drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/rcupdate.h>
  19 #include <linux/file.h>
  20 #include <linux/slab.h>
  21
  22 #include <linux/net.h>
  23 #include <linux/if_packet.h>
  24 #include <linux/if_arp.h>
  25 #include <linux/if_tun.h>
  26 #include <linux/if_macvlan.h>
  27 #include <linux/if_vlan.h>
  28
  29 #include <net/sock.h>
  30
  31 #include "vhost.h"
  32
  33 static int experimental_zcopytx;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Experimental Zero Copy TX");
  36
  37 /* Max number of bytes transferred before requeueing the job.
  38  * Using this limit prevents one virtqueue from starving others. */
  39 #define VHOST_NET_WEIGHT 0x80000
  40
  41 /* MAX number of TX used buffers for outstanding zerocopy */
  42 #define VHOST_MAX_PEND 128
  43 #define VHOST_GOODCOPY_LEN 256
  44
  45 /*
  46  * For transmit, used buffer len is unused; we override it to track buffer
  47  * status internally; used for zerocopy tx only.
  48  */
  49 /* Lower device DMA failed */
  50 #define VHOST_DMA_FAILED_LEN    3
  51 /* Lower device DMA done */
  52 #define VHOST_DMA_DONE_LEN      2
  53 /* Lower device DMA in progress */
  54 #define VHOST_DMA_IN_PROGRESS   1
  55 /* Buffer unused */
  56 #define VHOST_DMA_CLEAR_LEN     0
  57
  58 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
  59
  60 enum {
  61         VHOST_NET_VQ_RX = 0,
  62         VHOST_NET_VQ_TX = 1,
  63         VHOST_NET_VQ_MAX = 2,
  64 };
  65
  66 enum vhost_net_poll_state {
  67         VHOST_NET_POLL_DISABLED = 0,
  68         VHOST_NET_POLL_STARTED = 1,
  69         VHOST_NET_POLL_STOPPED = 2,
  70 };
  71
  72 struct vhost_net {
  73         struct vhost_dev dev;
  74         struct vhost_virtqueue vqs[VHOST_NET_VQ_MAX];
  75         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  76         /* Tells us whether we are polling a socket for TX.
  77          * We only do this when socket buffer fills up.
  78          * Protected by tx vq lock. */
  79         enum vhost_net_poll_state tx_poll_state;
  80         /* Number of TX recently submitted.
  81          * Protected by tx vq lock. */
  82         unsigned tx_packets;
  83         /* Number of times zerocopy TX recently failed.
  84          * Protected by tx vq lock. */
  85         unsigned tx_zcopy_err;
  86         /* Flush in progress. Protected by tx vq lock. */
  87         bool tx_flush;
  88 };
  89
  90 static void vhost_net_tx_packet(struct vhost_net *net)
  91 {
  92         ++net->tx_packets;
  93         if (net->tx_packets < 1024)
  94                 return;
  95         net->tx_packets = 0;
  96         net->tx_zcopy_err = 0;
  97 }
  98
  99 static void vhost_net_tx_err(struct vhost_net *net)
 100 {
 101         ++net->tx_zcopy_err;
 102 }
 103
 104 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 105 {
 106         /* TX flush waits for outstanding DMAs to be done.
 107          * Don't start new DMAs.
 108          */
 109         return !net->tx_flush &&
 110                 net->tx_packets / 64 >= net->tx_zcopy_err;
 111 }
 112
 113 static bool vhost_sock_zcopy(struct socket *sock)
 114 {
 115         return unlikely(experimental_zcopytx) &&
 116                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 117 }
 118
 119 /* Pop first len bytes from iovec. Return number of segments used. */
 120 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 121                           size_t len, int iov_count)
 122 {
 123         int seg = 0;
 124         size_t size;
 125
 126         while (len && seg < iov_count) {
 127                 size = min(from->iov_len, len);
 128                 to->iov_base = from->iov_base;
 129                 to->iov_len = size;
 130                 from->iov_len -= size;
 131                 from->iov_base += size;
 132                 len -= size;
 133                 ++from;
 134                 ++to;
 135                 ++seg;
 136         }
 137         return seg;
 138 }
 139 /* Copy iovec entries for len bytes from iovec. */
 140 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 141                            size_t len, int iovcount)
 142 {
 143         int seg = 0;
 144         size_t size;
 145
 146         while (len && seg < iovcount) {
 147                 size = min(from->iov_len, len);
 148                 to->iov_base = from->iov_base;
 149                 to->iov_len = size;
 150                 len -= size;
 151                 ++from;
 152                 ++to;
 153                 ++seg;
 154         }
 155 }
 156
 157 /* Caller must have TX VQ lock */
 158 static void tx_poll_stop(struct vhost_net *net)
 159 {
 160         if (likely(net->tx_poll_state != VHOST_NET_POLL_STARTED))
 161                 return;
 162         vhost_poll_stop(net->poll + VHOST_NET_VQ_TX);
 163         net->tx_poll_state = VHOST_NET_POLL_STOPPED;
 164 }
 165
 166 /* Caller must have TX VQ lock */
 167 static void tx_poll_start(struct vhost_net *net, struct socket *sock)
 168 {
 169         if (unlikely(net->tx_poll_state != VHOST_NET_POLL_STOPPED))
 170                 return;
 171         vhost_poll_start(net->poll + VHOST_NET_VQ_TX, sock->file);
 172         net->tx_poll_state = VHOST_NET_POLL_STARTED;
 173 }
 174
 175 /* In case of DMA done not in order in lower device driver for some reason.
 176  * upend_idx is used to track end of used idx, done_idx is used to track head
 177  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 178  * guest used idx.
 179  */
 180 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 181                                       struct vhost_virtqueue *vq)
 182 {
 183         int i;
 184         int j = 0;
 185
 186         for (i = vq->done_idx; i != vq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 187                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 188                         vhost_net_tx_err(net);
 189                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 190                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 191                         vhost_add_used_and_signal(vq->dev, vq,
 192                                                   vq->heads[i].id, 0);
 193                         ++j;
 194                 } else
 195                         break;
 196         }
 197         if (j)
 198                 vq->done_idx = i;
 199         return j;
 200 }
 201
 202 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 203 {
 204         struct vhost_ubuf_ref *ubufs = ubuf->ctx;
 205         struct vhost_virtqueue *vq = ubufs->vq;
 206         int cnt = atomic_read(&ubufs->kref.refcount);
 207
 208         /*
 209          * Trigger polling thread if guest stopped submitting new buffers:
 210          * in this case, the refcount after decrement will eventually reach 1
 211          * so here it is 2.
 212          * We also trigger polling periodically after each 16 packets
 213          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 214          * less than 10% of times).
 215          */
 216         if (cnt <= 2 || !(cnt % 16))
 217                 vhost_poll_queue(&vq->poll);
 218         /* set len to mark this desc buffers done DMA */
 219         vq->heads[ubuf->desc].len = success ?
 220                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 221         vhost_ubuf_put(ubufs);
 222 }
 223
 224 /* Expects to be always run from workqueue - which acts as
 225  * read-size critical section for our kind of RCU. */
 226 static void handle_tx(struct vhost_net *net)
 227 {
 228         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_TX];
 229         unsigned out, in, s;
 230         int head;
 231         struct msghdr msg = {
 232                 .msg_name = NULL,
 233                 .msg_namelen = 0,
 234                 .msg_control = NULL,
 235                 .msg_controllen = 0,
 236                 .msg_iov = vq->iov,
 237                 .msg_flags = MSG_DONTWAIT,
 238         };
 239         size_t len, total_len = 0;
 240         int err, wmem;
 241         size_t hdr_size;
 242         struct socket *sock;
 243         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 244         bool zcopy, zcopy_used;
 245
 246         /* TODO: check that we are running from vhost_worker? */
 247         sock = rcu_dereference_check(vq->private_data, 1);
 248         if (!sock)
 249                 return;
 250
 251         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 252         if (wmem >= sock->sk->sk_sndbuf) {
 253                 mutex_lock(&vq->mutex);
 254                 tx_poll_start(net, sock);
 255                 mutex_unlock(&vq->mutex);
 256                 return;
 257         }
 258
 259         mutex_lock(&vq->mutex);
 260         vhost_disable_notify(&net->dev, vq);
 261
 262         if (wmem < sock->sk->sk_sndbuf / 2)
 263                 tx_poll_stop(net);
 264         hdr_size = vq->vhost_hlen;
 265         zcopy = vq->ubufs;
 266
 267         for (;;) {
 268                 /* Release DMAs done buffers first */
 269                 if (zcopy)
 270                         vhost_zerocopy_signal_used(net, vq);
 271
 272                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 273                                          ARRAY_SIZE(vq->iov),
 274                                          &out, &in,
 275                                          NULL, NULL);
 276                 /* On error, stop handling until the next kick. */
 277                 if (unlikely(head < 0))
 278                         break;
 279                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 280                 if (head == vq->num) {
 281                         int num_pends;
 282
 283                         wmem = atomic_read(&sock->sk->sk_wmem_alloc);
 284                         if (wmem >= sock->sk->sk_sndbuf * 3 / 4) {
 285                                 tx_poll_start(net, sock);
 286                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 287                                 break;
 288                         }
 289                         /* If more outstanding DMAs, queue the work.
 290                          * Handle upend_idx wrap around
 291                          */
 292                         num_pends = likely(vq->upend_idx >= vq->done_idx) ?
 293                                     (vq->upend_idx - vq->done_idx) :
 294                                     (vq->upend_idx + UIO_MAXIOV - vq->done_idx);
 295                         if (unlikely(num_pends > VHOST_MAX_PEND)) {
 296                                 tx_poll_start(net, sock);
 297                                 set_bit(SOCK_ASYNC_NOSPACE, &sock->flags);
 298                                 break;
 299                         }
 300                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 301                                 vhost_disable_notify(&net->dev, vq);
 302                                 continue;
 303                         }
 304                         break;
 305                 }
 306                 if (in) {
 307                         vq_err(vq, "Unexpected descriptor format for TX: "
 308                                "out %d, int %d\n", out, in);
 309                         break;
 310                 }
 311                 /* Skip header. TODO: support TSO. */
 312                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 313                 msg.msg_iovlen = out;
 314                 len = iov_length(vq->iov, out);
 315                 /* Sanity check */
 316                 if (!len) {
 317                         vq_err(vq, "Unexpected header len for TX: "
 318                                "%zd expected %zd\n",
 319                                iov_length(vq->hdr, s), hdr_size);
 320                         break;
 321                 }
 322                 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
 323                                        vq->upend_idx != vq->done_idx);
 324
 325                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 326                 if (zcopy_used) {
 327                         vq->heads[vq->upend_idx].id = head;
 328                         if (!vhost_net_tx_select_zcopy(net) ||
 329                             len < VHOST_GOODCOPY_LEN) {
 330                                 /* copy don't need to wait for DMA done */
 331                                 vq->heads[vq->upend_idx].len =
 332                                                         VHOST_DMA_DONE_LEN;
 333                                 msg.msg_control = NULL;
 334                                 msg.msg_controllen = 0;
 335                                 ubufs = NULL;
 336                         } else {
 337                                 struct ubuf_info *ubuf = &vq->ubuf_info[head];
 338
 339                                 vq->heads[vq->upend_idx].len =
 340                                         VHOST_DMA_IN_PROGRESS;
 341                                 ubuf->callback = vhost_zerocopy_callback;
 342                                 ubuf->ctx = vq->ubufs;
 343                                 ubuf->desc = vq->upend_idx;
 344                                 msg.msg_control = ubuf;
 345                                 msg.msg_controllen = sizeof(ubuf);
 346                                 ubufs = vq->ubufs;
 347                                 kref_get(&ubufs->kref);
 348                         }
 349                         vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
 350                 }
 351                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 352                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 353                 if (unlikely(err < 0)) {
 354                         if (zcopy_used) {
 355                                 if (ubufs)
 356                                         vhost_ubuf_put(ubufs);
 357                                 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
 358                                         UIO_MAXIOV;
 359                         }
 360                         vhost_discard_vq_desc(vq, 1);
 361                         if (err == -EAGAIN || err == -ENOBUFS)
 362                                 tx_poll_start(net, sock);
 363                         break;
 364                 }
 365                 if (err != len)
 366                         pr_debug("Truncated TX packet: "
 367                                  " len %d != %zd\n", err, len);
 368                 if (!zcopy_used)
 369                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 370                 else
 371                         vhost_zerocopy_signal_used(net, vq);
 372                 total_len += len;
 373                 vhost_net_tx_packet(net);
 374                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 375                         vhost_poll_queue(&vq->poll);
 376                         break;
 377                 }
 378         }
 379
 380         mutex_unlock(&vq->mutex);
 381 }
 382
 383 static int peek_head_len(struct sock *sk)
 384 {
 385         struct sk_buff *head;
 386         int len = 0;
 387         unsigned long flags;
 388
 389         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 390         head = skb_peek(&sk->sk_receive_queue);
 391         if (likely(head)) {
 392                 len = head->len;
 393                 if (vlan_tx_tag_present(head))
 394                         len += VLAN_HLEN;
 395         }
 396
 397         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 398         return len;
 399 }
 400
 401 /* This is a multi-buffer version of vhost_get_desc, that works if
 402  *      vq has read descriptors only.
 403  * @vq          - the relevant virtqueue
 404  * @datalen     - data length we'll be reading
 405  * @iovcount    - returned count of io vectors we fill
 406  * @log         - vhost log
 407  * @log_num     - log offset
 408  * @quota       - headcount quota, 1 for big buffer
 409  *      returns number of buffer heads allocated, negative on error
 410  */
 411 static int get_rx_bufs(struct vhost_virtqueue *vq,
 412                        struct vring_used_elem *heads,
 413                        int datalen,
 414                        unsigned *iovcount,
 415                        struct vhost_log *log,
 416                        unsigned *log_num,
 417                        unsigned int quota)
 418 {
 419         unsigned int out, in;
 420         int seg = 0;
 421         int headcount = 0;
 422         unsigned d;
 423         int r, nlogs = 0;
 424
 425         while (datalen > 0 && headcount < quota) {
 426                 if (unlikely(seg >= UIO_MAXIOV)) {
 427                         r = -ENOBUFS;
 428                         goto err;
 429                 }
 430                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 431                                       ARRAY_SIZE(vq->iov) - seg, &out,
 432                                       &in, log, log_num);
 433                 if (d == vq->num) {
 434                         r = 0;
 435                         goto err;
 436                 }
 437                 if (unlikely(out || in <= 0)) {
 438                         vq_err(vq, "unexpected descriptor format for RX: "
 439                                 "out %d, in %d\n", out, in);
 440                         r = -EINVAL;
 441                         goto err;
 442                 }
 443                 if (unlikely(log)) {
 444                         nlogs += *log_num;
 445                         log += *log_num;
 446                 }
 447                 heads[headcount].id = d;
 448                 heads[headcount].len = iov_length(vq->iov + seg, in);
 449                 datalen -= heads[headcount].len;
 450                 ++headcount;
 451                 seg += in;
 452         }
 453         heads[headcount - 1].len += datalen;
 454         *iovcount = seg;
 455         if (unlikely(log))
 456                 *log_num = nlogs;
 457         return headcount;
 458 err:
 459         vhost_discard_vq_desc(vq, headcount);
 460         return r;
 461 }
 462
 463 /* Expects to be always run from workqueue - which acts as
 464  * read-size critical section for our kind of RCU. */
 465 static void handle_rx(struct vhost_net *net)
 466 {
 467         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 468         unsigned uninitialized_var(in), log;
 469         struct vhost_log *vq_log;
 470         struct msghdr msg = {
 471                 .msg_name = NULL,
 472                 .msg_namelen = 0,
 473                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 474                 .msg_controllen = 0,
 475                 .msg_iov = vq->iov,
 476                 .msg_flags = MSG_DONTWAIT,
 477         };
 478         struct virtio_net_hdr_mrg_rxbuf hdr = {
 479                 .hdr.flags = 0,
 480                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 481         };
 482         size_t total_len = 0;
 483         int err, mergeable;
 484         s16 headcount;
 485         size_t vhost_hlen, sock_hlen;
 486         size_t vhost_len, sock_len;
 487         /* TODO: check that we are running from vhost_worker? */
 488         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 489
 490         if (!sock)
 491                 return;
 492
 493         mutex_lock(&vq->mutex);
 494         vhost_disable_notify(&net->dev, vq);
 495         vhost_hlen = vq->vhost_hlen;
 496         sock_hlen = vq->sock_hlen;
 497
 498         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 499                 vq->log : NULL;
 500         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 501
 502         while ((sock_len = peek_head_len(sock->sk))) {
 503                 sock_len += sock_hlen;
 504                 vhost_len = sock_len + vhost_hlen;
 505                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 506                                         &in, vq_log, &log,
 507                                         likely(mergeable) ? UIO_MAXIOV : 1);
 508                 /* On error, stop handling until the next kick. */
 509                 if (unlikely(headcount < 0))
 510                         break;
 511                 /* OK, now we need to know about added descriptors. */
 512                 if (!headcount) {
 513                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 514                                 /* They have slipped one in as we were
 515                                  * doing that: check again. */
 516                                 vhost_disable_notify(&net->dev, vq);
 517                                 continue;
 518                         }
 519                         /* Nothing new?  Wait for eventfd to tell us
 520                          * they refilled. */
 521                         break;
 522                 }
 523                 /* We don't need to be notified again. */
 524                 if (unlikely((vhost_hlen)))
 525                         /* Skip header. TODO: support TSO. */
 526                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 527                 else
 528                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 529                          * needed because recvmsg can modify msg_iov. */
 530                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 531                 msg.msg_iovlen = in;
 532                 err = sock->ops->recvmsg(NULL, sock, &msg,
 533                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 534                 /* Userspace might have consumed the packet meanwhile:
 535                  * it's not supposed to do this usually, but might be hard
 536                  * to prevent. Discard data we got (if any) and keep going. */
 537                 if (unlikely(err != sock_len)) {
 538                         pr_debug("Discarded rx packet: "
 539                                  " len %d, expected %zd\n", err, sock_len);
 540                         vhost_discard_vq_desc(vq, headcount);
 541                         continue;
 542                 }
 543                 if (unlikely(vhost_hlen) &&
 544                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 545                                       vhost_hlen)) {
 546                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 547                                vq->iov->iov_base);
 548                         break;
 549                 }
 550                 /* TODO: Should check and handle checksum. */
 551                 if (likely(mergeable) &&
 552                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 553                                       offsetof(typeof(hdr), num_buffers),
 554                                       sizeof hdr.num_buffers)) {
 555                         vq_err(vq, "Failed num_buffers write");
 556                         vhost_discard_vq_desc(vq, headcount);
 557                         break;
 558                 }
 559                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 560                                             headcount);
 561                 if (unlikely(vq_log))
 562                         vhost_log_write(vq, vq_log, log, vhost_len);
 563                 total_len += vhost_len;
 564                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 565                         vhost_poll_queue(&vq->poll);
 566                         break;
 567                 }
 568         }
 569
 570         mutex_unlock(&vq->mutex);
 571 }
 572
 573 static void handle_tx_kick(struct vhost_work *work)
 574 {
 575         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 576                                                   poll.work);
 577         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 578
 579         handle_tx(net);
 580 }
 581
 582 static void handle_rx_kick(struct vhost_work *work)
 583 {
 584         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 585                                                   poll.work);
 586         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 587
 588         handle_rx(net);
 589 }
 590
 591 static void handle_tx_net(struct vhost_work *work)
 592 {
 593         struct vhost_net *net = container_of(work, struct vhost_net,
 594                                              poll[VHOST_NET_VQ_TX].work);
 595         handle_tx(net);
 596 }
 597
 598 static void handle_rx_net(struct vhost_work *work)
 599 {
 600         struct vhost_net *net = container_of(work, struct vhost_net,
 601                                              poll[VHOST_NET_VQ_RX].work);
 602         handle_rx(net);
 603 }
 604
 605 static int vhost_net_open(struct inode *inode, struct file *f)
 606 {
 607         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 608         struct vhost_dev *dev;
 609         int r;
 610
 611         if (!n)
 612                 return -ENOMEM;
 613
 614         dev = &n->dev;
 615         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 616         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 617         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 618         if (r < 0) {
 619                 kfree(n);
 620                 return r;
 621         }
 622
 623         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 624         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 625         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 626
 627         f->private_data = n;
 628
 629         return 0;
 630 }
 631
 632 static void vhost_net_disable_vq(struct vhost_net *n,
 633                                  struct vhost_virtqueue *vq)
 634 {
 635         if (!vq->private_data)
 636                 return;
 637         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 638                 tx_poll_stop(n);
 639                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 640         } else
 641                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 642 }
 643
 644 static void vhost_net_enable_vq(struct vhost_net *n,
 645                                 struct vhost_virtqueue *vq)
 646 {
 647         struct socket *sock;
 648
 649         sock = rcu_dereference_protected(vq->private_data,
 650                                          lockdep_is_held(&vq->mutex));
 651         if (!sock)
 652                 return;
 653         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 654                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 655                 tx_poll_start(n, sock);
 656         } else
 657                 vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 658 }
 659
 660 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 661                                         struct vhost_virtqueue *vq)
 662 {
 663         struct socket *sock;
 664
 665         mutex_lock(&vq->mutex);
 666         sock = rcu_dereference_protected(vq->private_data,
 667                                          lockdep_is_held(&vq->mutex));
 668         vhost_net_disable_vq(n, vq);
 669         rcu_assign_pointer(vq->private_data, NULL);
 670         mutex_unlock(&vq->mutex);
 671         return sock;
 672 }
 673
 674 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 675                            struct socket **rx_sock)
 676 {
 677         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 678         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 679 }
 680
 681 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 682 {
 683         vhost_poll_flush(n->poll + index);
 684         vhost_poll_flush(&n->dev.vqs[index].poll);
 685 }
 686
 687 static void vhost_net_flush(struct vhost_net *n)
 688 {
 689         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 690         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 691         if (n->dev.vqs[VHOST_NET_VQ_TX].ubufs) {
 692                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 693                 n->tx_flush = true;
 694                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 695                 /* Wait for all lower device DMAs done. */
 696                 vhost_ubuf_put_and_wait(n->dev.vqs[VHOST_NET_VQ_TX].ubufs);
 697                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 698                 n->tx_flush = false;
 699                 kref_init(&n->dev.vqs[VHOST_NET_VQ_TX].ubufs->kref);
 700                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 701         }
 702 }
 703
 704 static int vhost_net_release(struct inode *inode, struct file *f)
 705 {
 706         struct vhost_net *n = f->private_data;
 707         struct socket *tx_sock;
 708         struct socket *rx_sock;
 709
 710         vhost_net_stop(n, &tx_sock, &rx_sock);
 711         vhost_net_flush(n);
 712         vhost_dev_stop(&n->dev);
 713         vhost_dev_cleanup(&n->dev, false);
 714         if (tx_sock)
 715                 fput(tx_sock->file);
 716         if (rx_sock)
 717                 fput(rx_sock->file);
 718         /* We do an extra flush before freeing memory,
 719          * since jobs can re-queue themselves. */
 720         vhost_net_flush(n);
 721         kfree(n);
 722         return 0;
 723 }
 724
 725 static struct socket *get_raw_socket(int fd)
 726 {
 727         struct {
 728                 struct sockaddr_ll sa;
 729                 char  buf[MAX_ADDR_LEN];
 730         } uaddr;
 731         int uaddr_len = sizeof uaddr, r;
 732         struct socket *sock = sockfd_lookup(fd, &r);
 733
 734         if (!sock)
 735                 return ERR_PTR(-ENOTSOCK);
 736
 737         /* Parameter checking */
 738         if (sock->sk->sk_type != SOCK_RAW) {
 739                 r = -ESOCKTNOSUPPORT;
 740                 goto err;
 741         }
 742
 743         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 744                                &uaddr_len, 0);
 745         if (r)
 746                 goto err;
 747
 748         if (uaddr.sa.sll_family != AF_PACKET) {
 749                 r = -EPFNOSUPPORT;
 750                 goto err;
 751         }
 752         return sock;
 753 err:
 754         fput(sock->file);
 755         return ERR_PTR(r);
 756 }
 757
 758 static struct socket *get_tap_socket(int fd)
 759 {
 760         struct file *file = fget(fd);
 761         struct socket *sock;
 762
 763         if (!file)
 764                 return ERR_PTR(-EBADF);
 765         sock = tun_get_socket(file);
 766         if (!IS_ERR(sock))
 767                 return sock;
 768         sock = macvtap_get_socket(file);
 769         if (IS_ERR(sock))
 770                 fput(file);
 771         return sock;
 772 }
 773
 774 static struct socket *get_socket(int fd)
 775 {
 776         struct socket *sock;
 777
 778         /* special case to disable backend */
 779         if (fd == -1)
 780                 return NULL;
 781         sock = get_raw_socket(fd);
 782         if (!IS_ERR(sock))
 783                 return sock;
 784         sock = get_tap_socket(fd);
 785         if (!IS_ERR(sock))
 786                 return sock;
 787         return ERR_PTR(-ENOTSOCK);
 788 }
 789
 790 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 791 {
 792         struct socket *sock, *oldsock;
 793         struct vhost_virtqueue *vq;
 794         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 795         int r;
 796
 797         mutex_lock(&n->dev.mutex);
 798         r = vhost_dev_check_owner(&n->dev);
 799         if (r)
 800                 goto err;
 801
 802         if (index >= VHOST_NET_VQ_MAX) {
 803                 r = -ENOBUFS;
 804                 goto err;
 805         }
 806         vq = n->vqs + index;
 807         mutex_lock(&vq->mutex);
 808
 809         /* Verify that ring has been setup correctly. */
 810         if (!vhost_vq_access_ok(vq)) {
 811                 r = -EFAULT;
 812                 goto err_vq;
 813         }
 814         sock = get_socket(fd);
 815         if (IS_ERR(sock)) {
 816                 r = PTR_ERR(sock);
 817                 goto err_vq;
 818         }
 819
 820         /* start polling new socket */
 821         oldsock = rcu_dereference_protected(vq->private_data,
 822                                             lockdep_is_held(&vq->mutex));
 823         if (sock != oldsock) {
 824                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 825                 if (IS_ERR(ubufs)) {
 826                         r = PTR_ERR(ubufs);
 827                         goto err_ubufs;
 828                 }
 829                 oldubufs = vq->ubufs;
 830                 vq->ubufs = ubufs;
 831                 vhost_net_disable_vq(n, vq);
 832                 rcu_assign_pointer(vq->private_data, sock);
 833                 vhost_net_enable_vq(n, vq);
 834
 835                 r = vhost_init_used(vq);
 836                 if (r)
 837                         goto err_vq;
 838
 839                 n->tx_packets = 0;
 840                 n->tx_zcopy_err = 0;
 841                 n->tx_flush = false;
 842         }
 843
 844         mutex_unlock(&vq->mutex);
 845
 846         if (oldubufs) {
 847                 vhost_ubuf_put_and_wait(oldubufs);
 848                 mutex_lock(&vq->mutex);
 849                 vhost_zerocopy_signal_used(n, vq);
 850                 mutex_unlock(&vq->mutex);
 851         }
 852
 853         if (oldsock) {
 854                 vhost_net_flush_vq(n, index);
 855                 fput(oldsock->file);
 856         }
 857
 858         mutex_unlock(&n->dev.mutex);
 859         return 0;
 860
 861 err_ubufs:
 862         fput(sock->file);
 863 err_vq:
 864         mutex_unlock(&vq->mutex);
 865 err:
 866         mutex_unlock(&n->dev.mutex);
 867         return r;
 868 }
 869
 870 static long vhost_net_reset_owner(struct vhost_net *n)
 871 {
 872         struct socket *tx_sock = NULL;
 873         struct socket *rx_sock = NULL;
 874         long err;
 875
 876         mutex_lock(&n->dev.mutex);
 877         err = vhost_dev_check_owner(&n->dev);
 878         if (err)
 879                 goto done;
 880         vhost_net_stop(n, &tx_sock, &rx_sock);
 881         vhost_net_flush(n);
 882         err = vhost_dev_reset_owner(&n->dev);
 883 done:
 884         mutex_unlock(&n->dev.mutex);
 885         if (tx_sock)
 886                 fput(tx_sock->file);
 887         if (rx_sock)
 888                 fput(rx_sock->file);
 889         return err;
 890 }
 891
 892 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 893 {
 894         size_t vhost_hlen, sock_hlen, hdr_len;
 895         int i;
 896
 897         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 898                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 899                         sizeof(struct virtio_net_hdr);
 900         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 901                 /* vhost provides vnet_hdr */
 902                 vhost_hlen = hdr_len;
 903                 sock_hlen = 0;
 904         } else {
 905                 /* socket provides vnet_hdr */
 906                 vhost_hlen = 0;
 907                 sock_hlen = hdr_len;
 908         }
 909         mutex_lock(&n->dev.mutex);
 910         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 911             !vhost_log_access_ok(&n->dev)) {
 912                 mutex_unlock(&n->dev.mutex);
 913                 return -EFAULT;
 914         }
 915         n->dev.acked_features = features;
 916         smp_wmb();
 917         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 918                 mutex_lock(&n->vqs[i].mutex);
 919                 n->vqs[i].vhost_hlen = vhost_hlen;
 920                 n->vqs[i].sock_hlen = sock_hlen;
 921                 mutex_unlock(&n->vqs[i].mutex);
 922         }
 923         vhost_net_flush(n);
 924         mutex_unlock(&n->dev.mutex);
 925         return 0;
 926 }
 927
 928 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 929                             unsigned long arg)
 930 {
 931         struct vhost_net *n = f->private_data;
 932         void __user *argp = (void __user *)arg;
 933         u64 __user *featurep = argp;
 934         struct vhost_vring_file backend;
 935         u64 features;
 936         int r;
 937
 938         switch (ioctl) {
 939         case VHOST_NET_SET_BACKEND:
 940                 if (copy_from_user(&backend, argp, sizeof backend))
 941                         return -EFAULT;
 942                 return vhost_net_set_backend(n, backend.index, backend.fd);
 943         case VHOST_GET_FEATURES:
 944                 features = VHOST_NET_FEATURES;
 945                 if (copy_to_user(featurep, &features, sizeof features))
 946                         return -EFAULT;
 947                 return 0;
 948         case VHOST_SET_FEATURES:
 949                 if (copy_from_user(&features, featurep, sizeof features))
 950                         return -EFAULT;
 951                 if (features & ~VHOST_NET_FEATURES)
 952                         return -EOPNOTSUPP;
 953                 return vhost_net_set_features(n, features);
 954         case VHOST_RESET_OWNER:
 955                 return vhost_net_reset_owner(n);
 956         default:
 957                 mutex_lock(&n->dev.mutex);
 958                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
 959                 if (r == -ENOIOCTLCMD)
 960                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
 961                 else
 962                         vhost_net_flush(n);
 963                 mutex_unlock(&n->dev.mutex);
 964                 return r;
 965         }
 966 }
 967
 968 #ifdef CONFIG_COMPAT
 969 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 970                                    unsigned long arg)
 971 {
 972         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 973 }
 974 #endif
 975
 976 static const struct file_operations vhost_net_fops = {
 977         .owner          = THIS_MODULE,
 978         .release        = vhost_net_release,
 979         .unlocked_ioctl = vhost_net_ioctl,
 980 #ifdef CONFIG_COMPAT
 981         .compat_ioctl   = vhost_net_compat_ioctl,
 982 #endif
 983         .open           = vhost_net_open,
 984         .llseek         = noop_llseek,
 985 };
 986
 987 static struct miscdevice vhost_net_misc = {
 988         .minor = VHOST_NET_MINOR,
 989         .name = "vhost-net",
 990         .fops = &vhost_net_fops,
 991 };
 992
 993 static int vhost_net_init(void)
 994 {
 995         if (experimental_zcopytx)
 996                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
 997         return misc_register(&vhost_net_misc);
 998 }
 999 module_init(vhost_net_init);
1000
1001 static void vhost_net_exit(void)
1002 {
1003         misc_deregister(&vhost_net_misc);
1004 }
1005 module_exit(vhost_net_exit);
1006
1007 MODULE_VERSION("0.0.1");
1008 MODULE_LICENSE("GPL v2");
1009 MODULE_AUTHOR("Michael S. Tsirkin");
1010 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1011 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1012 MODULE_ALIAS("devname:vhost-net");