *wait)
1357 {
1358         struct vhost_net *n = file->private_data;
1359         struct vhost_dev *dev = &n->dev;
1360
1361         return vhost_chr_poll(file, dev, wait);
1362 }
1363
1364 static const struct file_operations vhost_net_fops = {
1365         .owner          = THIS_MODULE,
1366         .release        = vhost_net_release,
1367         .read_iter      = vhost_net_chr_read_iter,
1368         .write_iter     = vhost_net_chr_write_iter,
1369         .poll           = vhost_net_chr_poll,
1370         .unlocked_ioctl = vhost_net_ioctl,
1371 #ifdef CONFIG_COMPAT
1372         .compat_ioctl   = vhost_net_compat_ioctl,
1373 #endif
1374         .open           = vhost_net_open,
1375         .llseek         = noop_llseek,
1376 };
1377
1378 static struct miscdevice vhost_net_misc = {
1379         .minor = VHOST_NET_MINOR,
1380         .name = "vhost-net",
1381         .fops = &vhost_net_fops,
1382 };
1383
1384 static int vhost_net_init(void)
1385 {
1386         if (experimental_zcopytx)
1387                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1388         return misc_register(&vhost_net_misc);
1389 }
1390 module_init(vhost_net_init);
1391
1392 static void vhost_net_exit(void)
1393 {
1394         misc_deregister(&vhost_net_misc);
1395 }
1396 module_exit(vhost_net_exit);
1397
1398 MODULE_VERSION("0.0.1");
1399 MODULE_LICENSE("GPL v2");
1400 MODULE_AUTHOR("Michael S. Tsirkin");
1401 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1402 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1403 MODULE_ALIAS("devname:vhost-net");