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