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;
 343                                 ubuf = vq->ubuf_info + vq->upend_idx;
 344
 345                                 vq->heads[vq->upend_idx].len =
 346                                         VHOST_DMA_IN_PROGRESS;
 347                                 ubuf->callback = vhost_zerocopy_callback;
 348                                 ubuf->ctx = vq->ubufs;
 349                                 ubuf->desc = vq->upend_idx;
 350                                 msg.msg_control = ubuf;
 351                                 msg.msg_controllen = sizeof(ubuf);
 352                                 ubufs = vq->ubufs;
 353                                 kref_get(&ubufs->kref);
 354                         }
 355                         vq->upend_idx = (vq->upend_idx + 1) % UIO_MAXIOV;
 356                 }
 357                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 358                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 359                 if (unlikely(err < 0)) {
 360                         if (zcopy_used) {
 361                                 if (ubufs)
 362                                         vhost_ubuf_put(ubufs);
 363                                 vq->upend_idx = ((unsigned)vq->upend_idx - 1) %
 364                                         UIO_MAXIOV;
 365                         }
 366                         vhost_discard_vq_desc(vq, 1);
 367                         if (err == -EAGAIN || err == -ENOBUFS)
 368                                 tx_poll_start(net, sock);
 369                         break;
 370                 }
 371                 if (err != len)
 372                         pr_debug("Truncated TX packet: "
 373                                  " len %d != %zd\n", err, len);
 374                 if (!zcopy_used)
 375                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 376                 else
 377                         vhost_zerocopy_signal_used(net, vq);
 378                 total_len += len;
 379                 vhost_net_tx_packet(net);
 380                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 381                         vhost_poll_queue(&vq->poll);
 382                         break;
 383                 }
 384         }
 385
 386         mutex_unlock(&vq->mutex);
 387 }
 388
 389 static int peek_head_len(struct sock *sk)
 390 {
 391         struct sk_buff *head;
 392         int len = 0;
 393         unsigned long flags;
 394
 395         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 396         head = skb_peek(&sk->sk_receive_queue);
 397         if (likely(head)) {
 398                 len = head->len;
 399                 if (vlan_tx_tag_present(head))
 400                         len += VLAN_HLEN;
 401         }
 402
 403         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 404         return len;
 405 }
 406
 407 /* This is a multi-buffer version of vhost_get_desc, that works if
 408  *      vq has read descriptors only.
 409  * @vq          - the relevant virtqueue
 410  * @datalen     - data length we'll be reading
 411  * @iovcount    - returned count of io vectors we fill
 412  * @log         - vhost log
 413  * @log_num     - log offset
 414  * @quota       - headcount quota, 1 for big buffer
 415  *      returns number of buffer heads allocated, negative on error
 416  */
 417 static int get_rx_bufs(struct vhost_virtqueue *vq,
 418                        struct vring_used_elem *heads,
 419                        int datalen,
 420                        unsigned *iovcount,
 421                        struct vhost_log *log,
 422                        unsigned *log_num,
 423                        unsigned int quota)
 424 {
 425         unsigned int out, in;
 426         int seg = 0;
 427         int headcount = 0;
 428         unsigned d;
 429         int r, nlogs = 0;
 430
 431         while (datalen > 0 && headcount < quota) {
 432                 if (unlikely(seg >= UIO_MAXIOV)) {
 433                         r = -ENOBUFS;
 434                         goto err;
 435                 }
 436                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 437                                       ARRAY_SIZE(vq->iov) - seg, &out,
 438                                       &in, log, log_num);
 439                 if (d == vq->num) {
 440                         r = 0;
 441                         goto err;
 442                 }
 443                 if (unlikely(out || in <= 0)) {
 444                         vq_err(vq, "unexpected descriptor format for RX: "
 445                                 "out %d, in %d\n", out, in);
 446                         r = -EINVAL;
 447                         goto err;
 448                 }
 449                 if (unlikely(log)) {
 450                         nlogs += *log_num;
 451                         log += *log_num;
 452                 }
 453                 heads[headcount].id = d;
 454                 heads[headcount].len = iov_length(vq->iov + seg, in);
 455                 datalen -= heads[headcount].len;
 456                 ++headcount;
 457                 seg += in;
 458         }
 459         heads[headcount - 1].len += datalen;
 460         *iovcount = seg;
 461         if (unlikely(log))
 462                 *log_num = nlogs;
 463         return headcount;
 464 err:
 465         vhost_discard_vq_desc(vq, headcount);
 466         return r;
 467 }
 468
 469 /* Expects to be always run from workqueue - which acts as
 470  * read-size critical section for our kind of RCU. */
 471 static void handle_rx(struct vhost_net *net)
 472 {
 473         struct vhost_virtqueue *vq = &net->dev.vqs[VHOST_NET_VQ_RX];
 474         unsigned uninitialized_var(in), log;
 475         struct vhost_log *vq_log;
 476         struct msghdr msg = {
 477                 .msg_name = NULL,
 478                 .msg_namelen = 0,
 479                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 480                 .msg_controllen = 0,
 481                 .msg_iov = vq->iov,
 482                 .msg_flags = MSG_DONTWAIT,
 483         };
 484         struct virtio_net_hdr_mrg_rxbuf hdr = {
 485                 .hdr.flags = 0,
 486                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 487         };
 488         size_t total_len = 0;
 489         int err, mergeable;
 490         s16 headcount;
 491         size_t vhost_hlen, sock_hlen;
 492         size_t vhost_len, sock_len;
 493         /* TODO: check that we are running from vhost_worker? */
 494         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 495
 496         if (!sock)
 497                 return;
 498
 499         mutex_lock(&vq->mutex);
 500         vhost_disable_notify(&net->dev, vq);
 501         vhost_hlen = vq->vhost_hlen;
 502         sock_hlen = vq->sock_hlen;
 503
 504         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 505                 vq->log : NULL;
 506         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 507
 508         while ((sock_len = peek_head_len(sock->sk))) {
 509                 sock_len += sock_hlen;
 510                 vhost_len = sock_len + vhost_hlen;
 511                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 512                                         &in, vq_log, &log,
 513                                         likely(mergeable) ? UIO_MAXIOV : 1);
 514                 /* On error, stop handling until the next kick. */
 515                 if (unlikely(headcount < 0))
 516                         break;
 517                 /* OK, now we need to know about added descriptors. */
 518                 if (!headcount) {
 519                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 520                                 /* They have slipped one in as we were
 521                                  * doing that: check again. */
 522                                 vhost_disable_notify(&net->dev, vq);
 523                                 continue;
 524                         }
 525                         /* Nothing new?  Wait for eventfd to tell us
 526                          * they refilled. */
 527                         break;
 528                 }
 529                 /* We don't need to be notified again. */
 530                 if (unlikely((vhost_hlen)))
 531                         /* Skip header. TODO: support TSO. */
 532                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 533                 else
 534                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 535                          * needed because recvmsg can modify msg_iov. */
 536                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 537                 msg.msg_iovlen = in;
 538                 err = sock->ops->recvmsg(NULL, sock, &msg,
 539                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 540                 /* Userspace might have consumed the packet meanwhile:
 541                  * it's not supposed to do this usually, but might be hard
 542                  * to prevent. Discard data we got (if any) and keep going. */
 543                 if (unlikely(err != sock_len)) {
 544                         pr_debug("Discarded rx packet: "
 545                                  " len %d, expected %zd\n", err, sock_len);
 546                         vhost_discard_vq_desc(vq, headcount);
 547                         continue;
 548                 }
 549                 if (unlikely(vhost_hlen) &&
 550                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 551                                       vhost_hlen)) {
 552                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 553                                vq->iov->iov_base);
 554                         break;
 555                 }
 556                 /* TODO: Should check and handle checksum. */
 557                 if (likely(mergeable) &&
 558                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 559                                       offsetof(typeof(hdr), num_buffers),
 560                                       sizeof hdr.num_buffers)) {
 561                         vq_err(vq, "Failed num_buffers write");
 562                         vhost_discard_vq_desc(vq, headcount);
 563                         break;
 564                 }
 565                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 566                                             headcount);
 567                 if (unlikely(vq_log))
 568                         vhost_log_write(vq, vq_log, log, vhost_len);
 569                 total_len += vhost_len;
 570                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 571                         vhost_poll_queue(&vq->poll);
 572                         break;
 573                 }
 574         }
 575
 576         mutex_unlock(&vq->mutex);
 577 }
 578
 579 static void handle_tx_kick(struct vhost_work *work)
 580 {
 581         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 582                                                   poll.work);
 583         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 584
 585         handle_tx(net);
 586 }
 587
 588 static void handle_rx_kick(struct vhost_work *work)
 589 {
 590         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 591                                                   poll.work);
 592         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 593
 594         handle_rx(net);
 595 }
 596
 597 static void handle_tx_net(struct vhost_work *work)
 598 {
 599         struct vhost_net *net = container_of(work, struct vhost_net,
 600                                              poll[VHOST_NET_VQ_TX].work);
 601         handle_tx(net);
 602 }
 603
 604 static void handle_rx_net(struct vhost_work *work)
 605 {
 606         struct vhost_net *net = container_of(work, struct vhost_net,
 607                                              poll[VHOST_NET_VQ_RX].work);
 608         handle_rx(net);
 609 }
 610
 611 static int vhost_net_open(struct inode *inode, struct file *f)
 612 {
 613         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 614         struct vhost_dev *dev;
 615         int r;
 616
 617         if (!n)
 618                 return -ENOMEM;
 619
 620         dev = &n->dev;
 621         n->vqs[VHOST_NET_VQ_TX].handle_kick = handle_tx_kick;
 622         n->vqs[VHOST_NET_VQ_RX].handle_kick = handle_rx_kick;
 623         r = vhost_dev_init(dev, n->vqs, VHOST_NET_VQ_MAX);
 624         if (r < 0) {
 625                 kfree(n);
 626                 return r;
 627         }
 628
 629         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 630         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 631         n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 632
 633         f->private_data = n;
 634
 635         return 0;
 636 }
 637
 638 static void vhost_net_disable_vq(struct vhost_net *n,
 639                                  struct vhost_virtqueue *vq)
 640 {
 641         if (!vq->private_data)
 642                 return;
 643         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 644                 tx_poll_stop(n);
 645                 n->tx_poll_state = VHOST_NET_POLL_DISABLED;
 646         } else
 647                 vhost_poll_stop(n->poll + VHOST_NET_VQ_RX);
 648 }
 649
 650 static int vhost_net_enable_vq(struct vhost_net *n,
 651                                 struct vhost_virtqueue *vq)
 652 {
 653         struct socket *sock;
 654         int ret;
 655
 656         sock = rcu_dereference_protected(vq->private_data,
 657                                          lockdep_is_held(&vq->mutex));
 658         if (!sock)
 659                 return 0;
 660         if (vq == n->vqs + VHOST_NET_VQ_TX) {
 661                 n->tx_poll_state = VHOST_NET_POLL_STOPPED;
 662                 ret = tx_poll_start(n, sock);
 663         } else
 664                 ret = vhost_poll_start(n->poll + VHOST_NET_VQ_RX, sock->file);
 665
 666         return ret;
 667 }
 668
 669 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 670                                         struct vhost_virtqueue *vq)
 671 {
 672         struct socket *sock;
 673
 674         mutex_lock(&vq->mutex);
 675         sock = rcu_dereference_protected(vq->private_data,
 676                                          lockdep_is_held(&vq->mutex));
 677         vhost_net_disable_vq(n, vq);
 678         rcu_assign_pointer(vq->private_data, NULL);
 679         mutex_unlock(&vq->mutex);
 680         return sock;
 681 }
 682
 683 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 684                            struct socket **rx_sock)
 685 {
 686         *tx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_TX);
 687         *rx_sock = vhost_net_stop_vq(n, n->vqs + VHOST_NET_VQ_RX);
 688 }
 689
 690 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 691 {
 692         vhost_poll_flush(n->poll + index);
 693         vhost_poll_flush(&n->dev.vqs[index].poll);
 694 }
 695
 696 static void vhost_net_flush(struct vhost_net *n)
 697 {
 698         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 699         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 700         if (n->dev.vqs[VHOST_NET_VQ_TX].ubufs) {
 701                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 702                 n->tx_flush = true;
 703                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 704                 /* Wait for all lower device DMAs done. */
 705                 vhost_ubuf_put_and_wait(n->dev.vqs[VHOST_NET_VQ_TX].ubufs);
 706                 mutex_lock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 707                 n->tx_flush = false;
 708                 kref_init(&n->dev.vqs[VHOST_NET_VQ_TX].ubufs->kref);
 709                 mutex_unlock(&n->dev.vqs[VHOST_NET_VQ_TX].mutex);
 710         }
 711 }
 712
 713 static int vhost_net_release(struct inode *inode, struct file *f)
 714 {
 715         struct vhost_net *n = f->private_data;
 716         struct socket *tx_sock;
 717         struct socket *rx_sock;
 718
 719         vhost_net_stop(n, &tx_sock, &rx_sock);
 720         vhost_net_flush(n);
 721         vhost_dev_stop(&n->dev);
 722         vhost_dev_cleanup(&n->dev, false);
 723         if (tx_sock)
 724                 fput(tx_sock->file);
 725         if (rx_sock)
 726                 fput(rx_sock->file);
 727         /* We do an extra flush before freeing memory,
 728          * since jobs can re-queue themselves. */
 729         vhost_net_flush(n);
 730         kfree(n);
 731         return 0;
 732 }
 733
 734 static struct socket *get_raw_socket(int fd)
 735 {
 736         struct {
 737                 struct sockaddr_ll sa;
 738                 char  buf[MAX_ADDR_LEN];
 739         } uaddr;
 740         int uaddr_len = sizeof uaddr, r;
 741         struct socket *sock = sockfd_lookup(fd, &r);
 742
 743         if (!sock)
 744                 return ERR_PTR(-ENOTSOCK);
 745
 746         /* Parameter checking */
 747         if (sock->sk->sk_type != SOCK_RAW) {
 748                 r = -ESOCKTNOSUPPORT;
 749                 goto err;
 750         }
 751
 752         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 753                                &uaddr_len, 0);
 754         if (r)
 755                 goto err;
 756
 757         if (uaddr.sa.sll_family != AF_PACKET) {
 758                 r = -EPFNOSUPPORT;
 759                 goto err;
 760         }
 761         return sock;
 762 err:
 763         fput(sock->file);
 764         return ERR_PTR(r);
 765 }
 766
 767 static struct socket *get_tap_socket(int fd)
 768 {
 769         struct file *file = fget(fd);
 770         struct socket *sock;
 771
 772         if (!file)
 773                 return ERR_PTR(-EBADF);
 774         sock = tun_get_socket(file);
 775         if (!IS_ERR(sock))
 776                 return sock;
 777         sock = macvtap_get_socket(file);
 778         if (IS_ERR(sock))
 779                 fput(file);
 780         return sock;
 781 }
 782
 783 static struct socket *get_socket(int fd)
 784 {
 785         struct socket *sock;
 786
 787         /* special case to disable backend */
 788         if (fd == -1)
 789                 return NULL;
 790         sock = get_raw_socket(fd);
 791         if (!IS_ERR(sock))
 792                 return sock;
 793         sock = get_tap_socket(fd);
 794         if (!IS_ERR(sock))
 795                 return sock;
 796         return ERR_PTR(-ENOTSOCK);
 797 }
 798
 799 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 800 {
 801         struct socket *sock, *oldsock;
 802         struct vhost_virtqueue *vq;
 803         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 804         int r;
 805
 806         mutex_lock(&n->dev.mutex);
 807         r = vhost_dev_check_owner(&n->dev);
 808         if (r)
 809                 goto err;
 810
 811         if (index >= VHOST_NET_VQ_MAX) {
 812                 r = -ENOBUFS;
 813                 goto err;
 814         }
 815         vq = n->vqs + index;
 816         mutex_lock(&vq->mutex);
 817
 818         /* Verify that ring has been setup correctly. */
 819         if (!vhost_vq_access_ok(vq)) {
 820                 r = -EFAULT;
 821                 goto err_vq;
 822         }
 823         sock = get_socket(fd);
 824         if (IS_ERR(sock)) {
 825                 r = PTR_ERR(sock);
 826                 goto err_vq;
 827         }
 828
 829         /* start polling new socket */
 830         oldsock = rcu_dereference_protected(vq->private_data,
 831                                             lockdep_is_held(&vq->mutex));
 832         if (sock != oldsock) {
 833                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 834                 if (IS_ERR(ubufs)) {
 835                         r = PTR_ERR(ubufs);
 836                         goto err_ubufs;
 837                 }
 838
 839                 vhost_net_disable_vq(n, vq);
 840                 rcu_assign_pointer(vq->private_data, sock);
 841                 r = vhost_init_used(vq);
 842                 if (r)
 843                         goto err_used;
 844                 r = vhost_net_enable_vq(n, vq);
 845                 if (r)
 846                         goto err_used;
 847
 848                 oldubufs = vq->ubufs;
 849                 vq->ubufs = ubufs;
 850
 851                 n->tx_packets = 0;
 852                 n->tx_zcopy_err = 0;
 853                 n->tx_flush = false;
 854         }
 855
 856         mutex_unlock(&vq->mutex);
 857
 858         if (oldubufs) {
 859                 vhost_ubuf_put_and_wait(oldubufs);
 860                 mutex_lock(&vq->mutex);
 861                 vhost_zerocopy_signal_used(n, vq);
 862                 mutex_unlock(&vq->mutex);
 863         }
 864
 865         if (oldsock) {
 866                 vhost_net_flush_vq(n, index);
 867                 fput(oldsock->file);
 868         }
 869
 870         mutex_unlock(&n->dev.mutex);
 871         return 0;
 872
 873 err_used:
 874         rcu_assign_pointer(vq->private_data, oldsock);
 875         vhost_net_enable_vq(n, vq);
 876         if (ubufs)
 877                 vhost_ubuf_put_and_wait(ubufs);
 878 err_ubufs:
 879         fput(sock->file);
 880 err_vq:
 881         mutex_unlock(&vq->mutex);
 882 err:
 883         mutex_unlock(&n->dev.mutex);
 884         return r;
 885 }
 886
 887 static long vhost_net_reset_owner(struct vhost_net *n)
 888 {
 889         struct socket *tx_sock = NULL;
 890         struct socket *rx_sock = NULL;
 891         long err;
 892
 893         mutex_lock(&n->dev.mutex);
 894         err = vhost_dev_check_owner(&n->dev);
 895         if (err)
 896                 goto done;
 897         vhost_net_stop(n, &tx_sock, &rx_sock);
 898         vhost_net_flush(n);
 899         err = vhost_dev_reset_owner(&n->dev);
 900 done:
 901         mutex_unlock(&n->dev.mutex);
 902         if (tx_sock)
 903                 fput(tx_sock->file);
 904         if (rx_sock)
 905                 fput(rx_sock->file);
 906         return err;
 907 }
 908
 909 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 910 {
 911         size_t vhost_hlen, sock_hlen, hdr_len;
 912         int i;
 913
 914         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 915                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 916                         sizeof(struct virtio_net_hdr);
 917         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 918                 /* vhost provides vnet_hdr */
 919                 vhost_hlen = hdr_len;
 920                 sock_hlen = 0;
 921         } else {
 922                 /* socket provides vnet_hdr */
 923                 vhost_hlen = 0;
 924                 sock_hlen = hdr_len;
 925         }
 926         mutex_lock(&n->dev.mutex);
 927         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 928             !vhost_log_access_ok(&n->dev)) {
 929                 mutex_unlock(&n->dev.mutex);
 930                 return -EFAULT;
 931         }
 932         n->dev.acked_features = features;
 933         smp_wmb();
 934         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 935                 mutex_lock(&n->vqs[i].mutex);
 936                 n->vqs[i].vhost_hlen = vhost_hlen;
 937                 n->vqs[i].sock_hlen = sock_hlen;
 938                 mutex_unlock(&n->vqs[i].mutex);
 939         }
 940         vhost_net_flush(n);
 941         mutex_unlock(&n->dev.mutex);
 942         return 0;
 943 }
 944
 945 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
 946                             unsigned long arg)
 947 {
 948         struct vhost_net *n = f->private_data;
 949         void __user *argp = (void __user *)arg;
 950         u64 __user *featurep = argp;
 951         struct vhost_vring_file backend;
 952         u64 features;
 953         int r;
 954
 955         switch (ioctl) {
 956         case VHOST_NET_SET_BACKEND:
 957                 if (copy_from_user(&backend, argp, sizeof backend))
 958                         return -EFAULT;
 959                 return vhost_net_set_backend(n, backend.index, backend.fd);
 960         case VHOST_GET_FEATURES:
 961                 features = VHOST_NET_FEATURES;
 962                 if (copy_to_user(featurep, &features, sizeof features))
 963                         return -EFAULT;
 964                 return 0;
 965         case VHOST_SET_FEATURES:
 966                 if (copy_from_user(&features, featurep, sizeof features))
 967                         return -EFAULT;
 968                 if (features & ~VHOST_NET_FEATURES)
 969                         return -EOPNOTSUPP;
 970                 return vhost_net_set_features(n, features);
 971         case VHOST_RESET_OWNER:
 972                 return vhost_net_reset_owner(n);
 973         default:
 974                 mutex_lock(&n->dev.mutex);
 975                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
 976                 if (r == -ENOIOCTLCMD)
 977                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
 978                 else
 979                         vhost_net_flush(n);
 980                 mutex_unlock(&n->dev.mutex);
 981                 return r;
 982         }
 983 }
 984
 985 #ifdef CONFIG_COMPAT
 986 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
 987                                    unsigned long arg)
 988 {
 989         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
 990 }
 991 #endif
 992
 993 static const struct file_operations vhost_net_fops = {
 994         .owner          = THIS_MODULE,
 995         .release        = vhost_net_release,
 996         .unlocked_ioctl = vhost_net_ioctl,
 997 #ifdef CONFIG_COMPAT
 998         .compat_ioctl   = vhost_net_compat_ioctl,
 999 #endif
1000         .open           = vhost_net_open,
1001         .llseek         = noop_llseek,
1002 };
1003
1004 static struct miscdevice vhost_net_misc = {
1005         .minor = VHOST_NET_MINOR,
1006         .name = "vhost-net",
1007         .fops = &vhost_net_fops,
1008 };
1009
1010 static int vhost_net_init(void)
1011 {
1012         if (experimental_zcopytx)
1013                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
1014         return misc_register(&vhost_net_misc);
1015 }
1016 module_init(vhost_net_init);
1017
1018 static void vhost_net_exit(void)
1019 {
1020         misc_deregister(&vhost_net_misc);
1021 }
1022 module_exit(vhost_net_exit);
1023
1024 MODULE_VERSION("0.0.1");
1025 MODULE_LICENSE("GPL v2");
1026 MODULE_AUTHOR("Michael S. Tsirkin");
1027 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1028 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1029 MODULE_ALIAS("devname:vhost-net");