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_FEATURES = VHOST_FEATURES |
  63                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  64                          (1ULL << VIRTIO_NET_F_MRG_RXBUF),
  65 };
  66
  67 enum {
  68         VHOST_NET_VQ_RX = 0,
  69         VHOST_NET_VQ_TX = 1,
  70         VHOST_NET_VQ_MAX = 2,
  71 };
  72
  73 struct vhost_ubuf_ref {
  74         struct kref kref;
  75         wait_queue_head_t wait;
  76         struct vhost_virtqueue *vq;
  77 };
  78
  79 struct vhost_net_virtqueue {
  80         struct vhost_virtqueue vq;
  81         /* hdr is used to store the virtio header.
  82          * Since each iovec has >= 1 byte length, we never need more than
  83          * header length entries to store the header. */
  84         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
  85         size_t vhost_hlen;
  86         size_t sock_hlen;
  87         /* vhost zerocopy support fields below: */
  88         /* last used idx for outstanding DMA zerocopy buffers */
  89         int upend_idx;
  90         /* first used idx for DMA done zerocopy buffers */
  91         int done_idx;
  92         /* an array of userspace buffers info */
  93         struct ubuf_info *ubuf_info;
  94         /* Reference counting for outstanding ubufs.
  95          * Protected by vq mutex. Writers must also take device mutex. */
  96         struct vhost_ubuf_ref *ubufs;
  97 };
  98
  99 struct vhost_net {
 100         struct vhost_dev dev;
 101         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 102         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 103         /* Number of TX recently submitted.
 104          * Protected by tx vq lock. */
 105         unsigned tx_packets;
 106         /* Number of times zerocopy TX recently failed.
 107          * Protected by tx vq lock. */
 108         unsigned tx_zcopy_err;
 109         /* Flush in progress. Protected by tx vq lock. */
 110         bool tx_flush;
 111 };
 112
 113 static unsigned vhost_zcopy_mask __read_mostly;
 114
 115 void vhost_enable_zcopy(int vq)
 116 {
 117         vhost_zcopy_mask |= 0x1 << vq;
 118 }
 119
 120 static void vhost_zerocopy_done_signal(struct kref *kref)
 121 {
 122         struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
 123                                                     kref);
 124         wake_up(&ubufs->wait);
 125 }
 126
 127 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
 128                                         bool zcopy)
 129 {
 130         struct vhost_ubuf_ref *ubufs;
 131         /* No zero copy backend? Nothing to count. */
 132         if (!zcopy)
 133                 return NULL;
 134         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 135         if (!ubufs)
 136                 return ERR_PTR(-ENOMEM);
 137         kref_init(&ubufs->kref);
 138         init_waitqueue_head(&ubufs->wait);
 139         ubufs->vq = vq;
 140         return ubufs;
 141 }
 142
 143 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
 144 {
 145         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
 146 }
 147
 148 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
 149 {
 150         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
 151         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
 152         kfree(ubufs);
 153 }
 154
 155 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 156 {
 157
 158         bool zcopy;
 159         int i;
 160
 161         for (i = 0; i < n->dev.nvqs; ++i) {
 162                 zcopy = vhost_zcopy_mask & (0x1 << i);
 163                 if (zcopy)
 164                         kfree(n->vqs[i].ubuf_info);
 165         }
 166 }
 167
 168 int vhost_net_set_ubuf_info(struct vhost_net *n)
 169 {
 170         bool zcopy;
 171         int i;
 172
 173         for (i = 0; i < n->dev.nvqs; ++i) {
 174                 zcopy = vhost_zcopy_mask & (0x1 << i);
 175                 if (!zcopy)
 176                         continue;
 177                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 178                                               UIO_MAXIOV, GFP_KERNEL);
 179                 if  (!n->vqs[i].ubuf_info)
 180                         goto err;
 181         }
 182         return 0;
 183
 184 err:
 185         while (i--) {
 186                 zcopy = vhost_zcopy_mask & (0x1 << i);
 187                 if (!zcopy)
 188                         continue;
 189                 kfree(n->vqs[i].ubuf_info);
 190         }
 191         return -ENOMEM;
 192 }
 193
 194 void vhost_net_vq_reset(struct vhost_net *n)
 195 {
 196         int i;
 197
 198         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 199                 n->vqs[i].done_idx = 0;
 200                 n->vqs[i].upend_idx = 0;
 201                 n->vqs[i].ubufs = NULL;
 202                 kfree(n->vqs[i].ubuf_info);
 203                 n->vqs[i].ubuf_info = NULL;
 204                 n->vqs[i].vhost_hlen = 0;
 205                 n->vqs[i].sock_hlen = 0;
 206         }
 207
 208 }
 209
 210 static void vhost_net_tx_packet(struct vhost_net *net)
 211 {
 212         ++net->tx_packets;
 213         if (net->tx_packets < 1024)
 214                 return;
 215         net->tx_packets = 0;
 216         net->tx_zcopy_err = 0;
 217 }
 218
 219 static void vhost_net_tx_err(struct vhost_net *net)
 220 {
 221         ++net->tx_zcopy_err;
 222 }
 223
 224 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 225 {
 226         /* TX flush waits for outstanding DMAs to be done.
 227          * Don't start new DMAs.
 228          */
 229         return !net->tx_flush &&
 230                 net->tx_packets / 64 >= net->tx_zcopy_err;
 231 }
 232
 233 static bool vhost_sock_zcopy(struct socket *sock)
 234 {
 235         return unlikely(experimental_zcopytx) &&
 236                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 237 }
 238
 239 /* Pop first len bytes from iovec. Return number of segments used. */
 240 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 241                           size_t len, int iov_count)
 242 {
 243         int seg = 0;
 244         size_t size;
 245
 246         while (len && seg < iov_count) {
 247                 size = min(from->iov_len, len);
 248                 to->iov_base = from->iov_base;
 249                 to->iov_len = size;
 250                 from->iov_len -= size;
 251                 from->iov_base += size;
 252                 len -= size;
 253                 ++from;
 254                 ++to;
 255                 ++seg;
 256         }
 257         return seg;
 258 }
 259 /* Copy iovec entries for len bytes from iovec. */
 260 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 261                            size_t len, int iovcount)
 262 {
 263         int seg = 0;
 264         size_t size;
 265
 266         while (len && seg < iovcount) {
 267                 size = min(from->iov_len, len);
 268                 to->iov_base = from->iov_base;
 269                 to->iov_len = size;
 270                 len -= size;
 271                 ++from;
 272                 ++to;
 273                 ++seg;
 274         }
 275 }
 276
 277 /* In case of DMA done not in order in lower device driver for some reason.
 278  * upend_idx is used to track end of used idx, done_idx is used to track head
 279  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 280  * guest used idx.
 281  */
 282 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 283                                       struct vhost_virtqueue *vq)
 284 {
 285         struct vhost_net_virtqueue *nvq =
 286                 container_of(vq, struct vhost_net_virtqueue, vq);
 287         int i;
 288         int j = 0;
 289
 290         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 291                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 292                         vhost_net_tx_err(net);
 293                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 294                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 295                         vhost_add_used_and_signal(vq->dev, vq,
 296                                                   vq->heads[i].id, 0);
 297                         ++j;
 298                 } else
 299                         break;
 300         }
 301         if (j)
 302                 nvq->done_idx = i;
 303         return j;
 304 }
 305
 306 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 307 {
 308         struct vhost_ubuf_ref *ubufs = ubuf->ctx;
 309         struct vhost_virtqueue *vq = ubufs->vq;
 310         int cnt = atomic_read(&ubufs->kref.refcount);
 311
 312         /*
 313          * Trigger polling thread if guest stopped submitting new buffers:
 314          * in this case, the refcount after decrement will eventually reach 1
 315          * so here it is 2.
 316          * We also trigger polling periodically after each 16 packets
 317          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 318          * less than 10% of times).
 319          */
 320         if (cnt <= 2 || !(cnt % 16))
 321                 vhost_poll_queue(&vq->poll);
 322         /* set len to mark this desc buffers done DMA */
 323         vq->heads[ubuf->desc].len = success ?
 324                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 325         vhost_ubuf_put(ubufs);
 326 }
 327
 328 /* Expects to be always run from workqueue - which acts as
 329  * read-size critical section for our kind of RCU. */
 330 static void handle_tx(struct vhost_net *net)
 331 {
 332         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 333         struct vhost_virtqueue *vq = &nvq->vq;
 334         unsigned out, in, s;
 335         int head;
 336         struct msghdr msg = {
 337                 .msg_name = NULL,
 338                 .msg_namelen = 0,
 339                 .msg_control = NULL,
 340                 .msg_controllen = 0,
 341                 .msg_iov = vq->iov,
 342                 .msg_flags = MSG_DONTWAIT,
 343         };
 344         size_t len, total_len = 0;
 345         int err;
 346         size_t hdr_size;
 347         struct socket *sock;
 348         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 349         bool zcopy, zcopy_used;
 350
 351         /* TODO: check that we are running from vhost_worker? */
 352         sock = rcu_dereference_check(vq->private_data, 1);
 353         if (!sock)
 354                 return;
 355
 356         mutex_lock(&vq->mutex);
 357         vhost_disable_notify(&net->dev, vq);
 358
 359         hdr_size = nvq->vhost_hlen;
 360         zcopy = nvq->ubufs;
 361
 362         for (;;) {
 363                 /* Release DMAs done buffers first */
 364                 if (zcopy)
 365                         vhost_zerocopy_signal_used(net, vq);
 366
 367                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 368                                          ARRAY_SIZE(vq->iov),
 369                                          &out, &in,
 370                                          NULL, NULL);
 371                 /* On error, stop handling until the next kick. */
 372                 if (unlikely(head < 0))
 373                         break;
 374                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 375                 if (head == vq->num) {
 376                         int num_pends;
 377
 378                         /* If more outstanding DMAs, queue the work.
 379                          * Handle upend_idx wrap around
 380                          */
 381                         num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
 382                                     (nvq->upend_idx - nvq->done_idx) :
 383                                     (nvq->upend_idx + UIO_MAXIOV -
 384                                      nvq->done_idx);
 385                         if (unlikely(num_pends > VHOST_MAX_PEND))
 386                                 break;
 387                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 388                                 vhost_disable_notify(&net->dev, vq);
 389                                 continue;
 390                         }
 391                         break;
 392                 }
 393                 if (in) {
 394                         vq_err(vq, "Unexpected descriptor format for TX: "
 395                                "out %d, int %d\n", out, in);
 396                         break;
 397                 }
 398                 /* Skip header. TODO: support TSO. */
 399                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 400                 msg.msg_iovlen = out;
 401                 len = iov_length(vq->iov, out);
 402                 /* Sanity check */
 403                 if (!len) {
 404                         vq_err(vq, "Unexpected header len for TX: "
 405                                "%zd expected %zd\n",
 406                                iov_length(nvq->hdr, s), hdr_size);
 407                         break;
 408                 }
 409                 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
 410                                        nvq->upend_idx != nvq->done_idx);
 411
 412                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 413                 if (zcopy_used) {
 414                         vq->heads[nvq->upend_idx].id = head;
 415                         if (!vhost_net_tx_select_zcopy(net) ||
 416                             len < VHOST_GOODCOPY_LEN) {
 417                                 /* copy don't need to wait for DMA done */
 418                                 vq->heads[nvq->upend_idx].len =
 419                                                         VHOST_DMA_DONE_LEN;
 420                                 msg.msg_control = NULL;
 421                                 msg.msg_controllen = 0;
 422                                 ubufs = NULL;
 423                         } else {
 424                                 struct ubuf_info *ubuf;
 425                                 ubuf = nvq->ubuf_info + nvq->upend_idx;
 426
 427                                 vq->heads[nvq->upend_idx].len =
 428                                         VHOST_DMA_IN_PROGRESS;
 429                                 ubuf->callback = vhost_zerocopy_callback;
 430                                 ubuf->ctx = nvq->ubufs;
 431                                 ubuf->desc = nvq->upend_idx;
 432                                 msg.msg_control = ubuf;
 433                                 msg.msg_controllen = sizeof(ubuf);
 434                                 ubufs = nvq->ubufs;
 435                                 kref_get(&ubufs->kref);
 436                         }
 437                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 438                 }
 439                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 440                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 441                 if (unlikely(err < 0)) {
 442                         if (zcopy_used) {
 443                                 if (ubufs)
 444                                         vhost_ubuf_put(ubufs);
 445                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 446                                         % UIO_MAXIOV;
 447                         }
 448                         vhost_discard_vq_desc(vq, 1);
 449                         break;
 450                 }
 451                 if (err != len)
 452                         pr_debug("Truncated TX packet: "
 453                                  " len %d != %zd\n", err, len);
 454                 if (!zcopy_used)
 455                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 456                 else
 457                         vhost_zerocopy_signal_used(net, vq);
 458                 total_len += len;
 459                 vhost_net_tx_packet(net);
 460                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 461                         vhost_poll_queue(&vq->poll);
 462                         break;
 463                 }
 464         }
 465
 466         mutex_unlock(&vq->mutex);
 467 }
 468
 469 static int peek_head_len(struct sock *sk)
 470 {
 471         struct sk_buff *head;
 472         int len = 0;
 473         unsigned long flags;
 474
 475         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 476         head = skb_peek(&sk->sk_receive_queue);
 477         if (likely(head)) {
 478                 len = head->len;
 479                 if (vlan_tx_tag_present(head))
 480                         len += VLAN_HLEN;
 481         }
 482
 483         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 484         return len;
 485 }
 486
 487 /* This is a multi-buffer version of vhost_get_desc, that works if
 488  *      vq has read descriptors only.
 489  * @vq          - the relevant virtqueue
 490  * @datalen     - data length we'll be reading
 491  * @iovcount    - returned count of io vectors we fill
 492  * @log         - vhost log
 493  * @log_num     - log offset
 494  * @quota       - headcount quota, 1 for big buffer
 495  *      returns number of buffer heads allocated, negative on error
 496  */
 497 static int get_rx_bufs(struct vhost_virtqueue *vq,
 498                        struct vring_used_elem *heads,
 499                        int datalen,
 500                        unsigned *iovcount,
 501                        struct vhost_log *log,
 502                        unsigned *log_num,
 503                        unsigned int quota)
 504 {
 505         unsigned int out, in;
 506         int seg = 0;
 507         int headcount = 0;
 508         unsigned d;
 509         int r, nlogs = 0;
 510
 511         while (datalen > 0 && headcount < quota) {
 512                 if (unlikely(seg >= UIO_MAXIOV)) {
 513                         r = -ENOBUFS;
 514                         goto err;
 515                 }
 516                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 517                                       ARRAY_SIZE(vq->iov) - seg, &out,
 518                                       &in, log, log_num);
 519                 if (d == vq->num) {
 520                         r = 0;
 521                         goto err;
 522                 }
 523                 if (unlikely(out || in <= 0)) {
 524                         vq_err(vq, "unexpected descriptor format for RX: "
 525                                 "out %d, in %d\n", out, in);
 526                         r = -EINVAL;
 527                         goto err;
 528                 }
 529                 if (unlikely(log)) {
 530                         nlogs += *log_num;
 531                         log += *log_num;
 532                 }
 533                 heads[headcount].id = d;
 534                 heads[headcount].len = iov_length(vq->iov + seg, in);
 535                 datalen -= heads[headcount].len;
 536                 ++headcount;
 537                 seg += in;
 538         }
 539         heads[headcount - 1].len += datalen;
 540         *iovcount = seg;
 541         if (unlikely(log))
 542                 *log_num = nlogs;
 543         return headcount;
 544 err:
 545         vhost_discard_vq_desc(vq, headcount);
 546         return r;
 547 }
 548
 549 /* Expects to be always run from workqueue - which acts as
 550  * read-size critical section for our kind of RCU. */
 551 static void handle_rx(struct vhost_net *net)
 552 {
 553         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 554         struct vhost_virtqueue *vq = &nvq->vq;
 555         unsigned uninitialized_var(in), log;
 556         struct vhost_log *vq_log;
 557         struct msghdr msg = {
 558                 .msg_name = NULL,
 559                 .msg_namelen = 0,
 560                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 561                 .msg_controllen = 0,
 562                 .msg_iov = vq->iov,
 563                 .msg_flags = MSG_DONTWAIT,
 564         };
 565         struct virtio_net_hdr_mrg_rxbuf hdr = {
 566                 .hdr.flags = 0,
 567                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 568         };
 569         size_t total_len = 0;
 570         int err, mergeable;
 571         s16 headcount;
 572         size_t vhost_hlen, sock_hlen;
 573         size_t vhost_len, sock_len;
 574         /* TODO: check that we are running from vhost_worker? */
 575         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 576
 577         if (!sock)
 578                 return;
 579
 580         mutex_lock(&vq->mutex);
 581         vhost_disable_notify(&net->dev, vq);
 582         vhost_hlen = nvq->vhost_hlen;
 583         sock_hlen = nvq->sock_hlen;
 584
 585         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 586                 vq->log : NULL;
 587         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 588
 589         while ((sock_len = peek_head_len(sock->sk))) {
 590                 sock_len += sock_hlen;
 591                 vhost_len = sock_len + vhost_hlen;
 592                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 593                                         &in, vq_log, &log,
 594                                         likely(mergeable) ? UIO_MAXIOV : 1);
 595                 /* On error, stop handling until the next kick. */
 596                 if (unlikely(headcount < 0))
 597                         break;
 598                 /* OK, now we need to know about added descriptors. */
 599                 if (!headcount) {
 600                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 601                                 /* They have slipped one in as we were
 602                                  * doing that: check again. */
 603                                 vhost_disable_notify(&net->dev, vq);
 604                                 continue;
 605                         }
 606                         /* Nothing new?  Wait for eventfd to tell us
 607                          * they refilled. */
 608                         break;
 609                 }
 610                 /* We don't need to be notified again. */
 611                 if (unlikely((vhost_hlen)))
 612                         /* Skip header. TODO: support TSO. */
 613                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 614                 else
 615                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 616                          * needed because recvmsg can modify msg_iov. */
 617                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 618                 msg.msg_iovlen = in;
 619                 err = sock->ops->recvmsg(NULL, sock, &msg,
 620                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 621                 /* Userspace might have consumed the packet meanwhile:
 622                  * it's not supposed to do this usually, but might be hard
 623                  * to prevent. Discard data we got (if any) and keep going. */
 624                 if (unlikely(err != sock_len)) {
 625                         pr_debug("Discarded rx packet: "
 626                                  " len %d, expected %zd\n", err, sock_len);
 627                         vhost_discard_vq_desc(vq, headcount);
 628                         continue;
 629                 }
 630                 if (unlikely(vhost_hlen) &&
 631                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 632                                       vhost_hlen)) {
 633                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 634                                vq->iov->iov_base);
 635                         break;
 636                 }
 637                 /* TODO: Should check and handle checksum. */
 638                 if (likely(mergeable) &&
 639                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 640                                       offsetof(typeof(hdr), num_buffers),
 641                                       sizeof hdr.num_buffers)) {
 642                         vq_err(vq, "Failed num_buffers write");
 643                         vhost_discard_vq_desc(vq, headcount);
 644                         break;
 645                 }
 646                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 647                                             headcount);
 648                 if (unlikely(vq_log))
 649                         vhost_log_write(vq, vq_log, log, vhost_len);
 650                 total_len += vhost_len;
 651                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 652                         vhost_poll_queue(&vq->poll);
 653                         break;
 654                 }
 655         }
 656
 657         mutex_unlock(&vq->mutex);
 658 }
 659
 660 static void handle_tx_kick(struct vhost_work *work)
 661 {
 662         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 663                                                   poll.work);
 664         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 665
 666         handle_tx(net);
 667 }
 668
 669 static void handle_rx_kick(struct vhost_work *work)
 670 {
 671         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 672                                                   poll.work);
 673         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 674
 675         handle_rx(net);
 676 }
 677
 678 static void handle_tx_net(struct vhost_work *work)
 679 {
 680         struct vhost_net *net = container_of(work, struct vhost_net,
 681                                              poll[VHOST_NET_VQ_TX].work);
 682         handle_tx(net);
 683 }
 684
 685 static void handle_rx_net(struct vhost_work *work)
 686 {
 687         struct vhost_net *net = container_of(work, struct vhost_net,
 688                                              poll[VHOST_NET_VQ_RX].work);
 689         handle_rx(net);
 690 }
 691
 692 static int vhost_net_open(struct inode *inode, struct file *f)
 693 {
 694         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 695         struct vhost_dev *dev;
 696         struct vhost_virtqueue **vqs;
 697         int r, i;
 698
 699         if (!n)
 700                 return -ENOMEM;
 701         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 702         if (!vqs) {
 703                 kfree(n);
 704                 return -ENOMEM;
 705         }
 706
 707         dev = &n->dev;
 708         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 709         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 710         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 711         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 712         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 713                 n->vqs[i].ubufs = NULL;
 714                 n->vqs[i].ubuf_info = NULL;
 715                 n->vqs[i].upend_idx = 0;
 716                 n->vqs[i].done_idx = 0;
 717                 n->vqs[i].vhost_hlen = 0;
 718                 n->vqs[i].sock_hlen = 0;
 719         }
 720         r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 721         if (r < 0) {
 722                 kfree(n);
 723                 kfree(vqs);
 724                 return r;
 725         }
 726
 727         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 728         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 729
 730         f->private_data = n;
 731
 732         return 0;
 733 }
 734
 735 static void vhost_net_disable_vq(struct vhost_net *n,
 736                                  struct vhost_virtqueue *vq)
 737 {
 738         struct vhost_net_virtqueue *nvq =
 739                 container_of(vq, struct vhost_net_virtqueue, vq);
 740         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 741         if (!vq->private_data)
 742                 return;
 743         vhost_poll_stop(poll);
 744 }
 745
 746 static int vhost_net_enable_vq(struct vhost_net *n,
 747                                 struct vhost_virtqueue *vq)
 748 {
 749         struct vhost_net_virtqueue *nvq =
 750                 container_of(vq, struct vhost_net_virtqueue, vq);
 751         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 752         struct socket *sock;
 753
 754         sock = rcu_dereference_protected(vq->private_data,
 755                                          lockdep_is_held(&vq->mutex));
 756         if (!sock)
 757                 return 0;
 758
 759         return vhost_poll_start(poll, sock->file);
 760 }
 761
 762 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 763                                         struct vhost_virtqueue *vq)
 764 {
 765         struct socket *sock;
 766
 767         mutex_lock(&vq->mutex);
 768         sock = rcu_dereference_protected(vq->private_data,
 769                                          lockdep_is_held(&vq->mutex));
 770         vhost_net_disable_vq(n, vq);
 771         rcu_assign_pointer(vq->private_data, NULL);
 772         mutex_unlock(&vq->mutex);
 773         return sock;
 774 }
 775
 776 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 777                            struct socket **rx_sock)
 778 {
 779         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 780         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 781 }
 782
 783 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 784 {
 785         vhost_poll_flush(n->poll + index);
 786         vhost_poll_flush(&n->vqs[index].vq.poll);
 787 }
 788
 789 static void vhost_net_flush(struct vhost_net *n)
 790 {
 791         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 792         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 793         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 794                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 795                 n->tx_flush = true;
 796                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 797                 /* Wait for all lower device DMAs done. */
 798                 vhost_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 799                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 800                 n->tx_flush = false;
 801                 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
 802                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 803         }
 804 }
 805
 806 static int vhost_net_release(struct inode *inode, struct file *f)
 807 {
 808         struct vhost_net *n = f->private_data;
 809         struct socket *tx_sock;
 810         struct socket *rx_sock;
 811
 812         vhost_net_stop(n, &tx_sock, &rx_sock);
 813         vhost_net_flush(n);
 814         vhost_dev_stop(&n->dev);
 815         vhost_dev_cleanup(&n->dev, false);
 816         vhost_net_vq_reset(n);
 817         if (tx_sock)
 818                 fput(tx_sock->file);
 819         if (rx_sock)
 820                 fput(rx_sock->file);
 821         /* We do an extra flush before freeing memory,
 822          * since jobs can re-queue themselves. */
 823         vhost_net_flush(n);
 824         kfree(n->dev.vqs);
 825         kfree(n);
 826         return 0;
 827 }
 828
 829 static struct socket *get_raw_socket(int fd)
 830 {
 831         struct {
 832                 struct sockaddr_ll sa;
 833                 char  buf[MAX_ADDR_LEN];
 834         } uaddr;
 835         int uaddr_len = sizeof uaddr, r;
 836         struct socket *sock = sockfd_lookup(fd, &r);
 837
 838         if (!sock)
 839                 return ERR_PTR(-ENOTSOCK);
 840
 841         /* Parameter checking */
 842         if (sock->sk->sk_type != SOCK_RAW) {
 843                 r = -ESOCKTNOSUPPORT;
 844                 goto err;
 845         }
 846
 847         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 848                                &uaddr_len, 0);
 849         if (r)
 850                 goto err;
 851
 852         if (uaddr.sa.sll_family != AF_PACKET) {
 853                 r = -EPFNOSUPPORT;
 854                 goto err;
 855         }
 856         return sock;
 857 err:
 858         fput(sock->file);
 859         return ERR_PTR(r);
 860 }
 861
 862 static struct socket *get_tap_socket(int fd)
 863 {
 864         struct file *file = fget(fd);
 865         struct socket *sock;
 866
 867         if (!file)
 868                 return ERR_PTR(-EBADF);
 869         sock = tun_get_socket(file);
 870         if (!IS_ERR(sock))
 871                 return sock;
 872         sock = macvtap_get_socket(file);
 873         if (IS_ERR(sock))
 874                 fput(file);
 875         return sock;
 876 }
 877
 878 static struct socket *get_socket(int fd)
 879 {
 880         struct socket *sock;
 881
 882         /* special case to disable backend */
 883         if (fd == -1)
 884                 return NULL;
 885         sock = get_raw_socket(fd);
 886         if (!IS_ERR(sock))
 887                 return sock;
 888         sock = get_tap_socket(fd);
 889         if (!IS_ERR(sock))
 890                 return sock;
 891         return ERR_PTR(-ENOTSOCK);
 892 }
 893
 894 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 895 {
 896         struct socket *sock, *oldsock;
 897         struct vhost_virtqueue *vq;
 898         struct vhost_net_virtqueue *nvq;
 899         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 900         int r;
 901
 902         mutex_lock(&n->dev.mutex);
 903         r = vhost_dev_check_owner(&n->dev);
 904         if (r)
 905                 goto err;
 906
 907         if (index >= VHOST_NET_VQ_MAX) {
 908                 r = -ENOBUFS;
 909                 goto err;
 910         }
 911         vq = &n->vqs[index].vq;
 912         nvq = &n->vqs[index];
 913         mutex_lock(&vq->mutex);
 914
 915         /* Verify that ring has been setup correctly. */
 916         if (!vhost_vq_access_ok(vq)) {
 917                 r = -EFAULT;
 918                 goto err_vq;
 919         }
 920         sock = get_socket(fd);
 921         if (IS_ERR(sock)) {
 922                 r = PTR_ERR(sock);
 923                 goto err_vq;
 924         }
 925
 926         /* start polling new socket */
 927         oldsock = rcu_dereference_protected(vq->private_data,
 928                                             lockdep_is_held(&vq->mutex));
 929         if (sock != oldsock) {
 930                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 931                 if (IS_ERR(ubufs)) {
 932                         r = PTR_ERR(ubufs);
 933                         goto err_ubufs;
 934                 }
 935
 936                 vhost_net_disable_vq(n, vq);
 937                 rcu_assign_pointer(vq->private_data, sock);
 938                 r = vhost_init_used(vq);
 939                 if (r)
 940                         goto err_used;
 941                 r = vhost_net_enable_vq(n, vq);
 942                 if (r)
 943                         goto err_used;
 944
 945                 oldubufs = nvq->ubufs;
 946                 nvq->ubufs = ubufs;
 947
 948                 n->tx_packets = 0;
 949                 n->tx_zcopy_err = 0;
 950                 n->tx_flush = false;
 951         }
 952
 953         mutex_unlock(&vq->mutex);
 954
 955         if (oldubufs) {
 956                 vhost_ubuf_put_and_wait(oldubufs);
 957                 mutex_lock(&vq->mutex);
 958                 vhost_zerocopy_signal_used(n, vq);
 959                 mutex_unlock(&vq->mutex);
 960         }
 961
 962         if (oldsock) {
 963                 vhost_net_flush_vq(n, index);
 964                 fput(oldsock->file);
 965         }
 966
 967         mutex_unlock(&n->dev.mutex);
 968         return 0;
 969
 970 err_used:
 971         rcu_assign_pointer(vq->private_data, oldsock);
 972         vhost_net_enable_vq(n, vq);
 973         if (ubufs)
 974                 vhost_ubuf_put_and_wait(ubufs);
 975 err_ubufs:
 976         fput(sock->file);
 977 err_vq:
 978         mutex_unlock(&vq->mutex);
 979 err:
 980         mutex_unlock(&n->dev.mutex);
 981         return r;
 982 }
 983
 984 static long vhost_net_reset_owner(struct vhost_net *n)
 985 {
 986         struct socket *tx_sock = NULL;
 987         struct socket *rx_sock = NULL;
 988         long err;
 989         struct vhost_memory *memory;
 990
 991         mutex_lock(&n->dev.mutex);
 992         err = vhost_dev_check_owner(&n->dev);
 993         if (err)
 994                 goto done;
 995         memory = vhost_dev_reset_owner_prepare();
 996         if (!memory) {
 997                 err = -ENOMEM;
 998                 goto done;
 999         }
1000         vhost_net_stop(n, &tx_sock, &rx_sock);
1001         vhost_net_flush(n);
1002         vhost_dev_reset_owner(&n->dev, memory);
1003         vhost_net_vq_reset(n);
1004 done:
1005         mutex_unlock(&n->dev.mutex);
1006         if (tx_sock)
1007                 fput(tx_sock->file);
1008         if (rx_sock)
1009                 fput(rx_sock->file);
1010         return err;
1011 }
1012
1013 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1014 {
1015         size_t vhost_hlen, sock_hlen, hdr_len;
1016         int i;
1017
1018         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1019                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1020                         sizeof(struct virtio_net_hdr);
1021         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1022                 /* vhost provides vnet_hdr */
1023                 vhost_hlen = hdr_len;
1024                 sock_hlen = 0;
1025         } else {
1026                 /* socket provides vnet_hdr */
1027                 vhost_hlen = 0;
1028                 sock_hlen = hdr_len;
1029         }
1030         mutex_lock(&n->dev.mutex);
1031         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1032             !vhost_log_access_ok(&n->dev)) {
1033                 mutex_unlock(&n->dev.mutex);
1034                 return -EFAULT;
1035         }
1036         n->dev.acked_features = features;
1037         smp_wmb();
1038         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1039                 mutex_lock(&n->vqs[i].vq.mutex);
1040                 n->vqs[i].vhost_hlen = vhost_hlen;
1041                 n->vqs[i].sock_hlen = sock_hlen;
1042                 mutex_unlock(&n->vqs[i].vq.mutex);
1043         }
1044         vhost_net_flush(n);
1045         mutex_unlock(&n->dev.mutex);
1046         return 0;
1047 }
1048
1049 static long vhost_net_set_owner(struct vhost_net *n)
1050 {
1051         int r;
1052
1053         mutex_lock(&n->dev.mutex);
1054         r = vhost_net_set_ubuf_info(n);
1055         if (r)
1056                 goto out;
1057         r = vhost_dev_set_owner(&n->dev);
1058         if (r)
1059                 vhost_net_clear_ubuf_info(n);
1060         vhost_net_flush(n);
1061 out:
1062         mutex_unlock(&n->dev.mutex);
1063         return r;
1064 }
1065
1066 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1067                             unsigned long arg)
1068 {
1069         struct vhost_net *n = f->private_data;
1070         void __user *argp = (void __user *)arg;
1071         u64 __user *featurep = argp;
1072         struct vhost_vring_file backend;
1073         u64 features;
1074         int r;
1075
1076         switch (ioctl) {
1077         case VHOST_NET_SET_BACKEND:
1078                 if (copy_from_user(&backend, argp, sizeof backend))
1079                         return -EFAULT;
1080                 return vhost_net_set_backend(n, backend.index, backend.fd);
1081         case VHOST_GET_FEATURES:
1082                 features = VHOST_NET_FEATURES;
1083                 if (copy_to_user(featurep, &features, sizeof features))
1084                         return -EFAULT;
1085                 return 0;
1086         case VHOST_SET_FEATURES:
1087                 if (copy_from_user(&features, featurep, sizeof features))
1088                         return -EFAULT;
1089                 if (features & ~VHOST_NET_FEATURES)
1090                         return -EOPNOTSUPP;
1091                 return vhost_net_set_features(n, features);
1092         case VHOST_RESET_OWNER:
1093                 return vhost_net_reset_owner(n);
1094         case VHOST_SET_OWNER:
1095                 return vhost_net_set_owner(n);
1096         default:
1097                 mutex_lock(&n->dev.mutex);
1098                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1099                 if (r == -ENOIOCTLCMD)
1100                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1101                 else
1102                         vhost_net_flush(n);
1103                 mutex_unlock(&n->dev.mutex);
1104                 return r;
1105         }
1106 }
1107
1108 #ifdef CONFIG_COMPAT
1109 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1110                                    unsigned long arg)
1111 {
1112         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1113 }
1114 #endif
1115
1116 static const struct file_operations vhost_net_fops = {
1117         .owner          = THIS_MODULE,
1118         .release        = vhost_net_release,
1119         .unlocked_ioctl = vhost_net_ioctl,
1120 #ifdef CONFIG_COMPAT
1121         .compat_ioctl   = vhost_net_compat_ioctl,
1122 #endif
1123         .open           = vhost_net_open,
1124         .llseek         = noop_llseek,
1125 };
1126
1127 static struct miscdevice vhost_net_misc = {
1128         .minor = VHOST_NET_MINOR,
1129         .name = "vhost-net",
1130         .fops = &vhost_net_fops,
1131 };
1132
1133 static int vhost_net_init(void)
1134 {
1135         if (experimental_zcopytx)
1136                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
1137         return misc_register(&vhost_net_misc);
1138 }
1139 module_init(vhost_net_init);
1140
1141 static void vhost_net_exit(void)
1142 {
1143         misc_deregister(&vhost_net_misc);
1144 }
1145 module_exit(vhost_net_exit);
1146
1147 MODULE_VERSION("0.0.1");
1148 MODULE_LICENSE("GPL v2");
1149 MODULE_AUTHOR("Michael S. Tsirkin");
1150 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1151 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1152 MODULE_ALIAS("devname:vhost-net");