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;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
  36
  37 /* Max number of bytes transferred before requeueing the job.
  38  * Using this limit prevents one virtqueue from starving others. */
  39 #define VHOST_NET_WEIGHT 0x80000
  40
  41 /* MAX number of TX used buffers for outstanding zerocopy */
  42 #define VHOST_MAX_PEND 128
  43 #define VHOST_GOODCOPY_LEN 256
  44
  45 /*
  46  * For transmit, used buffer len is unused; we override it to track buffer
  47  * status internally; used for zerocopy tx only.
  48  */
  49 /* Lower device DMA failed */
  50 #define VHOST_DMA_FAILED_LEN    3
  51 /* Lower device DMA done */
  52 #define VHOST_DMA_DONE_LEN      2
  53 /* Lower device DMA in progress */
  54 #define VHOST_DMA_IN_PROGRESS   1
  55 /* Buffer unused */
  56 #define VHOST_DMA_CLEAR_LEN     0
  57
  58 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
  59
  60 enum {
  61         VHOST_NET_VQ_RX = 0,
  62         VHOST_NET_VQ_TX = 1,
  63         VHOST_NET_VQ_MAX = 2,
  64 };
  65
  66 enum vhost_net_poll_state {
  67         VHOST_NET_POLL_DISABLED = 0,
  68         VHOST_NET_POLL_STARTED = 1,
  69         VHOST_NET_POLL_STOPPED = 2,
  70 };
  71
  72 struct vhost_net {
  73         struct vhost_dev dev;
  74         struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  75         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  76         /* Tells us whether we are polling a socket for TX.
  77          * We only do this when socket buffer fills up.
  78          * Protected by tx vq lock. */
  79         enum vhost_net_poll_state tx_poll_state;
  80         /* Number of TX recently submitted.
  81          * Protected by tx vq lock. */
  82         unsigned tx_packets;
  83         /* Number of times zerocopy TX recently failed.
  84          * Protected by tx vq lock. */
  85         unsigned tx_zcopy_err;
  86 };
  87
  88 static void vhost_net_tx_packet(struct vhost_net *net)
  89 {
  90         ++net->tx_packets;
  91         if (net->tx_packets < 1024)
  92                 return;
  93         net->tx_packets = 0;
  94         net->tx_zcopy_err = 0;
  95 }
  96
  97 static void vhost_net_tx_err(struct vhost_net *net)
  98 {
  99         ++net->tx_zcopy_err;
 100 }
 101
 102 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 103 {
 104         return net->tx_packets / 64 >= net->tx_zcopy_err;
 105 }
 106
 107 static bool vhost_sock_zcopy(struct socket *sock)
 108 {
 109         return unlikely(experimental_zcopytx) &&
 110                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 111 }
 112
 113 /* Pop first len bytes from iovec. Return number of segments used. */
 114 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 115                           size_t len, int iov_count)
 116 {
 117         int seg = 0;
 118         size_t size;
 119
 120         while (len && seg < iov_count) {
 121                 size = min(from->iov_len, len);
 122                 to->iov_base = from->iov_base;
 123                 to->iov_len = size;
 124                 from->iov_len -= size;
 125                 from->iov_base += size;
 126                 len -= size;
 127                 ++from;
 128                 ++to;
 129                 ++seg;
 130         }
 131         return seg;
 132 }
 133 /* Copy iovec entries for len bytes from iovec. */
 134 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 135                            size_t len, int iovcount)
 136 {
 137         int seg = 0;
 138         size_t size;
 139
 140         while (len && seg < iovcount) {
 141                 size = min(from->iov_len, len);
 142                 to->iov_base = from->iov_base;
 143                 to->iov_len = size;
 144                 len -= size;
 145                 ++from;
 146                 ++to;
 147                 ++seg;
 148         }
 149 }
 150
 151 /* Caller must have TX VQ lock */
 152 static void tx_poll_stop(struct vhost_net *net)
 153 {
 154         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
 155                 return;
 156         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 157         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 158 }
 159
 160 /* Caller must have TX VQ lock */
 161 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 162 {
 163         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 164                 return;
 165         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 166         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 167 }
 168
 169 /* In case of DMA done not in order in lower device driver for some reason.
 170  * upend_idx is used to track end of used idx, done_idx is used to track head
 171  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 172  * guest used idx.
 173  */
 174 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 175                                       struct vhost_virtqueue *vq)
 176 {
 177         int i;
 178         int j = 0;
 179
 180         for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 181                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 182                         vhost_net_tx_err(net);
 183                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 184                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 185                         vhost_add_used_and_signal(vq->dev, vq,
 186                                                   vq->heads[i].id, 0);
 187                         ++j;
 188                 } else
 189                         break;
 190         }
 191         if (j)
 192                 vq->done_idx = i;
 193         return j;
 194 }
 195
 196 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 197 {
 198         struct vhost_ubuf_ref *ubufs = ubuf->ctx;
 199         struct vhost_virtqueue *vq = ubufs->vq;
 200         int cnt = atomic_read(&ubufs->kref.refcount);
 201
 202         /*
 203          * Trigger polling thread if guest stopped submitting new buffers:
 204          * in this case, the refcount after decrement will eventually reach 1
 205          * so here it is 2.
 206          * We also trigger polling periodically after each 16 packets
 207          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 208          * less than 10% of times).
 209          */
 210         if (cnt <= 2 || !(cnt % 16))
 211                 vhost_poll_queue(&vq->poll);
 212         /* set len to mark this desc buffers done DMA */
 213         vq->heads[ubuf->desc].len = success ?
 214                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 215         vhost_ubuf_put(ubufs);
 216 }
 217
 218 /* Expects to be always run from workqueue - which acts as
 219  * read-size critical section for our kind of RCU. */
 220 static void handle_tx(struct vhost_net *net)
 221 {
 222         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 223         unsigned out, in, s;
 224         int head;
 225         struct msghdr msg = {
 226                 .msg_name = NULL,
 227                 .msg_namelen = 0,
 228                 .msg_control = NULL,
 229                 .msg_controllen = 0,
 230                 .msg_iov = vq->iov,
 231                 .msg_flags = MSG_DONTWAIT,
 232         };
 233         size_t len, total_len = 0;
 234         int err, wmem;
 235         size_t hdr_size;
 236         struct socket *sock;
 237         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 238         bool zcopy;
 239
 240         /* TODO: check that we are running from vhost_worker? */
 241         sock = rcu_dereference_check(vq->private_data, 1);
 242         if (!sock)
 243                 return;
 244
 245         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 246         if (wmem >= sock->sk->sk_sndbuf) {
 247                 mutex_lock(&vq->mutex);
 248                 tx_poll_start(net, sock);
 249                 mutex_unlock(&vq->mutex);
 250                 return;
 251         }
 252
 253         mutex_lock(&vq->mutex);
 254         vhost_disable_notify(&net->dev, vq);
 255
 256         if (wmem < sock->sk->sk_sndbuf / 2)
 257                 tx_poll_stop(net);
 258         hdr_size = vq->vhost_hlen;
 259         zcopy = vq->ubufs;
 260
 261         for (;;) {
 262                 /* Release DMAs done buffers first */
 263                 if (zcopy)
 264                         vhost_zerocopy_signal_used(net, vq);
 265
 266                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 267                                          ARRAY_SIZE(vq->iov),
 268                                          &out, &in,
 269                                          NULL, NULL);
 270                 /* On error, stop handling until the next kick. */
 271                 if (unlikely(head < 0))
 272                         break;
 273                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 274                 if (head == vq->num) {
 275                         int num_pends;
 276
 277                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 278                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 279                                 tx_poll_start(net, sock);
 280                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 281                                 break;
 282                         }
 283                         /* If more outstanding DMAs, queue the work.
 284                          * Handle upend_idx wrap around
 285                          */
 286                         num_pends = likely(vq->upend_idx >= vq->done_idx) ?
 287                                     (vq->upend_idx - vq->done_idx) :
 288                                     (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
 289                         if (unlikely(num_pends > VHOST_MAX_PEND)) {
 290                                 tx_poll_start(net, sock);
 291                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 292                                 break;
 293                         }
 294                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 295                                 vhost_disable_notify(&net->dev, vq);
 296                                 continue;
 297                         }
 298                         break;
 299                 }
 300                 if (in) {
 301                         vq_err(vq, "Unexpected descriptor format for TX: "
 302                                "out %d, int %d\n", out, in);
 303                         break;
 304                 }
 305                 /* Skip header. TODO: support TSO. */
 306                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 307                 msg.msg_iovlen = out;
 308                 len = iov_length(vq->iov, out);
 309                 /* Sanity check */
 310                 if (!len) {
 311                         vq_err(vq, "Unexpected header len for TX: "
 312                                "%zd expected %zd\n",
 313                                iov_length(vq->hdr, s), hdr_size);
 314                         break;
 315                 }
 316                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 317                 if (zcopy) {
 318                         vq->heads[vq->upend_idx].id = head;
 319                         if (!vhost_net_tx_select_zcopy(net) ||
 320                             len < VHOST_GOODCOPY_LEN) {
 321                                 /* copy don't need to wait for DMA done */
 322                                 vq->heads[vq->upend_idx].len =
 323                                                         VHOST_DMA_DONE_LEN;
 324                                 msg.msg_control = NULL;
 325                                 msg.msg_controllen = 0;
 326                                 ubufs = NULL;
 327                         } else {
 328                                 struct ubuf_info *ubuf = &vq->ubuf_info[head];
 329
 330                                 vq->heads[vq->upend_idx].len =
 331                                         VHOST_DMA_IN_PROGRESS;
 332                                 ubuf->callback = vhost_zerocopy_callback;
 333                                 ubuf->ctx = vq->ubufs;
 334                                 ubuf->desc = vq->upend_idx;
 335                                 msg.msg_control = ubuf;
 336                                 msg.msg_controllen = sizeof(ubuf);
 337                                 ubufs = vq->ubufs;
 338                                 kref_get(&ubufs->kref);
 339                         }
 340                         vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
 341                 }
 342                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 343                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 344                 if (unlikely(err < 0)) {
 345                         if (zcopy) {
 346                                 if (ubufs)
 347                                         vhost_ubuf_put(ubufs);
 348                                 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
 349                                         UIO_MAXIOV;
 350                         }
 351                         vhost_discard_vq_desc(vq, 1);
 352                         if (err == -EAGAIN || err == -ENOBUFS)
 353                                 tx_poll_start(net, sock);
 354                         break;
 355                 }
 356                 if (err != len)
 357                         pr_debug("Truncated TX packet: "
 358                                  " len %d != %zd\n", err, len);
 359                 if (!zcopy)
 360                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 361                 else
 362                         vhost_zerocopy_signal_used(net, vq);
 363                 total_len += len;
 364                 vhost_net_tx_packet(net);
 365                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 366                         vhost_poll_queue(&vq->poll);
 367                         break;
 368                 }
 369         }
 370
 371         mutex_unlock(&vq->mutex);
 372 }
 373
 374 static int peek_head_len(struct sock *sk)
 375 {
 376         struct sk_buff *head;
 377         int len = 0;
 378         unsigned long flags;
 379
 380         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 381         head = skb_peek(&sk->sk_receive_queue);
 382         if (likely(head)) {
 383                 len = head->len;
 384                 if (vlan_tx_tag_present(head))
 385                         len += VLAN_HLEN;
 386         }
 387
 388         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 389         return len;
 390 }
 391
 392 /* This is a multi-buffer version of vhost_get_desc, that works if
 393  *      vq has read descriptors only.
 394  * @vq          - the relevant virtqueue
 395  * @datalen     - data length we'll be reading
 396  * @iovcount    - returned count of io vectors we fill
 397  * @log         - vhost log
 398  * @log_num     - log offset
 399  * @quota       - headcount quota, 1 for big buffer
 400  *      returns number of buffer heads allocated, negative on error
 401  */
 402 static int get_rx_bufs(struct vhost_virtqueue *vq,
 403                        struct vring_used_elem *heads,
 404                        int datalen,
 405                        unsigned *iovcount,
 406                        struct vhost_log *log,
 407                        unsigned *log_num,
 408                        unsigned int quota)
 409 {
 410         unsigned int out, in;
 411         int seg = 0;
 412         int headcount = 0;
 413         unsigned d;
 414         int r, nlogs = 0;
 415
 416         while (datalen > 0 && headcount < quota) {
 417                 if (unlikely(seg >= UIO_MAXIOV)) {
 418                         r = -ENOBUFS;
 419                         goto err;
 420                 }
 421                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 422                                       ARRAY_SIZE(vq->iov) - seg, &out,
 423                                       &in, log, log_num);
 424                 if (d == vq->num) {
 425                         r = 0;
 426                         goto err;
 427                 }
 428                 if (unlikely(out || in <= 0)) {
 429                         vq_err(vq, "unexpected descriptor format for RX: "
 430                                 "out %d, in %d\n", out, in);
 431                         r = -EINVAL;
 432                         goto err;
 433                 }
 434                 if (unlikely(log)) {
 435                         nlogs += *log_num;
 436                         log += *log_num;
 437                 }
 438                 heads[headcount].id = d;
 439                 heads[headcount].len = iov_length(vq->iov + seg, in);
 440                 datalen -= heads[headcount].len;
 441                 ++headcount;
 442                 seg += in;
 443         }
 444         heads[headcount - 1].len += datalen;
 445         *iovcount = seg;
 446         if (unlikely(log))
 447                 *log_num = nlogs;
 448         return headcount;
 449 err:
 450         vhost_discard_vq_desc(vq, headcount);
 451         return r;
 452 }
 453
 454 /* Expects to be always run from workqueue - which acts as
 455  * read-size critical section for our kind of RCU. */
 456 static void handle_rx(struct vhost_net *net)
 457 {
 458         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 459         unsigned uninitialized_var(in), log;
 460         struct vhost_log *vq_log;
 461         struct msghdr msg = {
 462                 .msg_name = NULL,
 463                 .msg_namelen = 0,
 464                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 465                 .msg_controllen = 0,
 466                 .msg_iov = vq->iov,
 467                 .msg_flags = MSG_DONTWAIT,
 468         };
 469         struct virtio_net_hdr_mrg_rxbuf hdr = {
 470                 .hdr.flags = 0,
 471                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 472         };
 473         size_t total_len = 0;
 474         int err, mergeable;
 475         s16 headcount;
 476         size_t vhost_hlen, sock_hlen;
 477         size_t vhost_len, sock_len;
 478         /* TODO: check that we are running from vhost_worker? */
 479         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 480
 481         if (!sock)
 482                 return;
 483
 484         mutex_lock(&vq->mutex);
 485         vhost_disable_notify(&net->dev, vq);
 486         vhost_hlen = vq->vhost_hlen;
 487         sock_hlen = vq->sock_hlen;
 488
 489         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 490                 vq->log : NULL;
 491         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 492
 493         while ((sock_len = peek_head_len(sock->sk))) {
 494                 sock_len += sock_hlen;
 495                 vhost_len = sock_len + vhost_hlen;
 496                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 497                                         &in, vq_log, &log,
 498                                         likely(mergeable) ? UIO_MAXIOV : 1);
 499                 /* On error, stop handling until the next kick. */
 500                 if (unlikely(headcount < 0))
 501                         break;
 502                 /* OK, now we need to know about added descriptors. */
 503                 if (!headcount) {
 504                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 505                                 /* They have slipped one in as we were
 506                                  * doing that: check again. */
 507                                 vhost_disable_notify(&net->dev, vq);
 508                                 continue;
 509                         }
 510                         /* Nothing new?  Wait for eventfd to tell us
 511                          * they refilled. */
 512                         break;
 513                 }
 514                 /* We don't need to be notified again. */
 515                 if (unlikely((vhost_hlen)))
 516                         /* Skip header. TODO: support TSO. */
 517                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 518                 else
 519                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 520                          * needed because recvmsg can modify msg_iov. */
 521                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 522                 msg.msg_iovlen = in;
 523                 err = sock->ops->recvmsg(NULL, sock, &msg,
 524                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 525                 /* Userspace might have consumed the packet meanwhile:
 526                  * it's not supposed to do this usually, but might be hard
 527                  * to prevent. Discard data we got (if any) and keep going. */
 528                 if (unlikely(err != sock_len)) {
 529                         pr_debug("Discarded rx packet: "
 530                                  " len %d, expected %zd\n", err, sock_len);
 531                         vhost_discard_vq_desc(vq, headcount);
 532                         continue;
 533                 }
 534                 if (unlikely(vhost_hlen) &&
 535                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 536                                       vhost_hlen)) {
 537                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 538                                vq->iov->iov_base);
 539                         break;
 540                 }
 541                 /* TODO: Should check and handle checksum. */
 542                 if (likely(mergeable) &&
 543                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 544                                       offsetof(typeof(hdr), num_buffers),
 545                                       sizeof hdr.num_buffers)) {
 546                         vq_err(vq, "Failed num_buffers write");
 547                         vhost_discard_vq_desc(vq, headcount);
 548                         break;
 549                 }
 550                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 551                                             headcount);
 552                 if (unlikely(vq_log))
 553                         vhost_log_write(vq, vq_log, log, vhost_len);
 554                 total_len += vhost_len;
 555                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 556                         vhost_poll_queue(&vq->poll);
 557                         break;
 558                 }
 559         }
 560
 561         mutex_unlock(&vq->mutex);
 562 }
 563
 564 static void handle_tx_kick(struct vhost_work *work)
 565 {
 566         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 567                                                   poll.work);
 568         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 569
 570         handle_tx(net);
 571 }
 572
 573 static void handle_rx_kick(struct vhost_work *work)
 574 {
 575         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 576                                                   poll.work);
 577         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 578
 579         handle_rx(net);
 580 }
 581
 582 static void handle_tx_net(struct vhost_work *work)
 583 {
 584         struct vhost_net *net = container_of(work, struct vhost_net,
 585                                              poll[VHOST_NET_VQ_TX].work);
 586         handle_tx(net);
 587 }
 588
 589 static void handle_rx_net(struct vhost_work *work)
 590 {
 591         struct vhost_net *net = container_of(work, struct vhost_net,
 592                                              poll[VHOST_NET_VQ_RX].work);
 593         handle_rx(net);
 594 }
 595
 596 static int vhost_net_open(struct inode *inode, struct file *f)
 597 {
 598         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 599         struct vhost_dev *dev;
 600         int r;
 601
 602         if (!n)
 603                 return -ENOMEM;
 604
 605         dev = &n->dev;
 606         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 607         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 608         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 609         if (r < 0) {
 610                 kfree(n);
 611                 return r;
 612         }
 613
 614         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 615         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 616         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 617
 618         f->private_data = n;
 619
 620         return 0;
 621 }
 622
 623 static void vhost_net_disable_vq(struct vhost_net *n,
 624                                  struct vhost_virtqueue *vq)
 625 {
 626         if (!vq->private_data)
 627                 return;
 628         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 629                 tx_poll_stop(n);
 630                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 631         } else
 632                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 633 }
 634
 635 static void vhost_net_enable_vq(struct vhost_net *n,
 636                                 struct vhost_virtqueue *vq)
 637 {
 638         struct socket *sock;
 639
 640         sock = rcu_dereference_protected(vq->private_data,
 641                                          lockdep_is_held(&vq->mutex));
 642         if (!sock)
 643                 return;
 644         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 645                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 646                 tx_poll_start(n, sock);
 647         } else
 648                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 649 }
 650
 651 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 652                                         struct vhost_virtqueue *vq)
 653 {
 654         struct socket *sock;
 655
 656         mutex_lock(&vq->mutex);
 657         sock = rcu_dereference_protected(vq->private_data,
 658                                          lockdep_is_held(&vq->mutex));
 659         vhost_net_disable_vq(n, vq);
 660         rcu_assign_pointer(vq->private_data, NULL);
 661         mutex_unlock(&vq->mutex);
 662         return sock;
 663 }
 664
 665 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 666                            struct socket **rx_sock)
 667 {
 668         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 669         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 670 }
 671
 672 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 673 {
 674         vhost_poll_flush(n->poll + index);
 675         vhost_poll_flush(&n->dev.vqs[index].poll);
 676 }
 677
 678 static void vhost_net_flush(struct vhost_net *n)
 679 {
 680         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 681         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 682 }
 683
 684 static int vhost_net_release(struct inode *inode, struct file *f)
 685 {
 686         struct vhost_net *n = f->private_data;
 687         struct socket *tx_sock;
 688         struct socket *rx_sock;
 689         int i;
 690
 691         vhost_net_stop(n, &tx_sock, &rx_sock);
 692         vhost_net_flush(n);
 693         vhost_dev_stop(&n->dev);
 694         for (i = 0; i < n->dev.nvqs; ++i) {
 695                 /* Wait for all lower device DMAs done. */
 696                 if (n->dev.vqs[i].ubufs)
 697                         vhost_ubuf_put_and_wait(n->dev.vqs[i].ubufs);
 698
 699                 vhost_zerocopy_signal_used(n, &n->dev.vqs[i]);
 700         }
 701         vhost_dev_cleanup(&n->dev, false);
 702         if (tx_sock)
 703                 fput(tx_sock->file);
 704         if (rx_sock)
 705                 fput(rx_sock->file);
 706         /* We do an extra flush before freeing memory,
 707          * since jobs can re-queue themselves. */
 708         vhost_net_flush(n);
 709         kfree(n);
 710         return 0;
 711 }
 712
 713 static struct socket *get_raw_socket(int fd)
 714 {
 715         struct {
 716                 struct sockaddr_ll sa;
 717                 char  buf[MAX_ADDR_LEN];
 718         } uaddr;
 719         int uaddr_len = sizeof uaddr, r;
 720         struct socket *sock = sockfd_lookup(fd, &r);
 721
 722         if (!sock)
 723                 return ERR_PTR(-ENOTSOCK);
 724
 725         /* Parameter checking */
 726         if (sock->sk->sk_type != SOCK_RAW) {
 727                 r = -ESOCKTNOSUPPORT;
 728                 goto err;
 729         }
 730
 731         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 732                                &uaddr_len, 0);
 733         if (r)
 734                 goto err;
 735
 736         if (uaddr.sa.sll_family != AF_PACKET) {
 737                 r = -EPFNOSUPPORT;
 738                 goto err;
 739         }
 740         return sock;
 741 err:
 742         fput(sock->file);
 743         return ERR_PTR(r);
 744 }
 745
 746 static struct socket *get_tap_socket(int fd)
 747 {
 748         struct file *file = fget(fd);
 749         struct socket *sock;
 750
 751         if (!file)
 752                 return ERR_PTR(-EBADF);
 753         sock = tun_get_socket(file);
 754         if (!IS_ERR(sock))
 755                 return sock;
 756         sock = macvtap_get_socket(file);
 757         if (IS_ERR(sock))
 758                 fput(file);
 759         return sock;
 760 }
 761
 762 static struct socket *get_socket(int fd)
 763 {
 764         struct socket *sock;
 765
 766         /* special case to disable backend */
 767         if (fd == -1)
 768                 return NULL;
 769         sock = get_raw_socket(fd);
 770         if (!IS_ERR(sock))
 771                 return sock;
 772         sock = get_tap_socket(fd);
 773         if (!IS_ERR(sock))
 774                 return sock;
 775         return ERR_PTR(-ENOTSOCK);
 776 }
 777
 778 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 779 {
 780         struct socket *sock, *oldsock;
 781         struct vhost_virtqueue *vq;
 782         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 783         int r;
 784
 785         mutex_lock(&n->dev.mutex);
 786         r = vhost_dev_check_owner(&n->dev);
 787         if (r)
 788                 goto err;
 789
 790         if (index >= VHOST_NET_VQ_MAX) {
 791                 r = -ENOBUFS;
 792                 goto err;
 793         }
 794         vq = n->vqs + index;
 795         mutex_lock(&vq->mutex);
 796
 797         /* Verify that ring has been setup correctly. */
 798         if (!vhost_vq_access_ok(vq)) {
 799                 r = -EFAULT;
 800                 goto err_vq;
 801         }
 802         sock = get_socket(fd);
 803         if (IS_ERR(sock)) {
 804                 r = PTR_ERR(sock);
 805                 goto err_vq;
 806         }
 807
 808         /* start polling new socket */
 809         oldsock = rcu_dereference_protected(vq->private_data,
 810                                             lockdep_is_held(&vq->mutex));
 811         if (sock != oldsock) {
 812                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 813                 if (IS_ERR(ubufs)) {
 814                         r = PTR_ERR(ubufs);
 815                         goto err_ubufs;
 816                 }
 817                 oldubufs = vq->ubufs;
 818                 vq->ubufs = ubufs;
 819                 vhost_net_disable_vq(n, vq);
 820                 rcu_assign_pointer(vq->private_data, sock);
 821                 vhost_net_enable_vq(n, vq);
 822
 823                 r = vhost_init_used(vq);
 824                 if (r)
 825                         goto err_vq;
 826         }
 827
 828         mutex_unlock(&vq->mutex);
 829
 830         if (oldubufs) {
 831                 vhost_ubuf_put_and_wait(oldubufs);
 832                 mutex_lock(&vq->mutex);
 833                 vhost_zerocopy_signal_used(n, vq);
 834                 mutex_unlock(&vq->mutex);
 835         }
 836
 837         if (oldsock) {
 838                 vhost_net_flush_vq(n, index);
 839                 fput(oldsock->file);
 840         }
 841
 842         mutex_unlock(&n->dev.mutex);
 843         return 0;
 844
 845 err_ubufs:
 846         fput(sock->file);
 847 err_vq:
 848         mutex_unlock(&vq->mutex);
 849 err:
 850         mutex_unlock(&n->dev.mutex);
 851         return r;
 852 }
 853
 854 static long vhost_net_reset_owner(struct vhost_net *n)
 855 {
 856         struct socket *tx_sock = NULL;
 857         struct socket *rx_sock = NULL;
 858         long err;
 859
 860         mutex_lock(&n->dev.mutex);
 861         err = vhost_dev_check_owner(&n->dev);
 862         if (err)
 863                 goto done;
 864         vhost_net_stop(n, &tx_sock, &rx_sock);
 865         vhost_net_flush(n);
 866         err = vhost_dev_reset_owner(&n->dev);
 867 done:
 868         mutex_unlock(&n->dev.mutex);
 869         if (tx_sock)
 870                 fput(tx_sock->file);
 871         if (rx_sock)
 872                 fput(rx_sock->file);
 873         return err;
 874 }
 875
 876 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 877 {
 878         size_t vhost_hlen, sock_hlen, hdr_len;
 879         int i;
 880
 881         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 882                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 883                         sizeof(struct virtio_net_hdr);
 884         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 885                 /* vhost provides vnet_hdr */
 886                 vhost_hlen = hdr_len;
 887                 sock_hlen = 0;
 888         } else {
 889                 /* socket provides vnet_hdr */
 890                 vhost_hlen = 0;
 891                 sock_hlen = hdr_len;
 892         }
 893         mutex_lock(&n->dev.mutex);
 894         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 895             !vhost_log_access_ok(&n->dev)) {
 896                 mutex_unlock(&n->dev.mutex);
 897                 return -EFAULT;
 898         }
 899         n->dev.acked_features = features;
 900         smp_wmb();
 901         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 902                 mutex_lock(&n->vqs[i].mutex);
 903                 n->vqs[i].vhost_hlen = vhost_hlen;
 904                 n->vqs[i].sock_hlen = sock_hlen;
 905                 mutex_unlock(&n->vqs[i].mutex);
 906         }
 907         vhost_net_flush(n);
 908         mutex_unlock(&n->dev.mutex);
 909         return 0;
 910 }
 911
 912 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 913                             unsigned long arg)
 914 {
 915         struct vhost_net *n = f->private_data;
 916         void __user *argp = (void __user *)arg;
 917         u64 __user *featurep = argp;
 918         struct vhost_vring_file backend;
 919         u64 features;
 920         int r;
 921
 922         switch (ioctl) {
 923         case VHOST_NET_SET_BACKEND:
 924                 if (copy_from_user(&backend, argp, sizeof backend))
 925                         return -EFAULT;
 926                 return vhost_net_set_backend(n, backend.index, backend.fd);
 927         case VHOST_GET_FEATURES:
 928                 features = VHOST_NET_FEATURES;
 929                 if (copy_to_user(featurep, &features, sizeof features))
 930                         return -EFAULT;
 931                 return 0;
 932         case VHOST_SET_FEATURES:
 933                 if (copy_from_user(&features, featurep, sizeof features))
 934                         return -EFAULT;
 935                 if (features & ~VHOST_NET_FEATURES)
 936                         return -EOPNOTSUPP;
 937                 return vhost_net_set_features(n, features);
 938         case VHOST_RESET_OWNER:
 939                 return vhost_net_reset_owner(n);
 940         default:
 941                 mutex_lock(&n->dev.mutex);
 942                 r = vhost_dev_ioctl(&n->dev, ioctl, arg);
 943                 vhost_net_flush(n);
 944                 mutex_unlock(&n->dev.mutex);
 945                 return r;
 946         }
 947 }
 948
 949 #ifdef CONFIG_COMPAT
 950 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 951                                    unsigned long arg)
 952 {
 953         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 954 }
 955 #endif
 956
 957 static const struct file_operations vhost_net_fops = {
 958         .owner          = THIS_MODULE,
 959         .release        = vhost_net_release,
 960         .unlocked_ioctl = vhost_net_ioctl,
 961 #ifdef CONFIG_COMPAT
 962         .compat_ioctl   = vhost_net_compat_ioctl,
 963 #endif
 964         .open           = vhost_net_open,
 965         .llseek         = noop_llseek,
 966 };
 967
 968 static struct miscdevice vhost_net_misc = {
 969         .minor = VHOST_NET_MINOR,
 970         .name = "vhost-net",
 971         .fops = &vhost_net_fops,
 972 };
 973
 974 static int vhost_net_init(void)
 975 {
 976         if (experimental_zcopytx)
 977                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
 978         return misc_register(&vhost_net_misc);
 979 }
 980 module_init(vhost_net_init);
 981
 982 static void vhost_net_exit(void)
 983 {
 984         misc_deregister(&vhost_net_misc);
 985 }
 986 module_exit(vhost_net_exit);
 987
 988 MODULE_VERSION("0.0.1");
 989 MODULE_LICENSE("GPL v2");
 990 MODULE_AUTHOR("Michael S. Tsirkin");
 991 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
 992 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
 993 MODULE_ALIAS("devname:vhost-net");