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         void **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 ptr_ring *rx_ring;
 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 = ptr_ring_consume_batched(nvq->rx_ring, 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_ring && !vhost_net_buf_is_empty(rxq)) {
 171                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 172                                    vhost_net_buf_get_size(rxq),
 173                                    tun_ptr_free);
 174                 rxq->head = rxq->tail = 0;
 175         }
 176 }
 177
 178 static int vhost_net_buf_peek_len(void *ptr)
 179 {
 180         if (tun_is_xdp_buff(ptr)) {
 181                 struct xdp_buff *xdp = tun_ptr_to_xdp(ptr);
 182
 183                 return xdp->data_end - xdp->data;
 184         }
 185
 186         return __skb_array_len_with_tag(ptr);
 187 }
 188
 189 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 190 {
 191         struct vhost_net_buf *rxq = &nvq->rxq;
 192
 193         if (!vhost_net_buf_is_empty(rxq))
 194                 goto out;
 195
 196         if (!vhost_net_buf_produce(nvq))
 197                 return 0;
 198
 199 out:
 200         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 201 }
 202
 203 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 204 {
 205         rxq->head = rxq->tail = 0;
 206 }
 207
 208 static void vhost_net_enable_zcopy(int vq)
 209 {
 210         vhost_net_zcopy_mask |= 0x1 << vq;
 211 }
 212
 213 static struct vhost_net_ubuf_ref *
 214 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 215 {
 216         struct vhost_net_ubuf_ref *ubufs;
 217         /* No zero copy backend? Nothing to count. */
 218         if (!zcopy)
 219                 return NULL;
 220         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 221         if (!ubufs)
 222                 return ERR_PTR(-ENOMEM);
 223         atomic_set(&ubufs->refcount, 1);
 224         init_waitqueue_head(&ubufs->wait);
 225         ubufs->vq = vq;
 226         return ubufs;
 227 }
 228
 229 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 230 {
 231         int r = atomic_sub_return(1, &ubufs->refcount);
 232         if (unlikely(!r))
 233                 wake_up(&ubufs->wait);
 234         return r;
 235 }
 236
 237 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 238 {
 239         vhost_net_ubuf_put(ubufs);
 240         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 241 }
 242
 243 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 244 {
 245         vhost_net_ubuf_put_and_wait(ubufs);
 246         kfree(ubufs);
 247 }
 248
 249 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 250 {
 251         int i;
 252
 253         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 254                 kfree(n->vqs[i].ubuf_info);
 255                 n->vqs[i].ubuf_info = NULL;
 256         }
 257 }
 258
 259 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 260 {
 261         bool zcopy;
 262         int i;
 263
 264         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 265                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 266                 if (!zcopy)
 267                         continue;
 268                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 269                                               UIO_MAXIOV, GFP_KERNEL);
 270                 if  (!n->vqs[i].ubuf_info)
 271                         goto err;
 272         }
 273         return 0;
 274
 275 err:
 276         vhost_net_clear_ubuf_info(n);
 277         return -ENOMEM;
 278 }
 279
 280 static void vhost_net_vq_reset(struct vhost_net *n)
 281 {
 282         int i;
 283
 284         vhost_net_clear_ubuf_info(n);
 285
 286         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 287                 n->vqs[i].done_idx = 0;
 288                 n->vqs[i].upend_idx = 0;
 289                 n->vqs[i].ubufs = NULL;
 290                 n->vqs[i].vhost_hlen = 0;
 291                 n->vqs[i].sock_hlen = 0;
 292                 vhost_net_buf_init(&n->vqs[i].rxq);
 293         }
 294
 295 }
 296
 297 static void vhost_net_tx_packet(struct vhost_net *net)
 298 {
 299         ++net->tx_packets;
 300         if (net->tx_packets < 1024)
 301                 return;
 302         net->tx_packets = 0;
 303         net->tx_zcopy_err = 0;
 304 }
 305
 306 static void vhost_net_tx_err(struct vhost_net *net)
 307 {
 308         ++net->tx_zcopy_err;
 309 }
 310
 311 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 312 {
 313         /* TX flush waits for outstanding DMAs to be done.
 314          * Don't start new DMAs.
 315          */
 316         return !net->tx_flush &&
 317                 net->tx_packets / 64 >= net->tx_zcopy_err;
 318 }
 319
 320 static bool vhost_sock_zcopy(struct socket *sock)
 321 {
 322         return unlikely(experimental_zcopytx) &&
 323                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 324 }
 325
 326 /* In case of DMA done not in order in lower device driver for some reason.
 327  * upend_idx is used to track end of used idx, done_idx is used to track head
 328  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 329  * guest used idx.
 330  */
 331 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 332                                        struct vhost_virtqueue *vq)
 333 {
 334         struct vhost_net_virtqueue *nvq =
 335                 container_of(vq, struct vhost_net_virtqueue, vq);
 336         int i, add;
 337         int j = 0;
 338
 339         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 340                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 341                         vhost_net_tx_err(net);
 342                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 343                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 344                         ++j;
 345                 } else
 346                         break;
 347         }
 348         while (j) {
 349                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 350                 vhost_add_used_and_signal_n(vq->dev, vq,
 351                                             &vq->heads[nvq->done_idx], add);
 352                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 353                 j -= add;
 354         }
 355 }
 356
 357 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 358 {
 359         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 360         struct vhost_virtqueue *vq = ubufs->vq;
 361         int cnt;
 362
 363         rcu_read_lock_bh();
 364
 365         /* set len to mark this desc buffers done DMA */
 366         vq->heads[ubuf->desc].len = success ?
 367                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 368         cnt = vhost_net_ubuf_put(ubufs);
 369
 370         /*
 371          * Trigger polling thread if guest stopped submitting new buffers:
 372          * in this case, the refcount after decrement will eventually reach 1.
 373          * We also trigger polling periodically after each 16 packets
 374          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 375          * less than 10% of times).
 376          */
 377         if (cnt <= 1 || !(cnt % 16))
 378                 vhost_poll_queue(&vq->poll);
 379
 380         rcu_read_unlock_bh();
 381 }
 382
 383 static inline unsigned long busy_clock(void)
 384 {
 385         return local_clock() >> 10;
 386 }
 387
 388 static bool vhost_can_busy_poll(struct vhost_dev *dev,
 389                                 unsigned long endtime)
 390 {
 391         return likely(!need_resched()) &&
 392                likely(!time_after(busy_clock(), endtime)) &&
 393                likely(!signal_pending(current)) &&
 394                !vhost_has_work(dev);
 395 }
 396
 397 static void vhost_net_disable_vq(struct vhost_net *n,
 398                                  struct vhost_virtqueue *vq)
 399 {
 400         struct vhost_net_virtqueue *nvq =
 401                 container_of(vq, struct vhost_net_virtqueue, vq);
 402         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 403         if (!vq->private_data)
 404                 return;
 405         vhost_poll_stop(poll);
 406 }
 407
 408 static int vhost_net_enable_vq(struct vhost_net *n,
 409                                 struct vhost_virtqueue *vq)
 410 {
 411         struct vhost_net_virtqueue *nvq =
 412                 container_of(vq, struct vhost_net_virtqueue, vq);
 413         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 414         struct socket *sock;
 415
 416         sock = vq->private_data;
 417         if (!sock)
 418                 return 0;
 419
 420         return vhost_poll_start(poll, sock->file);
 421 }
 422
 423 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 424                                     struct vhost_virtqueue *vq,
 425                                     struct iovec iov[], unsigned int iov_size,
 426                                     unsigned int *out_num, unsigned int *in_num)
 427 {
 428         unsigned long uninitialized_var(endtime);
 429         int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 430                                   out_num, in_num, NULL, NULL);
 431
 432         if (r == vq->num && vq->busyloop_timeout) {
 433                 preempt_disable();
 434                 endtime = busy_clock() + vq->busyloop_timeout;
 435                 while (vhost_can_busy_poll(vq->dev, endtime) &&
 436                        vhost_vq_avail_empty(vq->dev, vq))
 437                         cpu_relax();
 438                 preempt_enable();
 439                 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 440                                       out_num, in_num, NULL, NULL);
 441         }
 442
 443         return r;
 444 }
 445
 446 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 447 {
 448         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 449         struct vhost_virtqueue *vq = &nvq->vq;
 450
 451         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 452                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 453 }
 454
 455 /* Expects to be always run from workqueue - which acts as
 456  * read-size critical section for our kind of RCU. */
 457 static void handle_tx(struct vhost_net *net)
 458 {
 459         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 460         struct vhost_virtqueue *vq = &nvq->vq;
 461         unsigned out, in;
 462         int head;
 463         struct msghdr msg = {
 464                 .msg_name = NULL,
 465                 .msg_namelen = 0,
 466                 .msg_control = NULL,
 467                 .msg_controllen = 0,
 468                 .msg_flags = MSG_DONTWAIT,
 469         };
 470         size_t len, total_len = 0;
 471         int err;
 472         size_t hdr_size;
 473         struct socket *sock;
 474         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 475         bool zcopy, zcopy_used;
 476
 477         mutex_lock(&vq->mutex);
 478         sock = vq->private_data;
 479         if (!sock)
 480                 goto out;
 481
 482         if (!vq_iotlb_prefetch(vq))
 483                 goto out;
 484
 485         vhost_disable_notify(&net->dev, vq);
 486         vhost_net_disable_vq(net, vq);
 487
 488         hdr_size = nvq->vhost_hlen;
 489         zcopy = nvq->ubufs;
 490
 491         for (;;) {
 492                 /* Release DMAs done buffers first */
 493                 if (zcopy)
 494                         vhost_zerocopy_signal_used(net, vq);
 495
 496
 497                 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
 498                                                 ARRAY_SIZE(vq->iov),
 499                                                 &out, &in);
 500                 /* On error, stop handling until the next kick. */
 501                 if (unlikely(head < 0))
 502                         break;
 503                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 504                 if (head == vq->num) {
 505                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 506                                 vhost_disable_notify(&net->dev, vq);
 507                                 continue;
 508                         }
 509                         break;
 510                 }
 511                 if (in) {
 512                         vq_err(vq, "Unexpected descriptor format for TX: "
 513                                "out %d, int %d\n", out, in);
 514                         break;
 515                 }
 516                 /* Skip header. TODO: support TSO. */
 517                 len = iov_length(vq->iov, out);
 518                 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
 519                 iov_iter_advance(&msg.msg_iter, hdr_size);
 520                 /* Sanity check */
 521                 if (!msg_data_left(&msg)) {
 522                         vq_err(vq, "Unexpected header len for TX: "
 523                                "%zd expected %zd\n",
 524                                len, hdr_size);
 525                         break;
 526                 }
 527                 len = msg_data_left(&msg);
 528
 529                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 530                                    && !vhost_exceeds_maxpend(net)
 531                                    && vhost_net_tx_select_zcopy(net);
 532
 533                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 534                 if (zcopy_used) {
 535                         struct ubuf_info *ubuf;
 536                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 537
 538                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 539                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 540                         ubuf->callback = vhost_zerocopy_callback;
 541                         ubuf->ctx = nvq->ubufs;
 542                         ubuf->desc = nvq->upend_idx;
 543                         refcount_set(&ubuf->refcnt, 1);
 544                         msg.msg_control = ubuf;
 545                         msg.msg_controllen = sizeof(ubuf);
 546                         ubufs = nvq->ubufs;
 547                         atomic_inc(&ubufs->refcount);
 548                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 549                 } else {
 550                         msg.msg_control = NULL;
 551                         ubufs = NULL;
 552                 }
 553
 554                 total_len += len;
 555                 if (total_len < VHOST_NET_WEIGHT &&
 556                     !vhost_vq_avail_empty(&net->dev, vq) &&
 557                     likely(!vhost_exceeds_maxpend(net))) {
 558                         msg.msg_flags |= MSG_MORE;
 559                 } else {
 560                         msg.msg_flags &= ~MSG_MORE;
 561                 }
 562
 563                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 564                 err = sock->ops->sendmsg(sock, &msg, len);
 565                 if (unlikely(err < 0)) {
 566                         if (zcopy_used) {
 567                                 vhost_net_ubuf_put(ubufs);
 568                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 569                                         % UIO_MAXIOV;
 570                         }
 571                         vhost_discard_vq_desc(vq, 1);
 572                         vhost_net_enable_vq(net, vq);
 573                         break;
 574                 }
 575                 if (err != len)
 576                         pr_debug("Truncated TX packet: "
 577                                  " len %d != %zd\n", err, len);
 578                 if (!zcopy_used)
 579                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 580                 else
 581                         vhost_zerocopy_signal_used(net, vq);
 582                 vhost_net_tx_packet(net);
 583                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 584                         vhost_poll_queue(&vq->poll);
 585                         break;
 586                 }
 587         }
 588 out:
 589         mutex_unlock(&vq->mutex);
 590 }
 591
 592 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 593 {
 594         struct sk_buff *head;
 595         int len = 0;
 596         unsigned long flags;
 597
 598         if (rvq->rx_ring)
 599                 return vhost_net_buf_peek(rvq);
 600
 601         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 602         head = skb_peek(&sk->sk_receive_queue);
 603         if (likely(head)) {
 604                 len = head->len;
 605                 if (skb_vlan_tag_present(head))
 606                         len += VLAN_HLEN;
 607         }
 608
 609         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 610         return len;
 611 }
 612
 613 static int sk_has_rx_data(struct sock *sk)
 614 {
 615         struct socket *sock = sk->sk_socket;
 616
 617         if (sock->ops->peek_len)
 618                 return sock->ops->peek_len(sock);
 619
 620         return skb_queue_empty(&sk->sk_receive_queue);
 621 }
 622
 623 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
 624 {
 625         struct vhost_net_virtqueue *rvq = &net->vqs[VHOST_NET_VQ_RX];
 626         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 627         struct vhost_virtqueue *vq = &nvq->vq;
 628         unsigned long uninitialized_var(endtime);
 629         int len = peek_head_len(rvq, sk);
 630
 631         if (!len && vq->busyloop_timeout) {
 632                 /* Both tx vq and rx socket were polled here */
 633                 mutex_lock(&vq->mutex);
 634                 vhost_disable_notify(&net->dev, vq);
 635
 636                 preempt_disable();
 637                 endtime = busy_clock() + vq->busyloop_timeout;
 638
 639                 while (vhost_can_busy_poll(&net->dev, endtime) &&
 640                        !sk_has_rx_data(sk) &&
 641                        vhost_vq_avail_empty(&net->dev, vq))
 642                         cpu_relax();
 643
 644                 preempt_enable();
 645
 646                 if (!vhost_vq_avail_empty(&net->dev, vq))
 647                         vhost_poll_queue(&vq->poll);
 648                 else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 649                         vhost_disable_notify(&net->dev, vq);
 650                         vhost_poll_queue(&vq->poll);
 651                 }
 652
 653                 mutex_unlock(&vq->mutex);
 654
 655                 len = peek_head_len(rvq, sk);
 656         }
 657
 658         return len;
 659 }
 660
 661 /* This is a multi-buffer version of vhost_get_desc, that works if
 662  *      vq has read descriptors only.
 663  * @vq          - the relevant virtqueue
 664  * @datalen     - data length we'll be reading
 665  * @iovcount    - returned count of io vectors we fill
 666  * @log         - vhost log
 667  * @log_num     - log offset
 668  * @quota       - headcount quota, 1 for big buffer
 669  *      returns number of buffer heads allocated, negative on error
 670  */
 671 static int get_rx_bufs(struct vhost_virtqueue *vq,
 672                        struct vring_used_elem *heads,
 673                        int datalen,
 674                        unsigned *iovcount,
 675                        struct vhost_log *log,
 676                        unsigned *log_num,
 677                        unsigned int quota)
 678 {
 679         unsigned int out, in;
 680         int seg = 0;
 681         int headcount = 0;
 682         unsigned d;
 683         int r, nlogs = 0;
 684         /* len is always initialized before use since we are always called with
 685          * datalen > 0.
 686          */
 687         u32 uninitialized_var(len);
 688
 689         while (datalen > 0 && headcount < quota) {
 690                 if (unlikely(seg >= UIO_MAXIOV)) {
 691                         r = -ENOBUFS;
 692                         goto err;
 693                 }
 694                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 695                                       ARRAY_SIZE(vq->iov) - seg, &out,
 696                                       &in, log, log_num);
 697                 if (unlikely(r < 0))
 698                         goto err;
 699
 700                 d = r;
 701                 if (d == vq->num) {
 702                         r = 0;
 703                         goto err;
 704                 }
 705                 if (unlikely(out || in <= 0)) {
 706                         vq_err(vq, "unexpected descriptor format for RX: "
 707                                 "out %d, in %d\n", out, in);
 708                         r = -EINVAL;
 709                         goto err;
 710                 }
 711                 if (unlikely(log)) {
 712                         nlogs += *log_num;
 713                         log += *log_num;
 714                 }
 715                 heads[headcount].id = cpu_to_vhost32(vq, d);
 716                 len = iov_length(vq->iov + seg, in);
 717                 heads[headcount].len = cpu_to_vhost32(vq, len);
 718                 datalen -= len;
 719                 ++headcount;
 720                 seg += in;
 721         }
 722         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 723         *iovcount = seg;
 724         if (unlikely(log))
 725                 *log_num = nlogs;
 726
 727         /* Detect overrun */
 728         if (unlikely(datalen > 0)) {
 729                 r = UIO_MAXIOV + 1;
 730                 goto err;
 731         }
 732         return headcount;
 733 err:
 734         vhost_discard_vq_desc(vq, headcount);
 735         return r;
 736 }
 737
 738 /* Expects to be always run from workqueue - which acts as
 739  * read-size critical section for our kind of RCU. */
 740 static void handle_rx(struct vhost_net *net)
 741 {
 742         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 743         struct vhost_virtqueue *vq = &nvq->vq;
 744         unsigned uninitialized_var(in), log;
 745         struct vhost_log *vq_log;
 746         struct msghdr msg = {
 747                 .msg_name = NULL,
 748                 .msg_namelen = 0,
 749                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 750                 .msg_controllen = 0,
 751                 .msg_flags = MSG_DONTWAIT,
 752         };
 753         struct virtio_net_hdr hdr = {
 754                 .flags = 0,
 755                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 756         };
 757         size_t total_len = 0;
 758         int err, mergeable;
 759         s16 headcount, nheads = 0;
 760         size_t vhost_hlen, sock_hlen;
 761         size_t vhost_len, sock_len;
 762         struct socket *sock;
 763         struct iov_iter fixup;
 764         __virtio16 num_buffers;
 765
 766         mutex_lock(&vq->mutex);
 767         sock = vq->private_data;
 768         if (!sock)
 769                 goto out;
 770
 771         if (!vq_iotlb_prefetch(vq))
 772                 goto out;
 773
 774         vhost_disable_notify(&net->dev, vq);
 775         vhost_net_disable_vq(net, vq);
 776
 777         vhost_hlen = nvq->vhost_hlen;
 778         sock_hlen = nvq->sock_hlen;
 779
 780         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 781                 vq->log : NULL;
 782         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 783
 784         while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
 785                 sock_len += sock_hlen;
 786                 vhost_len = sock_len + vhost_hlen;
 787                 headcount = get_rx_bufs(vq, vq->heads + nheads, vhost_len,
 788                                         &in, vq_log, &log,
 789                                         likely(mergeable) ? UIO_MAXIOV : 1);
 790                 /* On error, stop handling until the next kick. */
 791                 if (unlikely(headcount < 0))
 792                         goto out;
 793                 /* OK, now we need to know about added descriptors. */
 794                 if (!headcount) {
 795                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 796                                 /* They have slipped one in as we were
 797                                  * doing that: check again. */
 798                                 vhost_disable_notify(&net->dev, vq);
 799                                 continue;
 800                         }
 801                         /* Nothing new?  Wait for eventfd to tell us
 802                          * they refilled. */
 803                         goto out;
 804                 }
 805                 if (nvq->rx_ring)
 806                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
 807                 /* On overrun, truncate and discard */
 808                 if (unlikely(headcount > UIO_MAXIOV)) {
 809                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 810                         err = sock->ops->recvmsg(sock, &msg,
 811                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 812                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 813                         continue;
 814                 }
 815                 /* We don't need to be notified again. */
 816                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
 817                 fixup = msg.msg_iter;
 818                 if (unlikely((vhost_hlen))) {
 819                         /* We will supply the header ourselves
 820                          * TODO: support TSO.
 821                          */
 822                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
 823                 }
 824                 err = sock->ops->recvmsg(sock, &msg,
 825                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 826                 /* Userspace might have consumed the packet meanwhile:
 827                  * it's not supposed to do this usually, but might be hard
 828                  * to prevent. Discard data we got (if any) and keep going. */
 829                 if (unlikely(err != sock_len)) {
 830                         pr_debug("Discarded rx packet: "
 831                                  " len %d, expected %zd\n", err, sock_len);
 832                         vhost_discard_vq_desc(vq, headcount);
 833                         continue;
 834                 }
 835                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
 836                 if (unlikely(vhost_hlen)) {
 837                         if (copy_to_iter(&hdr, sizeof(hdr),
 838                                          &fixup) != sizeof(hdr)) {
 839                                 vq_err(vq, "Unable to write vnet_hdr "
 840                                        "at addr %p\n", vq->iov->iov_base);
 841                                 goto out;
 842                         }
 843                 } else {
 844                         /* Header came from socket; we'll need to patch
 845                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
 846                          */
 847                         iov_iter_advance(&fixup, sizeof(hdr));
 848                 }
 849                 /* TODO: Should check and handle checksum. */
 850
 851                 num_buffers = cpu_to_vhost16(vq, headcount);
 852                 if (likely(mergeable) &&
 853                     copy_to_iter(&num_buffers, sizeof num_buffers,
 854                                  &fixup) != sizeof num_buffers) {
 855                         vq_err(vq, "Failed num_buffers write");
 856                         vhost_discard_vq_desc(vq, headcount);
 857                         goto out;
 858                 }
 859                 nheads += headcount;
 860                 if (nheads > VHOST_RX_BATCH) {
 861                         vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 862                                                     nheads);
 863                         nheads = 0;
 864                 }
 865                 if (unlikely(vq_log))
 866                         vhost_log_write(vq, vq_log, log, vhost_len);
 867                 total_len += vhost_len;
 868                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 869                         vhost_poll_queue(&vq->poll);
 870                         goto out;
 871                 }
 872         }
 873         vhost_net_enable_vq(net, vq);
 874 out:
 875         if (nheads)
 876                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 877                                             nheads);
 878         mutex_unlock(&vq->mutex);
 879 }
 880
 881 static void handle_tx_kick(struct vhost_work *work)
 882 {
 883         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 884                                                   poll.work);
 885         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 886
 887         handle_tx(net);
 888 }
 889
 890 static void handle_rx_kick(struct vhost_work *work)
 891 {
 892         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 893                                                   poll.work);
 894         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 895
 896         handle_rx(net);
 897 }
 898
 899 static void handle_tx_net(struct vhost_work *work)
 900 {
 901         struct vhost_net *net = container_of(work, struct vhost_net,
 902                                              poll[VHOST_NET_VQ_TX].work);
 903         handle_tx(net);
 904 }
 905
 906 static void handle_rx_net(struct vhost_work *work)
 907 {
 908         struct vhost_net *net = container_of(work, struct vhost_net,
 909                                              poll[VHOST_NET_VQ_RX].work);
 910         handle_rx(net);
 911 }
 912
 913 static int vhost_net_open(struct inode *inode, struct file *f)
 914 {
 915         struct vhost_net *n;
 916         struct vhost_dev *dev;
 917         struct vhost_virtqueue **vqs;
 918         void **queue;
 919         int i;
 920
 921         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
 922         if (!n)
 923                 return -ENOMEM;
 924         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 925         if (!vqs) {
 926                 kvfree(n);
 927                 return -ENOMEM;
 928         }
 929
 930         queue = kmalloc_array(VHOST_RX_BATCH, sizeof(void *),
 931                               GFP_KERNEL);
 932         if (!queue) {
 933                 kfree(vqs);
 934                 kvfree(n);
 935                 return -ENOMEM;
 936         }
 937         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
 938
 939         dev = &n->dev;
 940         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 941         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 942         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 943         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 944         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 945                 n->vqs[i].ubufs = NULL;
 946                 n->vqs[i].ubuf_info = NULL;
 947                 n->vqs[i].upend_idx = 0;
 948                 n->vqs[i].done_idx = 0;
 949                 n->vqs[i].vhost_hlen = 0;
 950                 n->vqs[i].sock_hlen = 0;
 951                 n->vqs[i].rx_ring = NULL;
 952                 vhost_net_buf_init(&n->vqs[i].rxq);
 953         }
 954         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 955
 956         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
 957         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
 958
 959         f->private_data = n;
 960
 961         return 0;
 962 }
 963
 964 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 965                                         struct vhost_virtqueue *vq)
 966 {
 967         struct socket *sock;
 968         struct vhost_net_virtqueue *nvq =
 969                 container_of(vq, struct vhost_net_virtqueue, vq);
 970
 971         mutex_lock(&vq->mutex);
 972         sock = vq->private_data;
 973         vhost_net_disable_vq(n, vq);
 974         vq->private_data = NULL;
 975         vhost_net_buf_unproduce(nvq);
 976         nvq->rx_ring = NULL;
 977         mutex_unlock(&vq->mutex);
 978         return sock;
 979 }
 980
 981 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 982                            struct socket **rx_sock)
 983 {
 984         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 985         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 986 }
 987
 988 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 989 {
 990         vhost_poll_flush(n->poll + index);
 991         vhost_poll_flush(&n->vqs[index].vq.poll);
 992 }
 993
 994 static void vhost_net_flush(struct vhost_net *n)
 995 {
 996         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 997         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 998         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 999                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1000                 n->tx_flush = true;
1001                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1002                 /* Wait for all lower device DMAs done. */
1003                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1004                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1005                 n->tx_flush = false;
1006                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1007                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1008         }
1009 }
1010
1011 static int vhost_net_release(struct inode *inode, struct file *f)
1012 {
1013         struct vhost_net *n = f->private_data;
1014         struct socket *tx_sock;
1015         struct socket *rx_sock;
1016
1017         vhost_net_stop(n, &tx_sock, &rx_sock);
1018         vhost_net_flush(n);
1019         vhost_dev_stop(&n->dev);
1020         vhost_dev_cleanup(&n->dev);
1021         vhost_net_vq_reset(n);
1022         if (tx_sock)
1023                 sockfd_put(tx_sock);
1024         if (rx_sock)
1025                 sockfd_put(rx_sock);
1026         /* Make sure no callbacks are outstanding */
1027         synchronize_rcu_bh();
1028         /* We do an extra flush before freeing memory,
1029          * since jobs can re-queue themselves. */
1030         vhost_net_flush(n);
1031         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1032         kfree(n->dev.vqs);
1033         kvfree(n);
1034         return 0;
1035 }
1036
1037 static struct socket *get_raw_socket(int fd)
1038 {
1039         struct {
1040                 struct sockaddr_ll sa;
1041                 char  buf[MAX_ADDR_LEN];
1042         } uaddr;
1043         int uaddr_len = sizeof uaddr, r;
1044         struct socket *sock = sockfd_lookup(fd, &r);
1045
1046         if (!sock)
1047                 return ERR_PTR(-ENOTSOCK);
1048
1049         /* Parameter checking */
1050         if (sock->sk->sk_type != SOCK_RAW) {
1051                 r = -ESOCKTNOSUPPORT;
1052                 goto err;
1053         }
1054
1055         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
1056                                &uaddr_len, 0);
1057         if (r)
1058                 goto err;
1059
1060         if (uaddr.sa.sll_family != AF_PACKET) {
1061                 r = -EPFNOSUPPORT;
1062                 goto err;
1063         }
1064         return sock;
1065 err:
1066         sockfd_put(sock);
1067         return ERR_PTR(r);
1068 }
1069
1070 static struct ptr_ring *get_tap_ptr_ring(int fd)
1071 {
1072         struct ptr_ring *ring;
1073         struct file *file = fget(fd);
1074
1075         if (!file)
1076                 return NULL;
1077         ring = tun_get_tx_ring(file);
1078         if (!IS_ERR(ring))
1079                 goto out;
1080         ring = tap_get_ptr_ring(file);
1081         if (!IS_ERR(ring))
1082                 goto out;
1083         ring = NULL;
1084 out:
1085         fput(file);
1086         return ring;
1087 }
1088
1089 static struct socket *get_tap_socket(int fd)
1090 {
1091         struct file *file = fget(fd);
1092         struct socket *sock;
1093
1094         if (!file)
1095                 return ERR_PTR(-EBADF);
1096         sock = tun_get_socket(file);
1097         if (!IS_ERR(sock))
1098                 return sock;
1099         sock = tap_get_socket(file);
1100         if (IS_ERR(sock))
1101                 fput(file);
1102         return sock;
1103 }
1104
1105 static struct socket *get_socket(int fd)
1106 {
1107         struct socket *sock;
1108
1109         /* special case to disable backend */
1110         if (fd == -1)
1111                 return NULL;
1112         sock = get_raw_socket(fd);
1113         if (!IS_ERR(sock))
1114                 return sock;
1115         sock = get_tap_socket(fd);
1116         if (!IS_ERR(sock))
1117                 return sock;
1118         return ERR_PTR(-ENOTSOCK);
1119 }
1120
1121 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1122 {
1123         struct socket *sock, *oldsock;
1124         struct vhost_virtqueue *vq;
1125         struct vhost_net_virtqueue *nvq;
1126         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1127         int r;
1128
1129         mutex_lock(&n->dev.mutex);
1130         r = vhost_dev_check_owner(&n->dev);
1131         if (r)
1132                 goto err;
1133
1134         if (index >= VHOST_NET_VQ_MAX) {
1135                 r = -ENOBUFS;
1136                 goto err;
1137         }
1138         vq = &n->vqs[index].vq;
1139         nvq = &n->vqs[index];
1140         mutex_lock(&vq->mutex);
1141
1142         /* Verify that ring has been setup correctly. */
1143         if (!vhost_vq_access_ok(vq)) {
1144                 r = -EFAULT;
1145                 goto err_vq;
1146         }
1147         sock = get_socket(fd);
1148         if (IS_ERR(sock)) {
1149                 r = PTR_ERR(sock);
1150                 goto err_vq;
1151         }
1152
1153         /* start polling new socket */
1154         oldsock = vq->private_data;
1155         if (sock != oldsock) {
1156                 ubufs = vhost_net_ubuf_alloc(vq,
1157                                              sock && vhost_sock_zcopy(sock));
1158                 if (IS_ERR(ubufs)) {
1159                         r = PTR_ERR(ubufs);
1160                         goto err_ubufs;
1161                 }
1162
1163                 vhost_net_disable_vq(n, vq);
1164                 vq->private_data = sock;
1165                 vhost_net_buf_unproduce(nvq);
1166                 r = vhost_vq_init_access(vq);
1167                 if (r)
1168                         goto err_used;
1169                 r = vhost_net_enable_vq(n, vq);
1170                 if (r)
1171                         goto err_used;
1172                 if (index == VHOST_NET_VQ_RX)
1173                         nvq->rx_ring = get_tap_ptr_ring(fd);
1174
1175                 oldubufs = nvq->ubufs;
1176                 nvq->ubufs = ubufs;
1177
1178                 n->tx_packets = 0;
1179                 n->tx_zcopy_err = 0;
1180                 n->tx_flush = false;
1181         }
1182
1183         mutex_unlock(&vq->mutex);
1184
1185         if (oldubufs) {
1186                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1187                 mutex_lock(&vq->mutex);
1188                 vhost_zerocopy_signal_used(n, vq);
1189                 mutex_unlock(&vq->mutex);
1190         }
1191
1192         if (oldsock) {
1193                 vhost_net_flush_vq(n, index);
1194                 sockfd_put(oldsock);
1195         }
1196
1197         mutex_unlock(&n->dev.mutex);
1198         return 0;
1199
1200 err_used:
1201         vq->private_data = oldsock;
1202         vhost_net_enable_vq(n, vq);
1203         if (ubufs)
1204                 vhost_net_ubuf_put_wait_and_free(ubufs);
1205 err_ubufs:
1206         sockfd_put(sock);
1207 err_vq:
1208         mutex_unlock(&vq->mutex);
1209 err:
1210         mutex_unlock(&n->dev.mutex);
1211         return r;
1212 }
1213
1214 static long vhost_net_reset_owner(struct vhost_net *n)
1215 {
1216         struct socket *tx_sock = NULL;
1217         struct socket *rx_sock = NULL;
1218         long err;
1219         struct vhost_umem *umem;
1220
1221         mutex_lock(&n->dev.mutex);
1222         err = vhost_dev_check_owner(&n->dev);
1223         if (err)
1224                 goto done;
1225         umem = vhost_dev_reset_owner_prepare();
1226         if (!umem) {
1227                 err = -ENOMEM;
1228                 goto done;
1229         }
1230         vhost_net_stop(n, &tx_sock, &rx_sock);
1231         vhost_net_flush(n);
1232         vhost_dev_stop(&n->dev);
1233         vhost_dev_reset_owner(&n->dev, umem);
1234         vhost_net_vq_reset(n);
1235 done:
1236         mutex_unlock(&n->dev.mutex);
1237         if (tx_sock)
1238                 sockfd_put(tx_sock);
1239         if (rx_sock)
1240                 sockfd_put(rx_sock);
1241         return err;
1242 }
1243
1244 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1245 {
1246         size_t vhost_hlen, sock_hlen, hdr_len;
1247         int i;
1248
1249         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1250                                (1ULL << VIRTIO_F_VERSION_1))) ?
1251                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1252                         sizeof(struct virtio_net_hdr);
1253         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1254                 /* vhost provides vnet_hdr */
1255                 vhost_hlen = hdr_len;
1256                 sock_hlen = 0;
1257         } else {
1258                 /* socket provides vnet_hdr */
1259                 vhost_hlen = 0;
1260                 sock_hlen = hdr_len;
1261         }
1262         mutex_lock(&n->dev.mutex);
1263         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1264             !vhost_log_access_ok(&n->dev))
1265                 goto out_unlock;
1266
1267         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1268                 if (vhost_init_device_iotlb(&n->dev, true))
1269                         goto out_unlock;
1270         }
1271
1272         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1273                 mutex_lock(&n->vqs[i].vq.mutex);
1274                 n->vqs[i].vq.acked_features = features;
1275                 n->vqs[i].vhost_hlen = vhost_hlen;
1276                 n->vqs[i].sock_hlen = sock_hlen;
1277                 mutex_unlock(&n->vqs[i].vq.mutex);
1278         }
1279         mutex_unlock(&n->dev.mutex);
1280         return 0;
1281
1282 out_unlock:
1283         mutex_unlock(&n->dev.mutex);
1284         return -EFAULT;
1285 }
1286
1287 static long vhost_net_set_owner(struct vhost_net *n)
1288 {
1289         int r;
1290
1291         mutex_lock(&n->dev.mutex);
1292         if (vhost_dev_has_owner(&n->dev)) {
1293                 r = -EBUSY;
1294                 goto out;
1295         }
1296         r = vhost_net_set_ubuf_info(n);
1297         if (r)
1298                 goto out;
1299         r = vhost_dev_set_owner(&n->dev);
1300         if (r)
1301                 vhost_net_clear_ubuf_info(n);
1302         vhost_net_flush(n);
1303 out:
1304         mutex_unlock(&n->dev.mutex);
1305         return r;
1306 }
1307
1308 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1309                             unsigned long arg)
1310 {
1311         struct vhost_net *n = f->private_data;
1312         void __user *argp = (void __user *)arg;
1313         u64 __user *featurep = argp;
1314         struct vhost_vring_file backend;
1315         u64 features;
1316         int r;
1317
1318         switch (ioctl) {
1319         case VHOST_NET_SET_BACKEND:
1320                 if (copy_from_user(&backend, argp, sizeof backend))
1321                         return -EFAULT;
1322                 return vhost_net_set_backend(n, backend.index, backend.fd);
1323         case VHOST_GET_FEATURES:
1324                 features = VHOST_NET_FEATURES;
1325                 if (copy_to_user(featurep, &features, sizeof features))
1326                         return -EFAULT;
1327                 return 0;
1328         case VHOST_SET_FEATURES:
1329                 if (copy_from_user(&features, featurep, sizeof features))
1330                         return -EFAULT;
1331                 if (features & ~VHOST_NET_FEATURES)
1332                         return -EOPNOTSUPP;
1333                 return vhost_net_set_features(n, features);
1334         case VHOST_RESET_OWNER:
1335                 return vhost_net_reset_owner(n);
1336         case VHOST_SET_OWNER:
1337                 return vhost_net_set_owner(n);
1338         default:
1339                 mutex_lock(&n->dev.mutex);
1340                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1341                 if (r == -ENOIOCTLCMD)
1342                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1343                 else
1344                         vhost_net_flush(n);
1345                 mutex_unlock(&n->dev.mutex);
1346                 return r;
1347         }
1348 }
1349
1350 #ifdef CONFIG_COMPAT
1351 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1352                                    unsigned long arg)
1353 {
1354         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1355 }
1356 #endif
1357
1358 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1359 {
1360         struct file *file = iocb->ki_filp;
1361         struct vhost_net *n = file->private_data;
1362         struct vhost_dev *dev = &n->dev;
1363         int noblock = file->f_flags & O_NONBLOCK;
1364
1365         return vhost_chr_read_iter(dev, to, noblock);
1366 }
1367
1368 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1369                                         struct iov_iter *from)
1370 {
1371         struct file *file = iocb->ki_filp;
1372         struct vhost_net *n = file->private_data;
1373         struct vhost_dev *dev = &n->dev;
1374
1375         return vhost_chr_write_iter(dev, from);
1376 }
1377
1378 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1379 {
1380         struct vhost_net *n = file->private_data;
1381         struct vhost_dev *dev = &n->dev;
1382
1383         return vhost_chr_poll(file, dev, wait);
1384 }
1385
1386 static const struct file_operations vhost_net_fops = {
1387         .owner          = THIS_MODULE,
1388         .release        = vhost_net_release,
1389         .read_iter      = vhost_net_chr_read_iter,
1390         .write_iter     = vhost_net_chr_write_iter,
1391         .poll           = vhost_net_chr_poll,
1392         .unlocked_ioctl = vhost_net_ioctl,
1393 #ifdef CONFIG_COMPAT
1394         .compat_ioctl   = vhost_net_compat_ioctl,
1395 #endif
1396         .open           = vhost_net_open,
1397         .llseek         = noop_llseek,
1398 };
1399
1400 static struct miscdevice vhost_net_misc = {
1401         .minor = VHOST_NET_MINOR,
1402         .name = "vhost-net",
1403         .fops = &vhost_net_fops,
1404 };
1405
1406 static int vhost_net_init(void)
1407 {
1408         if (experimental_zcopytx)
1409                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1410         return misc_register(&vhost_net_misc);
1411 }
1412 module_init(vhost_net_init);
1413
1414 static void vhost_net_exit(void)
1415 {
1416         misc_deregister(&vhost_net_misc);
1417 }
1418 module_exit(vhost_net_exit);
1419
1420 MODULE_VERSION("0.0.1");
1421 MODULE_LICENSE("GPL v2");
1422 MODULE_AUTHOR("Michael S. Tsirkin");
1423 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1424 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1425 MODULE_ALIAS("devname:vhost-net");