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/rcupdate.h>
  19 #include <linux/file.h>
  20 #include <linux/slab.h>
  21
  22 #include <linux/net.h>
  23 #include <linux/if_packet.h>
  24 #include <linux/if_arp.h>
  25 #include <linux/if_tun.h>
  26 #include <linux/if_macvlan.h>
  27 #include <linux/if_vlan.h>
  28
  29 #include <net/sock.h>
  30
  31 #include "vhost.h"
  32
  33 static int experimental_zcopytx = 1;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  36                                        " 1 -Enable; 0 - Disable");
  37
  38 /* Max number of bytes transferred before requeueing the job.
  39  * Using this limit prevents one virtqueue from starving others. */
  40 #define VHOST_NET_WEIGHT 0x80000
  41
  42 /* MAX number of TX used buffers for outstanding zerocopy */
  43 #define VHOST_MAX_PEND 128
  44 #define VHOST_GOODCOPY_LEN 256
  45
  46 /*
  47  * For transmit, used buffer len is unused; we override it to track buffer
  48  * status internally; used for zerocopy tx only.
  49  */
  50 /* Lower device DMA failed */
  51 #define VHOST_DMA_FAILED_LEN    3
  52 /* Lower device DMA done */
  53 #define VHOST_DMA_DONE_LEN      2
  54 /* Lower device DMA in progress */
  55 #define VHOST_DMA_IN_PROGRESS   1
  56 /* Buffer unused */
  57 #define VHOST_DMA_CLEAR_LEN     0
  58
  59 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
  60
  61 enum {
  62         VHOST_NET_VQ_RX = 0,
  63         VHOST_NET_VQ_TX = 1,
  64         VHOST_NET_VQ_MAX = 2,
  65 };
  66
  67 struct vhost_net {
  68         struct vhost_dev dev;
  69         struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  70         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  71         /* Number of TX recently submitted.
  72          * Protected by tx vq lock. */
  73         unsigned tx_packets;
  74         /* Number of times zerocopy TX recently failed.
  75          * Protected by tx vq lock. */
  76         unsigned tx_zcopy_err;
  77         /* Flush in progress. Protected by tx vq lock. */
  78         bool tx_flush;
  79 };
  80
  81 static void vhost_net_tx_packet(struct vhost_net *net)
  82 {
  83         ++net->tx_packets;
  84         if (net->tx_packets < 1024)
  85                 return;
  86         net->tx_packets = 0;
  87         net->tx_zcopy_err = 0;
  88 }
  89
  90 static void vhost_net_tx_err(struct vhost_net *net)
  91 {
  92         ++net->tx_zcopy_err;
  93 }
  94
  95 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
  96 {
  97         /* TX flush waits for outstanding DMAs to be done.
  98          * Don't start new DMAs.
  99          */
 100         return !net->tx_flush &&
 101                 net->tx_packets / 64 >= net->tx_zcopy_err;
 102 }
 103
 104 static bool vhost_sock_zcopy(struct socket *sock)
 105 {
 106         return unlikely(experimental_zcopytx) &&
 107                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 108 }
 109
 110 /* Pop first len bytes from iovec. Return number of segments used. */
 111 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 112                           size_t len, int iov_count)
 113 {
 114         int seg = 0;
 115         size_t size;
 116
 117         while (len && seg < iov_count) {
 118                 size = min(from->iov_len, len);
 119                 to->iov_base = from->iov_base;
 120                 to->iov_len = size;
 121                 from->iov_len -= size;
 122                 from->iov_base += size;
 123                 len -= size;
 124                 ++from;
 125                 ++to;
 126                 ++seg;
 127         }
 128         return seg;
 129 }
 130 /* Copy iovec entries for len bytes from iovec. */
 131 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 132                            size_t len, int iovcount)
 133 {
 134         int seg = 0;
 135         size_t size;
 136
 137         while (len && seg < iovcount) {
 138                 size = min(from->iov_len, len);
 139                 to->iov_base = from->iov_base;
 140                 to->iov_len = size;
 141                 len -= size;
 142                 ++from;
 143                 ++to;
 144                 ++seg;
 145         }
 146 }
 147
 148 /* In case of DMA done not in order in lower device driver for some reason.
 149  * upend_idx is used to track end of used idx, done_idx is used to track head
 150  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 151  * guest used idx.
 152  */
 153 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 154                                       struct vhost_virtqueue *vq)
 155 {
 156         int i;
 157         int j = 0;
 158
 159         for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 160                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 161                         vhost_net_tx_err(net);
 162                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 163                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 164                         vhost_add_used_and_signal(vq->dev, vq,
 165                                                   vq->heads[i].id, 0);
 166                         ++j;
 167                 } else
 168                         break;
 169         }
 170         if (j)
 171                 vq->done_idx = i;
 172         return j;
 173 }
 174
 175 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 176 {
 177         struct vhost_ubuf_ref *ubufs = ubuf->ctx;
 178         struct vhost_virtqueue *vq = ubufs->vq;
 179         int cnt = atomic_read(&ubufs->kref.refcount);
 180
 181         /*
 182          * Trigger polling thread if guest stopped submitting new buffers:
 183          * in this case, the refcount after decrement will eventually reach 1
 184          * so here it is 2.
 185          * We also trigger polling periodically after each 16 packets
 186          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 187          * less than 10% of times).
 188          */
 189         if (cnt <= 2 || !(cnt % 16))
 190                 vhost_poll_queue(&vq->poll);
 191         /* set len to mark this desc buffers done DMA */
 192         vq->heads[ubuf->desc].len = success ?
 193                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 194         vhost_ubuf_put(ubufs);
 195 }
 196
 197 /* Expects to be always run from workqueue - which acts as
 198  * read-size critical section for our kind of RCU. */
 199 static void handle_tx(struct vhost_net *net)
 200 {
 201         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 202         unsigned out, in, s;
 203         int head;
 204         struct msghdr msg = {
 205                 .msg_name = NULL,
 206                 .msg_namelen = 0,
 207                 .msg_control = NULL,
 208                 .msg_controllen = 0,
 209                 .msg_iov = vq->iov,
 210                 .msg_flags = MSG_DONTWAIT,
 211         };
 212         size_t len, total_len = 0;
 213         int err;
 214         size_t hdr_size;
 215         struct socket *sock;
 216         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 217         bool zcopy, zcopy_used;
 218
 219         /* TODO: check that we are running from vhost_worker? */
 220         sock = rcu_dereference_check(vq->private_data, 1);
 221         if (!sock)
 222                 return;
 223
 224         mutex_lock(&vq->mutex);
 225         vhost_disable_notify(&net->dev, vq);
 226
 227         hdr_size = vq->vhost_hlen;
 228         zcopy = vq->ubufs;
 229
 230         for (;;) {
 231                 /* Release DMAs done buffers first */
 232                 if (zcopy)
 233                         vhost_zerocopy_signal_used(net, vq);
 234
 235                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 236                                          ARRAY_SIZE(vq->iov),
 237                                          &out, &in,
 238                                          NULL, NULL);
 239                 /* On error, stop handling until the next kick. */
 240                 if (unlikely(head < 0))
 241                         break;
 242                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 243                 if (head == vq->num) {
 244                         int num_pends;
 245
 246                         /* If more outstanding DMAs, queue the work.
 247                          * Handle upend_idx wrap around
 248                          */
 249                         num_pends = likely(vq->upend_idx >= vq->done_idx) ?
 250                                     (vq->upend_idx - vq->done_idx) :
 251                                     (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
 252                         if (unlikely(num_pends > VHOST_MAX_PEND))
 253                                 break;
 254                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 255                                 vhost_disable_notify(&net->dev, vq);
 256                                 continue;
 257                         }
 258                         break;
 259                 }
 260                 if (in) {
 261                         vq_err(vq, "Unexpected descriptor format for TX: "
 262                                "out %d, int %d\n", out, in);
 263                         break;
 264                 }
 265                 /* Skip header. TODO: support TSO. */
 266                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 267                 msg.msg_iovlen = out;
 268                 len = iov_length(vq->iov, out);
 269                 /* Sanity check */
 270                 if (!len) {
 271                         vq_err(vq, "Unexpected header len for TX: "
 272                                "%zd expected %zd\n",
 273                                iov_length(vq->hdr, s), hdr_size);
 274                         break;
 275                 }
 276                 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
 277                                        vq->upend_idx != vq->done_idx);
 278
 279                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 280                 if (zcopy_used) {
 281                         vq->heads[vq->upend_idx].id = head;
 282                         if (!vhost_net_tx_select_zcopy(net) ||
 283                             len < VHOST_GOODCOPY_LEN) {
 284                                 /* copy don't need to wait for DMA done */
 285                                 vq->heads[vq->upend_idx].len =
 286                                                         VHOST_DMA_DONE_LEN;
 287                                 msg.msg_control = NULL;
 288                                 msg.msg_controllen = 0;
 289                                 ubufs = NULL;
 290                         } else {
 291                                 struct ubuf_info *ubuf;
 292                                 ubuf = vq->ubuf_info + vq->upend_idx;
 293
 294                                 vq->heads[vq->upend_idx].len =
 295                                         VHOST_DMA_IN_PROGRESS;
 296                                 ubuf->callback = vhost_zerocopy_callback;
 297                                 ubuf->ctx = vq->ubufs;
 298                                 ubuf->desc = vq->upend_idx;
 299                                 msg.msg_control = ubuf;
 300                                 msg.msg_controllen = sizeof(ubuf);
 301                                 ubufs = vq->ubufs;
 302                                 kref_get(&ubufs->kref);
 303                         }
 304                         vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
 305                 }
 306                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 307                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 308                 if (unlikely(err < 0)) {
 309                         if (zcopy_used) {
 310                                 if (ubufs)
 311                                         vhost_ubuf_put(ubufs);
 312                                 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
 313                                         UIO_MAXIOV;
 314                         }
 315                         vhost_discard_vq_desc(vq, 1);
 316                         break;
 317                 }
 318                 if (err != len)
 319                         pr_debug("Truncated TX packet: "
 320                                  " len %d != %zd\n", err, len);
 321                 if (!zcopy_used)
 322                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 323                 else
 324                         vhost_zerocopy_signal_used(net, vq);
 325                 total_len += len;
 326                 vhost_net_tx_packet(net);
 327                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 328                         vhost_poll_queue(&vq->poll);
 329                         break;
 330                 }
 331         }
 332
 333         mutex_unlock(&vq->mutex);
 334 }
 335
 336 static int peek_head_len(struct sock *sk)
 337 {
 338         struct sk_buff *head;
 339         int len = 0;
 340         unsigned long flags;
 341
 342         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 343         head = skb_peek(&sk->sk_receive_queue);
 344         if (likely(head)) {
 345                 len = head->len;
 346                 if (vlan_tx_tag_present(head))
 347                         len += VLAN_HLEN;
 348         }
 349
 350         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 351         return len;
 352 }
 353
 354 /* This is a multi-buffer version of vhost_get_desc, that works if
 355  *      vq has read descriptors only.
 356  * @vq          - the relevant virtqueue
 357  * @datalen     - data length we'll be reading
 358  * @iovcount    - returned count of io vectors we fill
 359  * @log         - vhost log
 360  * @log_num     - log offset
 361  * @quota       - headcount quota, 1 for big buffer
 362  *      returns number of buffer heads allocated, negative on error
 363  */
 364 static int get_rx_bufs(struct vhost_virtqueue *vq,
 365                        struct vring_used_elem *heads,
 366                        int datalen,
 367                        unsigned *iovcount,
 368                        struct vhost_log *log,
 369                        unsigned *log_num,
 370                        unsigned int quota)
 371 {
 372         unsigned int out, in;
 373         int seg = 0;
 374         int headcount = 0;
 375         unsigned d;
 376         int r, nlogs = 0;
 377
 378         while (datalen > 0 && headcount < quota) {
 379                 if (unlikely(seg >= UIO_MAXIOV)) {
 380                         r = -ENOBUFS;
 381                         goto err;
 382                 }
 383                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 384                                       ARRAY_SIZE(vq->iov) - seg, &out,
 385                                       &in, log, log_num);
 386                 if (d == vq->num) {
 387                         r = 0;
 388                         goto err;
 389                 }
 390                 if (unlikely(out || in <= 0)) {
 391                         vq_err(vq, "unexpected descriptor format for RX: "
 392                                 "out %d, in %d\n", out, in);
 393                         r = -EINVAL;
 394                         goto err;
 395                 }
 396                 if (unlikely(log)) {
 397                         nlogs += *log_num;
 398                         log += *log_num;
 399                 }
 400                 heads[headcount].id = d;
 401                 heads[headcount].len = iov_length(vq->iov + seg, in);
 402                 datalen -= heads[headcount].len;
 403                 ++headcount;
 404                 seg += in;
 405         }
 406         heads[headcount - 1].len += datalen;
 407         *iovcount = seg;
 408         if (unlikely(log))
 409                 *log_num = nlogs;
 410         return headcount;
 411 err:
 412         vhost_discard_vq_desc(vq, headcount);
 413         return r;
 414 }
 415
 416 /* Expects to be always run from workqueue - which acts as
 417  * read-size critical section for our kind of RCU. */
 418 static void handle_rx(struct vhost_net *net)
 419 {
 420         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 421         unsigned uninitialized_var(in), log;
 422         struct vhost_log *vq_log;
 423         struct msghdr msg = {
 424                 .msg_name = NULL,
 425                 .msg_namelen = 0,
 426                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 427                 .msg_controllen = 0,
 428                 .msg_iov = vq->iov,
 429                 .msg_flags = MSG_DONTWAIT,
 430         };
 431         struct virtio_net_hdr_mrg_rxbuf hdr = {
 432                 .hdr.flags = 0,
 433                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 434         };
 435         size_t total_len = 0;
 436         int err, mergeable;
 437         s16 headcount;
 438         size_t vhost_hlen, sock_hlen;
 439         size_t vhost_len, sock_len;
 440         /* TODO: check that we are running from vhost_worker? */
 441         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 442
 443         if (!sock)
 444                 return;
 445
 446         mutex_lock(&vq->mutex);
 447         vhost_disable_notify(&net->dev, vq);
 448         vhost_hlen = vq->vhost_hlen;
 449         sock_hlen = vq->sock_hlen;
 450
 451         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 452                 vq->log : NULL;
 453         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 454
 455         while ((sock_len = peek_head_len(sock->sk))) {
 456                 sock_len += sock_hlen;
 457                 vhost_len = sock_len + vhost_hlen;
 458                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 459                                         &in, vq_log, &log,
 460                                         likely(mergeable) ? UIO_MAXIOV : 1);
 461                 /* On error, stop handling until the next kick. */
 462                 if (unlikely(headcount < 0))
 463                         break;
 464                 /* OK, now we need to know about added descriptors. */
 465                 if (!headcount) {
 466                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 467                                 /* They have slipped one in as we were
 468                                  * doing that: check again. */
 469                                 vhost_disable_notify(&net->dev, vq);
 470                                 continue;
 471                         }
 472                         /* Nothing new?  Wait for eventfd to tell us
 473                          * they refilled. */
 474                         break;
 475                 }
 476                 /* We don't need to be notified again. */
 477                 if (unlikely((vhost_hlen)))
 478                         /* Skip header. TODO: support TSO. */
 479                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 480                 else
 481                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 482                          * needed because recvmsg can modify msg_iov. */
 483                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 484                 msg.msg_iovlen = in;
 485                 err = sock->ops->recvmsg(NULL, sock, &msg,
 486                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 487                 /* Userspace might have consumed the packet meanwhile:
 488                  * it's not supposed to do this usually, but might be hard
 489                  * to prevent. Discard data we got (if any) and keep going. */
 490                 if (unlikely(err != sock_len)) {
 491                         pr_debug("Discarded rx packet: "
 492                                  " len %d, expected %zd\n", err, sock_len);
 493                         vhost_discard_vq_desc(vq, headcount);
 494                         continue;
 495                 }
 496                 if (unlikely(vhost_hlen) &&
 497                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 498                                       vhost_hlen)) {
 499                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 500                                vq->iov->iov_base);
 501                         break;
 502                 }
 503                 /* TODO: Should check and handle checksum. */
 504                 if (likely(mergeable) &&
 505                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 506                                       offsetof(typeof(hdr), num_buffers),
 507                                       sizeof hdr.num_buffers)) {
 508                         vq_err(vq, "Failed num_buffers write");
 509                         vhost_discard_vq_desc(vq, headcount);
 510                         break;
 511                 }
 512                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 513                                             headcount);
 514                 if (unlikely(vq_log))
 515                         vhost_log_write(vq, vq_log, log, vhost_len);
 516                 total_len += vhost_len;
 517                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 518                         vhost_poll_queue(&vq->poll);
 519                         break;
 520                 }
 521         }
 522
 523         mutex_unlock(&vq->mutex);
 524 }
 525
 526 static void handle_tx_kick(struct vhost_work *work)
 527 {
 528         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 529                                                   poll.work);
 530         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 531
 532         handle_tx(net);
 533 }
 534
 535 static void handle_rx_kick(struct vhost_work *work)
 536 {
 537         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 538                                                   poll.work);
 539         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 540
 541         handle_rx(net);
 542 }
 543
 544 static void handle_tx_net(struct vhost_work *work)
 545 {
 546         struct vhost_net *net = container_of(work, struct vhost_net,
 547                                              poll[VHOST_NET_VQ_TX].work);
 548         handle_tx(net);
 549 }
 550
 551 static void handle_rx_net(struct vhost_work *work)
 552 {
 553         struct vhost_net *net = container_of(work, struct vhost_net,
 554                                              poll[VHOST_NET_VQ_RX].work);
 555         handle_rx(net);
 556 }
 557
 558 static int vhost_net_open(struct inode *inode, struct file *f)
 559 {
 560         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 561         struct vhost_dev *dev;
 562         int r;
 563
 564         if (!n)
 565                 return -ENOMEM;
 566
 567         dev = &n->dev;
 568         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 569         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 570         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 571         if (r < 0) {
 572                 kfree(n);
 573                 return r;
 574         }
 575
 576         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 577         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 578
 579         f->private_data = n;
 580
 581         return 0;
 582 }
 583
 584 static void vhost_net_disable_vq(struct vhost_net *n,
 585                                  struct vhost_virtqueue *vq)
 586 {
 587         struct vhost_poll *poll = n->poll + (vq - n->vqs);
 588         if (!vq->private_data)
 589                 return;
 590         vhost_poll_stop(poll);
 591 }
 592
 593 static int vhost_net_enable_vq(struct vhost_net *n,
 594                                 struct vhost_virtqueue *vq)
 595 {
 596         struct vhost_poll *poll = n->poll + (vq - n->vqs);
 597         struct socket *sock;
 598
 599         sock = rcu_dereference_protected(vq->private_data,
 600                                          lockdep_is_held(&vq->mutex));
 601         if (!sock)
 602                 return 0;
 603
 604         return vhost_poll_start(poll, sock->file);
 605 }
 606
 607 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 608                                         struct vhost_virtqueue *vq)
 609 {
 610         struct socket *sock;
 611
 612         mutex_lock(&vq->mutex);
 613         sock = rcu_dereference_protected(vq->private_data,
 614                                          lockdep_is_held(&vq->mutex));
 615         vhost_net_disable_vq(n, vq);
 616         rcu_assign_pointer(vq->private_data, NULL);
 617         mutex_unlock(&vq->mutex);
 618         return sock;
 619 }
 620
 621 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 622                            struct socket **rx_sock)
 623 {
 624         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 625         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 626 }
 627
 628 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 629 {
 630         vhost_poll_flush(n->poll + index);
 631         vhost_poll_flush(&n->dev.vqs[index].poll);
 632 }
 633
 634 static void vhost_net_flush(struct vhost_net *n)
 635 {
 636         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 637         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 638         if (n->dev.vqs[VHOST_NET_VQ_TX].ubufs) {
 639                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 640                 n->tx_flush = true;
 641                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 642                 /* Wait for all lower device DMAs done. */
 643                 vhost_ubuf_put_and_wait(n->dev.vqs[VHOST_NET_VQ_TX].ubufs);
 644                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 645                 n->tx_flush = false;
 646                 kref_init(&n->dev.vqs[VHOST_NET_VQ_TX].ubufs->kref);
 647                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 648         }
 649 }
 650
 651 static int vhost_net_release(struct inode *inode, struct file *f)
 652 {
 653         struct vhost_net *n = f->private_data;
 654         struct socket *tx_sock;
 655         struct socket *rx_sock;
 656
 657         vhost_net_stop(n, &tx_sock, &rx_sock);
 658         vhost_net_flush(n);
 659         vhost_dev_stop(&n->dev);
 660         vhost_dev_cleanup(&n->dev, false);
 661         if (tx_sock)
 662                 fput(tx_sock->file);
 663         if (rx_sock)
 664                 fput(rx_sock->file);
 665         /* We do an extra flush before freeing memory,
 666          * since jobs can re-queue themselves. */
 667         vhost_net_flush(n);
 668         kfree(n);
 669         return 0;
 670 }
 671
 672 static struct socket *get_raw_socket(int fd)
 673 {
 674         struct {
 675                 struct sockaddr_ll sa;
 676                 char  buf[MAX_ADDR_LEN];
 677         } uaddr;
 678         int uaddr_len = sizeof uaddr, r;
 679         struct socket *sock = sockfd_lookup(fd, &r);
 680
 681         if (!sock)
 682                 return ERR_PTR(-ENOTSOCK);
 683
 684         /* Parameter checking */
 685         if (sock->sk->sk_type != SOCK_RAW) {
 686                 r = -ESOCKTNOSUPPORT;
 687                 goto err;
 688         }
 689
 690         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 691                                &uaddr_len, 0);
 692         if (r)
 693                 goto err;
 694
 695         if (uaddr.sa.sll_family != AF_PACKET) {
 696                 r = -EPFNOSUPPORT;
 697                 goto err;
 698         }
 699         return sock;
 700 err:
 701         fput(sock->file);
 702         return ERR_PTR(r);
 703 }
 704
 705 static struct socket *get_tap_socket(int fd)
 706 {
 707         struct file *file = fget(fd);
 708         struct socket *sock;
 709
 710         if (!file)
 711                 return ERR_PTR(-EBADF);
 712         sock = tun_get_socket(file);
 713         if (!IS_ERR(sock))
 714                 return sock;
 715         sock = macvtap_get_socket(file);
 716         if (IS_ERR(sock))
 717                 fput(file);
 718         return sock;
 719 }
 720
 721 static struct socket *get_socket(int fd)
 722 {
 723         struct socket *sock;
 724
 725         /* special case to disable backend */
 726         if (fd == -1)
 727                 return NULL;
 728         sock = get_raw_socket(fd);
 729         if (!IS_ERR(sock))
 730                 return sock;
 731         sock = get_tap_socket(fd);
 732         if (!IS_ERR(sock))
 733                 return sock;
 734         return ERR_PTR(-ENOTSOCK);
 735 }
 736
 737 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 738 {
 739         struct socket *sock, *oldsock;
 740         struct vhost_virtqueue *vq;
 741         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 742         int r;
 743
 744         mutex_lock(&n->dev.mutex);
 745         r = vhost_dev_check_owner(&n->dev);
 746         if (r)
 747                 goto err;
 748
 749         if (index >= VHOST_NET_VQ_MAX) {
 750                 r = -ENOBUFS;
 751                 goto err;
 752         }
 753         vq = n->vqs + index;
 754         mutex_lock(&vq->mutex);
 755
 756         /* Verify that ring has been setup correctly. */
 757         if (!vhost_vq_access_ok(vq)) {
 758                 r = -EFAULT;
 759                 goto err_vq;
 760         }
 761         sock = get_socket(fd);
 762         if (IS_ERR(sock)) {
 763                 r = PTR_ERR(sock);
 764                 goto err_vq;
 765         }
 766
 767         /* start polling new socket */
 768         oldsock = rcu_dereference_protected(vq->private_data,
 769                                             lockdep_is_held(&vq->mutex));
 770         if (sock != oldsock) {
 771                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 772                 if (IS_ERR(ubufs)) {
 773                         r = PTR_ERR(ubufs);
 774                         goto err_ubufs;
 775                 }
 776
 777                 vhost_net_disable_vq(n, vq);
 778                 rcu_assign_pointer(vq->private_data, sock);
 779                 r = vhost_init_used(vq);
 780                 if (r)
 781                         goto err_used;
 782                 r = vhost_net_enable_vq(n, vq);
 783                 if (r)
 784                         goto err_used;
 785
 786                 oldubufs = vq->ubufs;
 787                 vq->ubufs = ubufs;
 788
 789                 n->tx_packets = 0;
 790                 n->tx_zcopy_err = 0;
 791                 n->tx_flush = false;
 792         }
 793
 794         mutex_unlock(&vq->mutex);
 795
 796         if (oldubufs) {
 797                 vhost_ubuf_put_and_wait(oldubufs);
 798                 mutex_lock(&vq->mutex);
 799                 vhost_zerocopy_signal_used(n, vq);
 800                 mutex_unlock(&vq->mutex);
 801         }
 802
 803         if (oldsock) {
 804                 vhost_net_flush_vq(n, index);
 805                 fput(oldsock->file);
 806         }
 807
 808         mutex_unlock(&n->dev.mutex);
 809         return 0;
 810
 811 err_used:
 812         rcu_assign_pointer(vq->private_data, oldsock);
 813         vhost_net_enable_vq(n, vq);
 814         if (ubufs)
 815                 vhost_ubuf_put_and_wait(ubufs);
 816 err_ubufs:
 817         fput(sock->file);
 818 err_vq:
 819         mutex_unlock(&vq->mutex);
 820 err:
 821         mutex_unlock(&n->dev.mutex);
 822         return r;
 823 }
 824
 825 static long vhost_net_reset_owner(struct vhost_net *n)
 826 {
 827         struct socket *tx_sock = NULL;
 828         struct socket *rx_sock = NULL;
 829         long err;
 830
 831         mutex_lock(&n->dev.mutex);
 832         err = vhost_dev_check_owner(&n->dev);
 833         if (err)
 834                 goto done;
 835         vhost_net_stop(n, &tx_sock, &rx_sock);
 836         vhost_net_flush(n);
 837         err = vhost_dev_reset_owner(&n->dev);
 838 done:
 839         mutex_unlock(&n->dev.mutex);
 840         if (tx_sock)
 841                 fput(tx_sock->file);
 842         if (rx_sock)
 843                 fput(rx_sock->file);
 844         return err;
 845 }
 846
 847 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 848 {
 849         size_t vhost_hlen, sock_hlen, hdr_len;
 850         int i;
 851
 852         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 853                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 854                         sizeof(struct virtio_net_hdr);
 855         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 856                 /* vhost provides vnet_hdr */
 857                 vhost_hlen = hdr_len;
 858                 sock_hlen = 0;
 859         } else {
 860                 /* socket provides vnet_hdr */
 861                 vhost_hlen = 0;
 862                 sock_hlen = hdr_len;
 863         }
 864         mutex_lock(&n->dev.mutex);
 865         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 866             !vhost_log_access_ok(&n->dev)) {
 867                 mutex_unlock(&n->dev.mutex);
 868                 return -EFAULT;
 869         }
 870         n->dev.acked_features = features;
 871         smp_wmb();
 872         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 873                 mutex_lock(&n->vqs[i].mutex);
 874                 n->vqs[i].vhost_hlen = vhost_hlen;
 875                 n->vqs[i].sock_hlen = sock_hlen;
 876                 mutex_unlock(&n->vqs[i].mutex);
 877         }
 878         vhost_net_flush(n);
 879         mutex_unlock(&n->dev.mutex);
 880         return 0;
 881 }
 882
 883 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 884                             unsigned long arg)
 885 {
 886         struct vhost_net *n = f->private_data;
 887         void __user *argp = (void __user *)arg;
 888         u64 __user *featurep = argp;
 889         struct vhost_vring_file backend;
 890         u64 features;
 891         int r;
 892
 893         switch (ioctl) {
 894         case VHOST_NET_SET_BACKEND:
 895                 if (copy_from_user(&backend, argp, sizeof backend))
 896                         return -EFAULT;
 897                 return vhost_net_set_backend(n, backend.index, backend.fd);
 898         case VHOST_GET_FEATURES:
 899                 features = VHOST_NET_FEATURES;
 900                 if (copy_to_user(featurep, &features, sizeof features))
 901                         return -EFAULT;
 902                 return 0;
 903         case VHOST_SET_FEATURES:
 904                 if (copy_from_user(&features, featurep, sizeof features))
 905                         return -EFAULT;
 906                 if (features & ~VHOST_NET_FEATURES)
 907                         return -EOPNOTSUPP;
 908                 return vhost_net_set_features(n, features);
 909         case VHOST_RESET_OWNER:
 910                 return vhost_net_reset_owner(n);
 911         default:
 912                 mutex_lock(&n->dev.mutex);
 913                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
 914                 if (r == -ENOIOCTLCMD)
 915                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
 916                 else
 917                         vhost_net_flush(n);
 918                 mutex_unlock(&n->dev.mutex);
 919                 return r;
 920         }
 921 }
 922
 923 #ifdef CONFIG_COMPAT
 924 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 925                                    unsigned long arg)
 926 {
 927         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 928 }
 929 #endif
 930
 931 static const struct file_operations vhost_net_fops = {
 932         .owner          = THIS_MODULE,
 933         .release        = vhost_net_release,
 934         .unlocked_ioctl = vhost_net_ioctl,
 935 #ifdef CONFIG_COMPAT
 936         .compat_ioctl   = vhost_net_compat_ioctl,
 937 #endif
 938         .open           = vhost_net_open,
 939         .llseek         = noop_llseek,
 940 };
 941
 942 static struct miscdevice vhost_net_misc = {
 943         .minor = VHOST_NET_MINOR,
 944         .name = "vhost-net",
 945         .fops = &vhost_net_fops,
 946 };
 947
 948 static int vhost_net_init(void)
 949 {
 950         if (experimental_zcopytx)
 951                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
 952         return misc_register(&vhost_net_misc);
 953 }
 954 module_init(vhost_net_init);
 955
 956 static void vhost_net_exit(void)
 957 {
 958         misc_deregister(&vhost_net_misc);
 959 }
 960 module_exit(vhost_net_exit);
 961
 962 MODULE_VERSION("0.0.1");
 963 MODULE_LICENSE("GPL v2");
 964 MODULE_AUTHOR("Michael S. Tsirkin");
 965 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
 966 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
 967 MODULE_ALIAS("devname:vhost-net");