Pileus Git - ~andy/linux/blob - 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         /* Flush in progress. Protected by tx vq lock. */
  87         bool tx_flush;
  88 };
  89
  90 static void vhost_net_tx_packet(struct vhost_net *net)
  91 {
  92         ++net->tx_packets;
  93         if (net->tx_packets < 1024)
  94                 return;
  95         net->tx_packets = 0;
  96         net->tx_zcopy_err = 0;
  97 }
  98
  99 static void vhost_net_tx_err(struct vhost_net *net)
 100 {
 101         ++net->tx_zcopy_err;
 102 }
 103
 104 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 105 {
 106         /* TX flush waits for outstanding DMAs to be done.
 107          * Don't start new DMAs.
 108          */
 109         return !net->tx_flush &&
 110                 net->tx_packets / 64 >= net->tx_zcopy_err;
 111 }
 112
 113 static bool vhost_sock_zcopy(struct socket *sock)
 114 {
 115         return unlikely(experimental_zcopytx) &&
 116                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 117 }
 118
 119 /* Pop first len bytes from iovec. Return number of segments used. */
 120 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 121                           size_t len, int iov_count)
 122 {
 123         int seg = 0;
 124         size_t size;
 125
 126         while (len && seg < iov_count) {
 127                 size = min(from->iov_len, len);
 128                 to->iov_base = from->iov_base;
 129                 to->iov_len = size;
 130                 from->iov_len -= size;
 131                 from->iov_base += size;
 132                 len -= size;
 133                 ++from;
 134                 ++to;
 135                 ++seg;
 136         }
 137         return seg;
 138 }
 139 /* Copy iovec entries for len bytes from iovec. */
 140 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 141                            size_t len, int iovcount)
 142 {
 143         int seg = 0;
 144         size_t size;
 145
 146         while (len && seg < iovcount) {
 147                 size = min(from->iov_len, len);
 148                 to->iov_base = from->iov_base;
 149                 to->iov_len = size;
 150                 len -= size;
 151                 ++from;
 152                 ++to;
 153                 ++seg;
 154         }
 155 }
 156
 157 /* Caller must have TX VQ lock */
 158 static void tx_poll_stop(struct vhost_net *net)
 159 {
 160         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
 161                 return;
 162         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 163         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 164 }
 165
 166 /* Caller must have TX VQ lock */
 167 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 168 {
 169         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 170                 return;
 171         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 172         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 173 }
 174
 175 /* In case of DMA done not in order in lower device driver for some reason.
 176  * upend_idx is used to track end of used idx, done_idx is used to track head
 177  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 178  * guest used idx.
 179  */
 180 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 181                                       struct vhost_virtqueue *vq)
 182 {
 183         int i;
 184         int j = 0;
 185
 186         for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 187                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 188                         vhost_net_tx_err(net);
 189                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 190                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 191                         vhost_add_used_and_signal(vq->dev, vq,
 192                                                   vq->heads[i].id, 0);
 193                         ++j;
 194                 } else
 195                         break;
 196         }
 197         if (j)
 198                 vq->done_idx = i;
 199         return j;
 200 }
 201
 202 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 203 {
 204         struct vhost_ubuf_ref *ubufs = ubuf->ctx;
 205         struct vhost_virtqueue *vq = ubufs->vq;
 206         int cnt = atomic_read(&ubufs->kref.refcount);
 207
 208         /*
 209          * Trigger polling thread if guest stopped submitting new buffers:
 210          * in this case, the refcount after decrement will eventually reach 1
 211          * so here it is 2.
 212          * We also trigger polling periodically after each 16 packets
 213          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 214          * less than 10% of times).
 215          */
 216         if (cnt <= 2 || !(cnt % 16))
 217                 vhost_poll_queue(&vq->poll);
 218         /* set len to mark this desc buffers done DMA */
 219         vq->heads[ubuf->desc].len = success ?
 220                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 221         vhost_ubuf_put(ubufs);
 222 }
 223
 224 /* Expects to be always run from workqueue - which acts as
 225  * read-size critical section for our kind of RCU. */
 226 static void handle_tx(struct vhost_net *net)
 227 {
 228         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 229         unsigned out, in, s;
 230         int head;
 231         struct msghdr msg = {
 232                 .msg_name = NULL,
 233                 .msg_namelen = 0,
 234                 .msg_control = NULL,
 235                 .msg_controllen = 0,
 236                 .msg_iov = vq->iov,
 237                 .msg_flags = MSG_DONTWAIT,
 238         };
 239         size_t len, total_len = 0;
 240         int err, wmem;
 241         size_t hdr_size;
 242         struct socket *sock;
 243         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 244         bool zcopy;
 245
 246         /* TODO: check that we are running from vhost_worker? */
 247         sock = rcu_dereference_check(vq->private_data, 1);
 248         if (!sock)
 249                 return;
 250
 251         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 252         if (wmem >= sock->sk->sk_sndbuf) {
 253                 mutex_lock(&vq->mutex);
 254                 tx_poll_start(net, sock);
 255                 mutex_unlock(&vq->mutex);
 256                 return;
 257         }
 258
 259         mutex_lock(&vq->mutex);
 260         vhost_disable_notify(&net->dev, vq);
 261
 262         if (wmem < sock->sk->sk_sndbuf / 2)
 263                 tx_poll_stop(net);
 264         hdr_size = vq->vhost_hlen;
 265         zcopy = vq->ubufs;
 266
 267         for (;;) {
 268                 /* Release DMAs done buffers first */
 269                 if (zcopy)
 270                         vhost_zerocopy_signal_used(net, vq);
 271
 272                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 273                                          ARRAY_SIZE(vq->iov),
 274                                          &out, &in,
 275                                          NULL, NULL);
 276                 /* On error, stop handling until the next kick. */
 277                 if (unlikely(head < 0))
 278                         break;
 279                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 280                 if (head == vq->num) {
 281                         int num_pends;
 282
 283                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 284                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 285                                 tx_poll_start(net, sock);
 286                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 287                                 break;
 288                         }
 289                         /* If more outstanding DMAs, queue the work.
 290                          * Handle upend_idx wrap around
 291                          */
 292                         num_pends = likely(vq->upend_idx >= vq->done_idx) ?
 293                                     (vq->upend_idx - vq->done_idx) :
 294                                     (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
 295                         if (unlikely(num_pends > VHOST_MAX_PEND)) {
 296                                 tx_poll_start(net, sock);
 297                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 298                                 break;
 299                         }
 300                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 301                                 vhost_disable_notify(&net->dev, vq);
 302                                 continue;
 303                         }
 304                         break;
 305                 }
 306                 if (in) {
 307                         vq_err(vq, "Unexpected descriptor format for TX: "
 308                                "out %d, int %d\n", out, in);
 309                         break;
 310                 }
 311                 /* Skip header. TODO: support TSO. */
 312                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 313                 msg.msg_iovlen = out;
 314                 len = iov_length(vq->iov, out);
 315                 /* Sanity check */
 316                 if (!len) {
 317                         vq_err(vq, "Unexpected header len for TX: "
 318                                "%zd expected %zd\n",
 319                                iov_length(vq->hdr, s), hdr_size);
 320                         break;
 321                 }
 322                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 323                 if (zcopy) {
 324                         vq->heads[vq->upend_idx].id = head;
 325                         if (!vhost_net_tx_select_zcopy(net) ||
 326                             len < VHOST_GOODCOPY_LEN) {
 327                                 /* copy don't need to wait for DMA done */
 328                                 vq->heads[vq->upend_idx].len =
 329                                                         VHOST_DMA_DONE_LEN;
 330                                 msg.msg_control = NULL;
 331                                 msg.msg_controllen = 0;
 332                                 ubufs = NULL;
 333                         } else {
 334                                 struct ubuf_info *ubuf = &vq->ubuf_info[head];
 335
 336                                 vq->heads[vq->upend_idx].len =
 337                                         VHOST_DMA_IN_PROGRESS;
 338                                 ubuf->callback = vhost_zerocopy_callback;
 339                                 ubuf->ctx = vq->ubufs;
 340                                 ubuf->desc = vq->upend_idx;
 341                                 msg.msg_control = ubuf;
 342                                 msg.msg_controllen = sizeof(ubuf);
 343                                 ubufs = vq->ubufs;
 344                                 kref_get(&ubufs->kref);
 345                         }
 346                         vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
 347                 }
 348                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 349                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 350                 if (unlikely(err < 0)) {
 351                         if (zcopy) {
 352                                 if (ubufs)
 353                                         vhost_ubuf_put(ubufs);
 354                                 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
 355                                         UIO_MAXIOV;
 356                         }
 357                         vhost_discard_vq_desc(vq, 1);
 358                         if (err == -EAGAIN || err == -ENOBUFS)
 359                                 tx_poll_start(net, sock);
 360                         break;
 361                 }
 362                 if (err != len)
 363                         pr_debug("Truncated TX packet: "
 364                                  " len %d != %zd\n", err, len);
 365                 if (!zcopy)
 366                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 367                 else
 368                         vhost_zerocopy_signal_used(net, vq);
 369                 total_len += len;
 370                 vhost_net_tx_packet(net);
 371                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 372                         vhost_poll_queue(&vq->poll);
 373                         break;
 374                 }
 375         }
 376
 377         mutex_unlock(&vq->mutex);
 378 }
 379
 380 static int peek_head_len(struct sock *sk)
 381 {
 382         struct sk_buff *head;
 383         int len = 0;
 384         unsigned long flags;
 385
 386         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 387         head = skb_peek(&sk->sk_receive_queue);
 388         if (likely(head)) {
 389                 len = head->len;
 390                 if (vlan_tx_tag_present(head))
 391                         len += VLAN_HLEN;
 392         }
 393
 394         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 395         return len;
 396 }
 397
 398 /* This is a multi-buffer version of vhost_get_desc, that works if
 399  *      vq has read descriptors only.
 400  * @vq          - the relevant virtqueue
 401  * @datalen     - data length we'll be reading
 402  * @iovcount    - returned count of io vectors we fill
 403  * @log         - vhost log
 404  * @log_num     - log offset
 405  * @quota       - headcount quota, 1 for big buffer
 406  *      returns number of buffer heads allocated, negative on error
 407  */
 408 static int get_rx_bufs(struct vhost_virtqueue *vq,
 409                        struct vring_used_elem *heads,
 410                        int datalen,
 411                        unsigned *iovcount,
 412                        struct vhost_log *log,
 413                        unsigned *log_num,
 414                        unsigned int quota)
 415 {
 416         unsigned int out, in;
 417         int seg = 0;
 418         int headcount = 0;
 419         unsigned d;
 420         int r, nlogs = 0;
 421
 422         while (datalen > 0 && headcount < quota) {
 423                 if (unlikely(seg >= UIO_MAXIOV)) {
 424                         r = -ENOBUFS;
 425                         goto err;
 426                 }
 427                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 428                                       ARRAY_SIZE(vq->iov) - seg, &out,
 429                                       &in, log, log_num);
 430                 if (d == vq->num) {
 431                         r = 0;
 432                         goto err;
 433                 }
 434                 if (unlikely(out || in <= 0)) {
 435                         vq_err(vq, "unexpected descriptor format for RX: "
 436                                 "out %d, in %d\n", out, in);
 437                         r = -EINVAL;
 438                         goto err;
 439                 }
 440                 if (unlikely(log)) {
 441                         nlogs += *log_num;
 442                         log += *log_num;
 443                 }
 444                 heads[headcount].id = d;
 445                 heads[headcount].len = iov_length(vq->iov + seg, in);
 446                 datalen -= heads[headcount].len;
 447                 ++headcount;
 448                 seg += in;
 449         }
 450         heads[headcount - 1].len += datalen;
 451         *iovcount = seg;
 452         if (unlikely(log))
 453                 *log_num = nlogs;
 454         return headcount;
 455 err:
 456         vhost_discard_vq_desc(vq, headcount);
 457         return r;
 458 }
 459
 460 /* Expects to be always run from workqueue - which acts as
 461  * read-size critical section for our kind of RCU. */
 462 static void handle_rx(struct vhost_net *net)
 463 {
 464         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 465         unsigned uninitialized_var(in), log;
 466         struct vhost_log *vq_log;
 467         struct msghdr msg = {
 468                 .msg_name = NULL,
 469                 .msg_namelen = 0,
 470                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 471                 .msg_controllen = 0,
 472                 .msg_iov = vq->iov,
 473                 .msg_flags = MSG_DONTWAIT,
 474         };
 475         struct virtio_net_hdr_mrg_rxbuf hdr = {
 476                 .hdr.flags = 0,
 477                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 478         };
 479         size_t total_len = 0;
 480         int err, mergeable;
 481         s16 headcount;
 482         size_t vhost_hlen, sock_hlen;
 483         size_t vhost_len, sock_len;
 484         /* TODO: check that we are running from vhost_worker? */
 485         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 486
 487         if (!sock)
 488                 return;
 489
 490         mutex_lock(&vq->mutex);
 491         vhost_disable_notify(&net->dev, vq);
 492         vhost_hlen = vq->vhost_hlen;
 493         sock_hlen = vq->sock_hlen;
 494
 495         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 496                 vq->log : NULL;
 497         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 498
 499         while ((sock_len = peek_head_len(sock->sk))) {
 500                 sock_len += sock_hlen;
 501                 vhost_len = sock_len + vhost_hlen;
 502                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 503                                         &in, vq_log, &log,
 504                                         likely(mergeable) ? UIO_MAXIOV : 1);
 505                 /* On error, stop handling until the next kick. */
 506                 if (unlikely(headcount < 0))
 507                         break;
 508                 /* OK, now we need to know about added descriptors. */
 509                 if (!headcount) {
 510                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 511                                 /* They have slipped one in as we were
 512                                  * doing that: check again. */
 513                                 vhost_disable_notify(&net->dev, vq);
 514                                 continue;
 515                         }
 516                         /* Nothing new?  Wait for eventfd to tell us
 517                          * they refilled. */
 518                         break;
 519                 }
 520                 /* We don't need to be notified again. */
 521                 if (unlikely((vhost_hlen)))
 522                         /* Skip header. TODO: support TSO. */
 523                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 524                 else
 525                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 526                          * needed because recvmsg can modify msg_iov. */
 527                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 528                 msg.msg_iovlen = in;
 529                 err = sock->ops->recvmsg(NULL, sock, &msg,
 530                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 531                 /* Userspace might have consumed the packet meanwhile:
 532                  * it's not supposed to do this usually, but might be hard
 533                  * to prevent. Discard data we got (if any) and keep going. */
 534                 if (unlikely(err != sock_len)) {
 535                         pr_debug("Discarded rx packet: "
 536                                  " len %d, expected %zd\n", err, sock_len);
 537                         vhost_discard_vq_desc(vq, headcount);
 538                         continue;
 539                 }
 540                 if (unlikely(vhost_hlen) &&
 541                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 542                                       vhost_hlen)) {
 543                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 544                                vq->iov->iov_base);
 545                         break;
 546                 }
 547                 /* TODO: Should check and handle checksum. */
 548                 if (likely(mergeable) &&
 549                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 550                                       offsetof(typeof(hdr), num_buffers),
 551                                       sizeof hdr.num_buffers)) {
 552                         vq_err(vq, "Failed num_buffers write");
 553                         vhost_discard_vq_desc(vq, headcount);
 554                         break;
 555                 }
 556                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 557                                             headcount);
 558                 if (unlikely(vq_log))
 559                         vhost_log_write(vq, vq_log, log, vhost_len);
 560                 total_len += vhost_len;
 561                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 562                         vhost_poll_queue(&vq->poll);
 563                         break;
 564                 }
 565         }
 566
 567         mutex_unlock(&vq->mutex);
 568 }
 569
 570 static void handle_tx_kick(struct vhost_work *work)
 571 {
 572         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 573                                                   poll.work);
 574         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 575
 576         handle_tx(net);
 577 }
 578
 579 static void handle_rx_kick(struct vhost_work *work)
 580 {
 581         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 582                                                   poll.work);
 583         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 584
 585         handle_rx(net);
 586 }
 587
 588 static void handle_tx_net(struct vhost_work *work)
 589 {
 590         struct vhost_net *net = container_of(work, struct vhost_net,
 591                                              poll[VHOST_NET_VQ_TX].work);
 592         handle_tx(net);
 593 }
 594
 595 static void handle_rx_net(struct vhost_work *work)
 596 {
 597         struct vhost_net *net = container_of(work, struct vhost_net,
 598                                              poll[VHOST_NET_VQ_RX].work);
 599         handle_rx(net);
 600 }
 601
 602 static int vhost_net_open(struct inode *inode, struct file *f)
 603 {
 604         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 605         struct vhost_dev *dev;
 606         int r;
 607
 608         if (!n)
 609                 return -ENOMEM;
 610
 611         dev = &n->dev;
 612         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 613         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 614         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 615         if (r < 0) {
 616                 kfree(n);
 617                 return r;
 618         }
 619
 620         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 621         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 622         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 623
 624         f->private_data = n;
 625
 626         return 0;
 627 }
 628
 629 static void vhost_net_disable_vq(struct vhost_net *n,
 630                                  struct vhost_virtqueue *vq)
 631 {
 632         if (!vq->private_data)
 633                 return;
 634         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 635                 tx_poll_stop(n);
 636                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 637         } else
 638                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 639 }
 640
 641 static void vhost_net_enable_vq(struct vhost_net *n,
 642                                 struct vhost_virtqueue *vq)
 643 {
 644         struct socket *sock;
 645
 646         sock = rcu_dereference_protected(vq->private_data,
 647                                          lockdep_is_held(&vq->mutex));
 648         if (!sock)
 649                 return;
 650         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 651                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 652                 tx_poll_start(n, sock);
 653         } else
 654                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 655 }
 656
 657 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 658                                         struct vhost_virtqueue *vq)
 659 {
 660         struct socket *sock;
 661
 662         mutex_lock(&vq->mutex);
 663         sock = rcu_dereference_protected(vq->private_data,
 664                                          lockdep_is_held(&vq->mutex));
 665         vhost_net_disable_vq(n, vq);
 666         rcu_assign_pointer(vq->private_data, NULL);
 667         mutex_unlock(&vq->mutex);
 668         return sock;
 669 }
 670
 671 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 672                            struct socket **rx_sock)
 673 {
 674         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 675         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 676 }
 677
 678 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 679 {
 680         vhost_poll_flush(n->poll + index);
 681         vhost_poll_flush(&n->dev.vqs[index].poll);
 682 }
 683
 684 static void vhost_net_flush(struct vhost_net *n)
 685 {
 686         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 687         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 688         if (n->dev.vqs[VHOST_NET_VQ_TX].ubufs) {
 689                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 690                 n->tx_flush = true;
 691                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 692                 /* Wait for all lower device DMAs done. */
 693                 vhost_ubuf_put_and_wait(n->dev.vqs[VHOST_NET_VQ_TX].ubufs);
 694                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 695                 n->tx_flush = false;
 696                 kref_init(&n->dev.vqs[VHOST_NET_VQ_TX].ubufs->kref);
 697                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 698         }
 699 }
 700
 701 static int vhost_net_release(struct inode *inode, struct file *f)
 702 {
 703         struct vhost_net *n = f->private_data;
 704         struct socket *tx_sock;
 705         struct socket *rx_sock;
 706
 707         vhost_net_stop(n, &tx_sock, &rx_sock);
 708         vhost_net_flush(n);
 709         vhost_dev_stop(&n->dev);
 710         vhost_dev_cleanup(&n->dev, false);
 711         if (tx_sock)
 712                 fput(tx_sock->file);
 713         if (rx_sock)
 714                 fput(rx_sock->file);
 715         /* We do an extra flush before freeing memory,
 716          * since jobs can re-queue themselves. */
 717         vhost_net_flush(n);
 718         kfree(n);
 719         return 0;
 720 }
 721
 722 static struct socket *get_raw_socket(int fd)
 723 {
 724         struct {
 725                 struct sockaddr_ll sa;
 726                 char  buf[MAX_ADDR_LEN];
 727         } uaddr;
 728         int uaddr_len = sizeof uaddr, r;
 729         struct socket *sock = sockfd_lookup(fd, &r);
 730
 731         if (!sock)
 732                 return ERR_PTR(-ENOTSOCK);
 733
 734         /* Parameter checking */
 735         if (sock->sk->sk_type != SOCK_RAW) {
 736                 r = -ESOCKTNOSUPPORT;
 737                 goto err;
 738         }
 739
 740         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 741                                &uaddr_len, 0);
 742         if (r)
 743                 goto err;
 744
 745         if (uaddr.sa.sll_family != AF_PACKET) {
 746                 r = -EPFNOSUPPORT;
 747                 goto err;
 748         }
 749         return sock;
 750 err:
 751         fput(sock->file);
 752         return ERR_PTR(r);
 753 }
 754
 755 static struct socket *get_tap_socket(int fd)
 756 {
 757         struct file *file = fget(fd);
 758         struct socket *sock;
 759
 760         if (!file)
 761                 return ERR_PTR(-EBADF);
 762         sock = tun_get_socket(file);
 763         if (!IS_ERR(sock))
 764                 return sock;
 765         sock = macvtap_get_socket(file);
 766         if (IS_ERR(sock))
 767                 fput(file);
 768         return sock;
 769 }
 770
 771 static struct socket *get_socket(int fd)
 772 {
 773         struct socket *sock;
 774
 775         /* special case to disable backend */
 776         if (fd == -1)
 777                 return NULL;
 778         sock = get_raw_socket(fd);
 779         if (!IS_ERR(sock))
 780                 return sock;
 781         sock = get_tap_socket(fd);
 782         if (!IS_ERR(sock))
 783                 return sock;
 784         return ERR_PTR(-ENOTSOCK);
 785 }
 786
 787 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 788 {
 789         struct socket *sock, *oldsock;
 790         struct vhost_virtqueue *vq;
 791         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 792         int r;
 793
 794         mutex_lock(&n->dev.mutex);
 795         r = vhost_dev_check_owner(&n->dev);
 796         if (r)
 797                 goto err;
 798
 799         if (index >= VHOST_NET_VQ_MAX) {
 800                 r = -ENOBUFS;
 801                 goto err;
 802         }
 803         vq = n->vqs + index;
 804         mutex_lock(&vq->mutex);
 805
 806         /* Verify that ring has been setup correctly. */
 807         if (!vhost_vq_access_ok(vq)) {
 808                 r = -EFAULT;
 809                 goto err_vq;
 810         }
 811         sock = get_socket(fd);
 812         if (IS_ERR(sock)) {
 813                 r = PTR_ERR(sock);
 814                 goto err_vq;
 815         }
 816
 817         /* start polling new socket */
 818         oldsock = rcu_dereference_protected(vq->private_data,
 819                                             lockdep_is_held(&vq->mutex));
 820         if (sock != oldsock) {
 821                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 822                 if (IS_ERR(ubufs)) {
 823                         r = PTR_ERR(ubufs);
 824                         goto err_ubufs;
 825                 }
 826                 oldubufs = vq->ubufs;
 827                 vq->ubufs = ubufs;
 828                 vhost_net_disable_vq(n, vq);
 829                 rcu_assign_pointer(vq->private_data, sock);
 830                 vhost_net_enable_vq(n, vq);
 831
 832                 r = vhost_init_used(vq);
 833                 if (r)
 834                         goto err_vq;
 835
 836                 n->tx_packets = 0;
 837                 n->tx_zcopy_err = 0;
 838                 n->tx_flush = false;
 839         }
 840
 841         mutex_unlock(&vq->mutex);
 842
 843         if (oldubufs) {
 844                 vhost_ubuf_put_and_wait(oldubufs);
 845                 mutex_lock(&vq->mutex);
 846                 vhost_zerocopy_signal_used(n, vq);
 847                 mutex_unlock(&vq->mutex);
 848         }
 849
 850         if (oldsock) {
 851                 vhost_net_flush_vq(n, index);
 852                 fput(oldsock->file);
 853         }
 854
 855         mutex_unlock(&n->dev.mutex);
 856         return 0;
 857
 858 err_ubufs:
 859         fput(sock->file);
 860 err_vq:
 861         mutex_unlock(&vq->mutex);
 862 err:
 863         mutex_unlock(&n->dev.mutex);
 864         return r;
 865 }
 866
 867 static long vhost_net_reset_owner(struct vhost_net *n)
 868 {
 869         struct socket *tx_sock = NULL;
 870         struct socket *rx_sock = NULL;
 871         long err;
 872
 873         mutex_lock(&n->dev.mutex);
 874         err = vhost_dev_check_owner(&n->dev);
 875         if (err)
 876                 goto done;
 877         vhost_net_stop(n, &tx_sock, &rx_sock);
 878         vhost_net_flush(n);
 879         err = vhost_dev_reset_owner(&n->dev);
 880 done:
 881         mutex_unlock(&n->dev.mutex);
 882         if (tx_sock)
 883                 fput(tx_sock->file);
 884         if (rx_sock)
 885                 fput(rx_sock->file);
 886         return err;
 887 }
 888
 889 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 890 {
 891         size_t vhost_hlen, sock_hlen, hdr_len;
 892         int i;
 893
 894         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 895                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 896                         sizeof(struct virtio_net_hdr);
 897         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 898                 /* vhost provides vnet_hdr */
 899                 vhost_hlen = hdr_len;
 900                 sock_hlen = 0;
 901         } else {
 902                 /* socket provides vnet_hdr */
 903                 vhost_hlen = 0;
 904                 sock_hlen = hdr_len;
 905         }
 906         mutex_lock(&n->dev.mutex);
 907         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 908             !vhost_log_access_ok(&n->dev)) {
 909                 mutex_unlock(&n->dev.mutex);
 910                 return -EFAULT;
 911         }
 912         n->dev.acked_features = features;
 913         smp_wmb();
 914         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 915                 mutex_lock(&n->vqs[i].mutex);
 916                 n->vqs[i].vhost_hlen = vhost_hlen;
 917                 n->vqs[i].sock_hlen = sock_hlen;
 918                 mutex_unlock(&n->vqs[i].mutex);
 919         }
 920         vhost_net_flush(n);
 921         mutex_unlock(&n->dev.mutex);
 922         return 0;
 923 }
 924
 925 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 926                             unsigned long arg)
 927 {
 928         struct vhost_net *n = f->private_data;
 929         void __user *argp = (void __user *)arg;
 930         u64 __user *featurep = argp;
 931         struct vhost_vring_file backend;
 932         u64 features;
 933         int r;
 934
 935         switch (ioctl) {
 936         case VHOST_NET_SET_BACKEND:
 937                 if (copy_from_user(&backend, argp, sizeof backend))
 938                         return -EFAULT;
 939                 return vhost_net_set_backend(n, backend.index, backend.fd);
 940         case VHOST_GET_FEATURES:
 941                 features = VHOST_NET_FEATURES;
 942                 if (copy_to_user(featurep, &features, sizeof features))
 943                         return -EFAULT;
 944                 return 0;
 945         case VHOST_SET_FEATURES:
 946                 if (copy_from_user(&features, featurep, sizeof features))
 947                         return -EFAULT;
 948                 if (features & ~VHOST_NET_FEATURES)
 949                         return -EOPNOTSUPP;
 950                 return vhost_net_set_features(n, features);
 951         case VHOST_RESET_OWNER:
 952                 return vhost_net_reset_owner(n);
 953         default:
 954                 mutex_lock(&n->dev.mutex);
 955                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
 956                 if (r == -ENOIOCTLCMD)
 957                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
 958                 else
 959                         vhost_net_flush(n);
 960                 mutex_unlock(&n->dev.mutex);
 961                 return r;
 962         }
 963 }
 964
 965 #ifdef CONFIG_COMPAT
 966 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 967                                    unsigned long arg)
 968 {
 969         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 970 }
 971 #endif
 972
 973 static const struct file_operations vhost_net_fops = {
 974         .owner          = THIS_MODULE,
 975         .release        = vhost_net_release,
 976         .unlocked_ioctl = vhost_net_ioctl,
 977 #ifdef CONFIG_COMPAT
 978         .compat_ioctl   = vhost_net_compat_ioctl,
 979 #endif
 980         .open           = vhost_net_open,
 981         .llseek         = noop_llseek,
 982 };
 983
 984 static struct miscdevice vhost_net_misc = {
 985         .minor = VHOST_NET_MINOR,
 986         .name = "vhost-net",
 987         .fops = &vhost_net_fops,
 988 };
 989
 990 static int vhost_net_init(void)
 991 {
 992         if (experimental_zcopytx)
 993                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
 994         return misc_register(&vhost_net_misc);
 995 }
 996 module_init(vhost_net_init);
 997
 998 static void vhost_net_exit(void)
 999 {
1000         misc_deregister(&vhost_net_misc);
1001 }
1002 module_exit(vhost_net_exit);
1003
1004 MODULE_VERSION("0.0.1");
1005 MODULE_LICENSE("GPL v2");
1006 MODULE_AUTHOR("Michael S. Tsirkin");
1007 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1008 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1009 MODULE_ALIAS("devname:vhost-net");