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 enum vhost_net_poll_state {
  68         VHOST_NET_POLL_DISABLED = 0,
  69         VHOST_NET_POLL_STARTED = 1,
  70         VHOST_NET_POLL_STOPPED = 2,
  71 };
  72
  73 struct vhost_net {
  74         struct vhost_dev dev;
  75         struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  76         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  77         /* Tells us whether we are polling a socket for TX.
  78          * We only do this when socket buffer fills up.
  79          * Protected by tx vq lock. */
  80         enum vhost_net_poll_state tx_poll_state;
  81         /* Number of TX recently submitted.
  82          * Protected by tx vq lock. */
  83         unsigned tx_packets;
  84         /* Number of times zerocopy TX recently failed.
  85          * Protected by tx vq lock. */
  86         unsigned tx_zcopy_err;
  87         /* Flush in progress. Protected by tx vq lock. */
  88         bool tx_flush;
  89 };
  90
  91 static void vhost_net_tx_packet(struct vhost_net *net)
  92 {
  93         ++net->tx_packets;
  94         if (net->tx_packets < 1024)
  95                 return;
  96         net->tx_packets = 0;
  97         net->tx_zcopy_err = 0;
  98 }
  99
 100 static void vhost_net_tx_err(struct vhost_net *net)
 101 {
 102         ++net->tx_zcopy_err;
 103 }
 104
 105 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 106 {
 107         /* TX flush waits for outstanding DMAs to be done.
 108          * Don't start new DMAs.
 109          */
 110         return !net->tx_flush &&
 111                 net->tx_packets / 64 >= net->tx_zcopy_err;
 112 }
 113
 114 static bool vhost_sock_zcopy(struct socket *sock)
 115 {
 116         return unlikely(experimental_zcopytx) &&
 117                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 118 }
 119
 120 /* Pop first len bytes from iovec. Return number of segments used. */
 121 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 122                           size_t len, int iov_count)
 123 {
 124         int seg = 0;
 125         size_t size;
 126
 127         while (len && seg < iov_count) {
 128                 size = min(from->iov_len, len);
 129                 to->iov_base = from->iov_base;
 130                 to->iov_len = size;
 131                 from->iov_len -= size;
 132                 from->iov_base += size;
 133                 len -= size;
 134                 ++from;
 135                 ++to;
 136                 ++seg;
 137         }
 138         return seg;
 139 }
 140 /* Copy iovec entries for len bytes from iovec. */
 141 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 142                            size_t len, int iovcount)
 143 {
 144         int seg = 0;
 145         size_t size;
 146
 147         while (len && seg < iovcount) {
 148                 size = min(from->iov_len, len);
 149                 to->iov_base = from->iov_base;
 150                 to->iov_len = size;
 151                 len -= size;
 152                 ++from;
 153                 ++to;
 154                 ++seg;
 155         }
 156 }
 157
 158 /* Caller must have TX VQ lock */
 159 static void tx_poll_stop(struct vhost_net *net)
 160 {
 161         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
 162                 return;
 163         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 164         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 165 }
 166
 167 /* Caller must have TX VQ lock */
 168 static int tx_poll_start(struct vhost_net *net, struct socket *sock)
 169 {
 170         int ret;
 171
 172         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 173                 return 0;
 174         ret = vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 175         if (!ret)
 176                 net->tx_poll_state = VHOST_NET_POLL_STARTED;
 177         return ret;
 178 }
 179
 180 /* In case of DMA done not in order in lower device driver for some reason.
 181  * upend_idx is used to track end of used idx, done_idx is used to track head
 182  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 183  * guest used idx.
 184  */
 185 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 186                                       struct vhost_virtqueue *vq)
 187 {
 188         int i;
 189         int j = 0;
 190
 191         for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 192                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 193                         vhost_net_tx_err(net);
 194                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 195                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 196                         vhost_add_used_and_signal(vq->dev, vq,
 197                                                   vq->heads[i].id, 0);
 198                         ++j;
 199                 } else
 200                         break;
 201         }
 202         if (j)
 203                 vq->done_idx = i;
 204         return j;
 205 }
 206
 207 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 208 {
 209         struct vhost_ubuf_ref *ubufs = ubuf->ctx;
 210         struct vhost_virtqueue *vq = ubufs->vq;
 211         int cnt = atomic_read(&ubufs->kref.refcount);
 212
 213         /*
 214          * Trigger polling thread if guest stopped submitting new buffers:
 215          * in this case, the refcount after decrement will eventually reach 1
 216          * so here it is 2.
 217          * We also trigger polling periodically after each 16 packets
 218          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 219          * less than 10% of times).
 220          */
 221         if (cnt <= 2 || !(cnt % 16))
 222                 vhost_poll_queue(&vq->poll);
 223         /* set len to mark this desc buffers done DMA */
 224         vq->heads[ubuf->desc].len = success ?
 225                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 226         vhost_ubuf_put(ubufs);
 227 }
 228
 229 /* Expects to be always run from workqueue - which acts as
 230  * read-size critical section for our kind of RCU. */
 231 static void handle_tx(struct vhost_net *net)
 232 {
 233         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 234         unsigned out, in, s;
 235         int head;
 236         struct msghdr msg = {
 237                 .msg_name = NULL,
 238                 .msg_namelen = 0,
 239                 .msg_control = NULL,
 240                 .msg_controllen = 0,
 241                 .msg_iov = vq->iov,
 242                 .msg_flags = MSG_DONTWAIT,
 243         };
 244         size_t len, total_len = 0;
 245         int err, wmem;
 246         size_t hdr_size;
 247         struct socket *sock;
 248         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 249         bool zcopy, zcopy_used;
 250
 251         /* TODO: check that we are running from vhost_worker? */
 252         sock = rcu_dereference_check(vq->private_data, 1);
 253         if (!sock)
 254                 return;
 255
 256         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 257         if (wmem >= sock->sk->sk_sndbuf) {
 258                 mutex_lock(&vq->mutex);
 259                 tx_poll_start(net, sock);
 260                 mutex_unlock(&vq->mutex);
 261                 return;
 262         }
 263
 264         mutex_lock(&vq->mutex);
 265         vhost_disable_notify(&net->dev, vq);
 266
 267         if (wmem < sock->sk->sk_sndbuf / 2)
 268                 tx_poll_stop(net);
 269         hdr_size = vq->vhost_hlen;
 270         zcopy = vq->ubufs;
 271
 272         for (;;) {
 273                 /* Release DMAs done buffers first */
 274                 if (zcopy)
 275                         vhost_zerocopy_signal_used(net, vq);
 276
 277                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 278                                          ARRAY_SIZE(vq->iov),
 279                                          &out, &in,
 280                                          NULL, NULL);
 281                 /* On error, stop handling until the next kick. */
 282                 if (unlikely(head < 0))
 283                         break;
 284                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 285                 if (head == vq->num) {
 286                         int num_pends;
 287
 288                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 289                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 290                                 tx_poll_start(net, sock);
 291                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 292                                 break;
 293                         }
 294                         /* If more outstanding DMAs, queue the work.
 295                          * Handle upend_idx wrap around
 296                          */
 297                         num_pends = likely(vq->upend_idx >= vq->done_idx) ?
 298                                     (vq->upend_idx - vq->done_idx) :
 299                                     (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
 300                         if (unlikely(num_pends > VHOST_MAX_PEND)) {
 301                                 tx_poll_start(net, sock);
 302                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 303                                 break;
 304                         }
 305                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 306                                 vhost_disable_notify(&net->dev, vq);
 307                                 continue;
 308                         }
 309                         break;
 310                 }
 311                 if (in) {
 312                         vq_err(vq, "Unexpected descriptor format for TX: "
 313                                "out %d, int %d\n", out, in);
 314                         break;
 315                 }
 316                 /* Skip header. TODO: support TSO. */
 317                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 318                 msg.msg_iovlen = out;
 319                 len = iov_length(vq->iov, out);
 320                 /* Sanity check */
 321                 if (!len) {
 322                         vq_err(vq, "Unexpected header len for TX: "
 323                                "%zd expected %zd\n",
 324                                iov_length(vq->hdr, s), hdr_size);
 325                         break;
 326                 }
 327                 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
 328                                        vq->upend_idx != vq->done_idx);
 329
 330                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 331                 if (zcopy_used) {
 332                         vq->heads[vq->upend_idx].id = head;
 333                         if (!vhost_net_tx_select_zcopy(net) ||
 334                             len < VHOST_GOODCOPY_LEN) {
 335                                 /* copy don't need to wait for DMA done */
 336                                 vq->heads[vq->upend_idx].len =
 337                                                         VHOST_DMA_DONE_LEN;
 338                                 msg.msg_control = NULL;
 339                                 msg.msg_controllen = 0;
 340                                 ubufs = NULL;
 341                         } else {
 342                                 struct ubuf_info *ubuf = &vq->ubuf_info[head];
 343
 344                                 vq->heads[vq->upend_idx].len =
 345                                         VHOST_DMA_IN_PROGRESS;
 346                                 ubuf->callback = vhost_zerocopy_callback;
 347                                 ubuf->ctx = vq->ubufs;
 348                                 ubuf->desc = vq->upend_idx;
 349                                 msg.msg_control = ubuf;
 350                                 msg.msg_controllen = sizeof(ubuf);
 351                                 ubufs = vq->ubufs;
 352                                 kref_get(&ubufs->kref);
 353                         }
 354                         vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
 355                 }
 356                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 357                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 358                 if (unlikely(err < 0)) {
 359                         if (zcopy_used) {
 360                                 if (ubufs)
 361                                         vhost_ubuf_put(ubufs);
 362                                 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
 363                                         UIO_MAXIOV;
 364                         }
 365                         vhost_discard_vq_desc(vq, 1);
 366                         if (err == -EAGAIN || err == -ENOBUFS)
 367                                 tx_poll_start(net, sock);
 368                         break;
 369                 }
 370                 if (err != len)
 371                         pr_debug("Truncated TX packet: "
 372                                  " len %d != %zd\n", err, len);
 373                 if (!zcopy_used)
 374                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 375                 else
 376                         vhost_zerocopy_signal_used(net, vq);
 377                 total_len += len;
 378                 vhost_net_tx_packet(net);
 379                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 380                         vhost_poll_queue(&vq->poll);
 381                         break;
 382                 }
 383         }
 384
 385         mutex_unlock(&vq->mutex);
 386 }
 387
 388 static int peek_head_len(struct sock *sk)
 389 {
 390         struct sk_buff *head;
 391         int len = 0;
 392         unsigned long flags;
 393
 394         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 395         head = skb_peek(&sk->sk_receive_queue);
 396         if (likely(head)) {
 397                 len = head->len;
 398                 if (vlan_tx_tag_present(head))
 399                         len += VLAN_HLEN;
 400         }
 401
 402         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 403         return len;
 404 }
 405
 406 /* This is a multi-buffer version of vhost_get_desc, that works if
 407  *      vq has read descriptors only.
 408  * @vq          - the relevant virtqueue
 409  * @datalen     - data length we'll be reading
 410  * @iovcount    - returned count of io vectors we fill
 411  * @log         - vhost log
 412  * @log_num     - log offset
 413  * @quota       - headcount quota, 1 for big buffer
 414  *      returns number of buffer heads allocated, negative on error
 415  */
 416 static int get_rx_bufs(struct vhost_virtqueue *vq,
 417                        struct vring_used_elem *heads,
 418                        int datalen,
 419                        unsigned *iovcount,
 420                        struct vhost_log *log,
 421                        unsigned *log_num,
 422                        unsigned int quota)
 423 {
 424         unsigned int out, in;
 425         int seg = 0;
 426         int headcount = 0;
 427         unsigned d;
 428         int r, nlogs = 0;
 429
 430         while (datalen > 0 && headcount < quota) {
 431                 if (unlikely(seg >= UIO_MAXIOV)) {
 432                         r = -ENOBUFS;
 433                         goto err;
 434                 }
 435                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 436                                       ARRAY_SIZE(vq->iov) - seg, &out,
 437                                       &in, log, log_num);
 438                 if (d == vq->num) {
 439                         r = 0;
 440                         goto err;
 441                 }
 442                 if (unlikely(out || in <= 0)) {
 443                         vq_err(vq, "unexpected descriptor format for RX: "
 444                                 "out %d, in %d\n", out, in);
 445                         r = -EINVAL;
 446                         goto err;
 447                 }
 448                 if (unlikely(log)) {
 449                         nlogs += *log_num;
 450                         log += *log_num;
 451                 }
 452                 heads[headcount].id = d;
 453                 heads[headcount].len = iov_length(vq->iov + seg, in);
 454                 datalen -= heads[headcount].len;
 455                 ++headcount;
 456                 seg += in;
 457         }
 458         heads[headcount - 1].len += datalen;
 459         *iovcount = seg;
 460         if (unlikely(log))
 461                 *log_num = nlogs;
 462         return headcount;
 463 err:
 464         vhost_discard_vq_desc(vq, headcount);
 465         return r;
 466 }
 467
 468 /* Expects to be always run from workqueue - which acts as
 469  * read-size critical section for our kind of RCU. */
 470 static void handle_rx(struct vhost_net *net)
 471 {
 472         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 473         unsigned uninitialized_var(in), log;
 474         struct vhost_log *vq_log;
 475         struct msghdr msg = {
 476                 .msg_name = NULL,
 477                 .msg_namelen = 0,
 478                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 479                 .msg_controllen = 0,
 480                 .msg_iov = vq->iov,
 481                 .msg_flags = MSG_DONTWAIT,
 482         };
 483         struct virtio_net_hdr_mrg_rxbuf hdr = {
 484                 .hdr.flags = 0,
 485                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 486         };
 487         size_t total_len = 0;
 488         int err, mergeable;
 489         s16 headcount;
 490         size_t vhost_hlen, sock_hlen;
 491         size_t vhost_len, sock_len;
 492         /* TODO: check that we are running from vhost_worker? */
 493         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 494
 495         if (!sock)
 496                 return;
 497
 498         mutex_lock(&vq->mutex);
 499         vhost_disable_notify(&net->dev, vq);
 500         vhost_hlen = vq->vhost_hlen;
 501         sock_hlen = vq->sock_hlen;
 502
 503         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 504                 vq->log : NULL;
 505         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 506
 507         while ((sock_len = peek_head_len(sock->sk))) {
 508                 sock_len += sock_hlen;
 509                 vhost_len = sock_len + vhost_hlen;
 510                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 511                                         &in, vq_log, &log,
 512                                         likely(mergeable) ? UIO_MAXIOV : 1);
 513                 /* On error, stop handling until the next kick. */
 514                 if (unlikely(headcount < 0))
 515                         break;
 516                 /* OK, now we need to know about added descriptors. */
 517                 if (!headcount) {
 518                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 519                                 /* They have slipped one in as we were
 520                                  * doing that: check again. */
 521                                 vhost_disable_notify(&net->dev, vq);
 522                                 continue;
 523                         }
 524                         /* Nothing new?  Wait for eventfd to tell us
 525                          * they refilled. */
 526                         break;
 527                 }
 528                 /* We don't need to be notified again. */
 529                 if (unlikely((vhost_hlen)))
 530                         /* Skip header. TODO: support TSO. */
 531                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 532                 else
 533                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 534                          * needed because recvmsg can modify msg_iov. */
 535                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 536                 msg.msg_iovlen = in;
 537                 err = sock->ops->recvmsg(NULL, sock, &msg,
 538                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 539                 /* Userspace might have consumed the packet meanwhile:
 540                  * it's not supposed to do this usually, but might be hard
 541                  * to prevent. Discard data we got (if any) and keep going. */
 542                 if (unlikely(err != sock_len)) {
 543                         pr_debug("Discarded rx packet: "
 544                                  " len %d, expected %zd\n", err, sock_len);
 545                         vhost_discard_vq_desc(vq, headcount);
 546                         continue;
 547                 }
 548                 if (unlikely(vhost_hlen) &&
 549                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 550                                       vhost_hlen)) {
 551                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 552                                vq->iov->iov_base);
 553                         break;
 554                 }
 555                 /* TODO: Should check and handle checksum. */
 556                 if (likely(mergeable) &&
 557                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 558                                       offsetof(typeof(hdr), num_buffers),
 559                                       sizeof hdr.num_buffers)) {
 560                         vq_err(vq, "Failed num_buffers write");
 561                         vhost_discard_vq_desc(vq, headcount);
 562                         break;
 563                 }
 564                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 565                                             headcount);
 566                 if (unlikely(vq_log))
 567                         vhost_log_write(vq, vq_log, log, vhost_len);
 568                 total_len += vhost_len;
 569                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 570                         vhost_poll_queue(&vq->poll);
 571                         break;
 572                 }
 573         }
 574
 575         mutex_unlock(&vq->mutex);
 576 }
 577
 578 static void handle_tx_kick(struct vhost_work *work)
 579 {
 580         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 581                                                   poll.work);
 582         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 583
 584         handle_tx(net);
 585 }
 586
 587 static void handle_rx_kick(struct vhost_work *work)
 588 {
 589         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 590                                                   poll.work);
 591         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 592
 593         handle_rx(net);
 594 }
 595
 596 static void handle_tx_net(struct vhost_work *work)
 597 {
 598         struct vhost_net *net = container_of(work, struct vhost_net,
 599                                              poll[VHOST_NET_VQ_TX].work);
 600         handle_tx(net);
 601 }
 602
 603 static void handle_rx_net(struct vhost_work *work)
 604 {
 605         struct vhost_net *net = container_of(work, struct vhost_net,
 606                                              poll[VHOST_NET_VQ_RX].work);
 607         handle_rx(net);
 608 }
 609
 610 static int vhost_net_open(struct inode *inode, struct file *f)
 611 {
 612         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 613         struct vhost_dev *dev;
 614         int r;
 615
 616         if (!n)
 617                 return -ENOMEM;
 618
 619         dev = &n->dev;
 620         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 621         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 622         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 623         if (r < 0) {
 624                 kfree(n);
 625                 return r;
 626         }
 627
 628         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 629         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 630         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 631
 632         f->private_data = n;
 633
 634         return 0;
 635 }
 636
 637 static void vhost_net_disable_vq(struct vhost_net *n,
 638                                  struct vhost_virtqueue *vq)
 639 {
 640         if (!vq->private_data)
 641                 return;
 642         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 643                 tx_poll_stop(n);
 644                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 645         } else
 646                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 647 }
 648
 649 static int vhost_net_enable_vq(struct vhost_net *n,
 650                                 struct vhost_virtqueue *vq)
 651 {
 652         struct socket *sock;
 653         int ret;
 654
 655         sock = rcu_dereference_protected(vq->private_data,
 656                                          lockdep_is_held(&vq->mutex));
 657         if (!sock)
 658                 return 0;
 659         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 660                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 661                 ret = tx_poll_start(n, sock);
 662         } else
 663                 ret = vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 664
 665         return ret;
 666 }
 667
 668 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 669                                         struct vhost_virtqueue *vq)
 670 {
 671         struct socket *sock;
 672
 673         mutex_lock(&vq->mutex);
 674         sock = rcu_dereference_protected(vq->private_data,
 675                                          lockdep_is_held(&vq->mutex));
 676         vhost_net_disable_vq(n, vq);
 677         rcu_assign_pointer(vq->private_data, NULL);
 678         mutex_unlock(&vq->mutex);
 679         return sock;
 680 }
 681
 682 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 683                            struct socket **rx_sock)
 684 {
 685         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 686         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 687 }
 688
 689 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 690 {
 691         vhost_poll_flush(n->poll + index);
 692         vhost_poll_flush(&n->dev.vqs[index].poll);
 693 }
 694
 695 static void vhost_net_flush(struct vhost_net *n)
 696 {
 697         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 698         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 699         if (n->dev.vqs[VHOST_NET_VQ_TX].ubufs) {
 700                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 701                 n->tx_flush = true;
 702                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 703                 /* Wait for all lower device DMAs done. */
 704                 vhost_ubuf_put_and_wait(n->dev.vqs[VHOST_NET_VQ_TX].ubufs);
 705                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 706                 n->tx_flush = false;
 707                 kref_init(&n->dev.vqs[VHOST_NET_VQ_TX].ubufs->kref);
 708                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 709         }
 710 }
 711
 712 static int vhost_net_release(struct inode *inode, struct file *f)
 713 {
 714         struct vhost_net *n = f->private_data;
 715         struct socket *tx_sock;
 716         struct socket *rx_sock;
 717
 718         vhost_net_stop(n, &tx_sock, &rx_sock);
 719         vhost_net_flush(n);
 720         vhost_dev_stop(&n->dev);
 721         vhost_dev_cleanup(&n->dev, false);
 722         if (tx_sock)
 723                 fput(tx_sock->file);
 724         if (rx_sock)
 725                 fput(rx_sock->file);
 726         /* We do an extra flush before freeing memory,
 727          * since jobs can re-queue themselves. */
 728         vhost_net_flush(n);
 729         kfree(n);
 730         return 0;
 731 }
 732
 733 static struct socket *get_raw_socket(int fd)
 734 {
 735         struct {
 736                 struct sockaddr_ll sa;
 737                 char  buf[MAX_ADDR_LEN];
 738         } uaddr;
 739         int uaddr_len = sizeof uaddr, r;
 740         struct socket *sock = sockfd_lookup(fd, &r);
 741
 742         if (!sock)
 743                 return ERR_PTR(-ENOTSOCK);
 744
 745         /* Parameter checking */
 746         if (sock->sk->sk_type != SOCK_RAW) {
 747                 r = -ESOCKTNOSUPPORT;
 748                 goto err;
 749         }
 750
 751         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 752                                &uaddr_len, 0);
 753         if (r)
 754                 goto err;
 755
 756         if (uaddr.sa.sll_family != AF_PACKET) {
 757                 r = -EPFNOSUPPORT;
 758                 goto err;
 759         }
 760         return sock;
 761 err:
 762         fput(sock->file);
 763         return ERR_PTR(r);
 764 }
 765
 766 static struct socket *get_tap_socket(int fd)
 767 {
 768         struct file *file = fget(fd);
 769         struct socket *sock;
 770
 771         if (!file)
 772                 return ERR_PTR(-EBADF);
 773         sock = tun_get_socket(file);
 774         if (!IS_ERR(sock))
 775                 return sock;
 776         sock = macvtap_get_socket(file);
 777         if (IS_ERR(sock))
 778                 fput(file);
 779         return sock;
 780 }
 781
 782 static struct socket *get_socket(int fd)
 783 {
 784         struct socket *sock;
 785
 786         /* special case to disable backend */
 787         if (fd == -1)
 788                 return NULL;
 789         sock = get_raw_socket(fd);
 790         if (!IS_ERR(sock))
 791                 return sock;
 792         sock = get_tap_socket(fd);
 793         if (!IS_ERR(sock))
 794                 return sock;
 795         return ERR_PTR(-ENOTSOCK);
 796 }
 797
 798 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 799 {
 800         struct socket *sock, *oldsock;
 801         struct vhost_virtqueue *vq;
 802         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 803         int r;
 804
 805         mutex_lock(&n->dev.mutex);
 806         r = vhost_dev_check_owner(&n->dev);
 807         if (r)
 808                 goto err;
 809
 810         if (index >= VHOST_NET_VQ_MAX) {
 811                 r = -ENOBUFS;
 812                 goto err;
 813         }
 814         vq = n->vqs + index;
 815         mutex_lock(&vq->mutex);
 816
 817         /* Verify that ring has been setup correctly. */
 818         if (!vhost_vq_access_ok(vq)) {
 819                 r = -EFAULT;
 820                 goto err_vq;
 821         }
 822         sock = get_socket(fd);
 823         if (IS_ERR(sock)) {
 824                 r = PTR_ERR(sock);
 825                 goto err_vq;
 826         }
 827
 828         /* start polling new socket */
 829         oldsock = rcu_dereference_protected(vq->private_data,
 830                                             lockdep_is_held(&vq->mutex));
 831         if (sock != oldsock) {
 832                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 833                 if (IS_ERR(ubufs)) {
 834                         r = PTR_ERR(ubufs);
 835                         goto err_ubufs;
 836                 }
 837
 838                 vhost_net_disable_vq(n, vq);
 839                 rcu_assign_pointer(vq->private_data, sock);
 840                 r = vhost_init_used(vq);
 841                 if (r)
 842                         goto err_used;
 843                 r = vhost_net_enable_vq(n, vq);
 844                 if (r)
 845                         goto err_used;
 846
 847                 oldubufs = vq->ubufs;
 848                 vq->ubufs = ubufs;
 849
 850                 n->tx_packets = 0;
 851                 n->tx_zcopy_err = 0;
 852                 n->tx_flush = false;
 853         }
 854
 855         mutex_unlock(&vq->mutex);
 856
 857         if (oldubufs) {
 858                 vhost_ubuf_put_and_wait(oldubufs);
 859                 mutex_lock(&vq->mutex);
 860                 vhost_zerocopy_signal_used(n, vq);
 861                 mutex_unlock(&vq->mutex);
 862         }
 863
 864         if (oldsock) {
 865                 vhost_net_flush_vq(n, index);
 866                 fput(oldsock->file);
 867         }
 868
 869         mutex_unlock(&n->dev.mutex);
 870         return 0;
 871
 872 err_used:
 873         rcu_assign_pointer(vq->private_data, oldsock);
 874         vhost_net_enable_vq(n, vq);
 875         if (ubufs)
 876                 vhost_ubuf_put_and_wait(ubufs);
 877 err_ubufs:
 878         fput(sock->file);
 879 err_vq:
 880         mutex_unlock(&vq->mutex);
 881 err:
 882         mutex_unlock(&n->dev.mutex);
 883         return r;
 884 }
 885
 886 static long vhost_net_reset_owner(struct vhost_net *n)
 887 {
 888         struct socket *tx_sock = NULL;
 889         struct socket *rx_sock = NULL;
 890         long err;
 891
 892         mutex_lock(&n->dev.mutex);
 893         err = vhost_dev_check_owner(&n->dev);
 894         if (err)
 895                 goto done;
 896         vhost_net_stop(n, &tx_sock, &rx_sock);
 897         vhost_net_flush(n);
 898         err = vhost_dev_reset_owner(&n->dev);
 899 done:
 900         mutex_unlock(&n->dev.mutex);
 901         if (tx_sock)
 902                 fput(tx_sock->file);
 903         if (rx_sock)
 904                 fput(rx_sock->file);
 905         return err;
 906 }
 907
 908 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 909 {
 910         size_t vhost_hlen, sock_hlen, hdr_len;
 911         int i;
 912
 913         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 914                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 915                         sizeof(struct virtio_net_hdr);
 916         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 917                 /* vhost provides vnet_hdr */
 918                 vhost_hlen = hdr_len;
 919                 sock_hlen = 0;
 920         } else {
 921                 /* socket provides vnet_hdr */
 922                 vhost_hlen = 0;
 923                 sock_hlen = hdr_len;
 924         }
 925         mutex_lock(&n->dev.mutex);
 926         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 927             !vhost_log_access_ok(&n->dev)) {
 928                 mutex_unlock(&n->dev.mutex);
 929                 return -EFAULT;
 930         }
 931         n->dev.acked_features = features;
 932         smp_wmb();
 933         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 934                 mutex_lock(&n->vqs[i].mutex);
 935                 n->vqs[i].vhost_hlen = vhost_hlen;
 936                 n->vqs[i].sock_hlen = sock_hlen;
 937                 mutex_unlock(&n->vqs[i].mutex);
 938         }
 939         vhost_net_flush(n);
 940         mutex_unlock(&n->dev.mutex);
 941         return 0;
 942 }
 943
 944 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 945                             unsigned long arg)
 946 {
 947         struct vhost_net *n = f->private_data;
 948         void __user *argp = (void __user *)arg;
 949         u64 __user *featurep = argp;
 950         struct vhost_vring_file backend;
 951         u64 features;
 952         int r;
 953
 954         switch (ioctl) {
 955         case VHOST_NET_SET_BACKEND:
 956                 if (copy_from_user(&backend, argp, sizeof backend))
 957                         return -EFAULT;
 958                 return vhost_net_set_backend(n, backend.index, backend.fd);
 959         case VHOST_GET_FEATURES:
 960                 features = VHOST_NET_FEATURES;
 961                 if (copy_to_user(featurep, &features, sizeof features))
 962                         return -EFAULT;
 963                 return 0;
 964         case VHOST_SET_FEATURES:
 965                 if (copy_from_user(&features, featurep, sizeof features))
 966                         return -EFAULT;
 967                 if (features & ~VHOST_NET_FEATURES)
 968                         return -EOPNOTSUPP;
 969                 return vhost_net_set_features(n, features);
 970         case VHOST_RESET_OWNER:
 971                 return vhost_net_reset_owner(n);
 972         default:
 973                 mutex_lock(&n->dev.mutex);
 974                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
 975                 if (r == -ENOIOCTLCMD)
 976                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
 977                 else
 978                         vhost_net_flush(n);
 979                 mutex_unlock(&n->dev.mutex);
 980                 return r;
 981         }
 982 }
 983
 984 #ifdef CONFIG_COMPAT
 985 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 986                                    unsigned long arg)
 987 {
 988         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 989 }
 990 #endif
 991
 992 static const struct file_operations vhost_net_fops = {
 993         .owner          = THIS_MODULE,
 994         .release        = vhost_net_release,
 995         .unlocked_ioctl = vhost_net_ioctl,
 996 #ifdef CONFIG_COMPAT
 997         .compat_ioctl   = vhost_net_compat_ioctl,
 998 #endif
 999         .open           = vhost_net_open,
1000         .llseek         = noop_llseek,
1001 };
1002
1003 static struct miscdevice vhost_net_misc = {
1004         .minor = VHOST_NET_MINOR,
1005         .name = "vhost-net",
1006         .fops = &vhost_net_fops,
1007 };
1008
1009 static int vhost_net_init(void)
1010 {
1011         if (experimental_zcopytx)
1012                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
1013         return misc_register(&vhost_net_misc);
1014 }
1015 module_init(vhost_net_init);
1016
1017 static void vhost_net_exit(void)
1018 {
1019         misc_deregister(&vhost_net_misc);
1020 }
1021 module_exit(vhost_net_exit);
1022
1023 MODULE_VERSION("0.0.1");
1024 MODULE_LICENSE("GPL v2");
1025 MODULE_AUTHOR("Michael S. Tsirkin");
1026 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1027 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1028 MODULE_ALIAS("devname:vhost-net");