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