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/file.h>
  19 #include <linux/slab.h>
  20
  21 #include <linux/net.h>
  22 #include <linux/if_packet.h>
  23 #include <linux/if_arp.h>
  24 #include <linux/if_tun.h>
  25 #include <linux/if_macvlan.h>
  26 #include <linux/if_vlan.h>
  27
  28 #include <net/sock.h>
  29
  30 #include "vhost.h"
  31
  32 static int experimental_zcopytx = 1;
  33 module_param(experimental_zcopytx, int, 0444);
  34 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  35                                        " 1 -Enable; 0 - Disable");
  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_FEATURES = VHOST_FEATURES |
  62                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  63                          (1ULL << VIRTIO_NET_F_MRG_RXBUF),
  64 };
  65
  66 enum {
  67         VHOST_NET_VQ_RX = 0,
  68         VHOST_NET_VQ_TX = 1,
  69         VHOST_NET_VQ_MAX = 2,
  70 };
  71
  72 struct vhost_net_ubuf_ref {
  73         /* refcount follows semantics similar to kref:
  74          *  0: object is released
  75          *  1: no outstanding ubufs
  76          * >1: outstanding ubufs
  77          */
  78         atomic_t refcount;
  79         wait_queue_head_t wait;
  80         struct vhost_virtqueue *vq;
  81 };
  82
  83 struct vhost_net_virtqueue {
  84         struct vhost_virtqueue vq;
  85         /* hdr is used to store the virtio header.
  86          * Since each iovec has >= 1 byte length, we never need more than
  87          * header length entries to store the header. */
  88         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
  89         size_t vhost_hlen;
  90         size_t sock_hlen;
  91         /* vhost zerocopy support fields below: */
  92         /* last used idx for outstanding DMA zerocopy buffers */
  93         int upend_idx;
  94         /* first used idx for DMA done zerocopy buffers */
  95         int done_idx;
  96         /* an array of userspace buffers info */
  97         struct ubuf_info *ubuf_info;
  98         /* Reference counting for outstanding ubufs.
  99          * Protected by vq mutex. Writers must also take device mutex. */
 100         struct vhost_net_ubuf_ref *ubufs;
 101 };
 102
 103 struct vhost_net {
 104         struct vhost_dev dev;
 105         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 106         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 107         /* Number of TX recently submitted.
 108          * Protected by tx vq lock. */
 109         unsigned tx_packets;
 110         /* Number of times zerocopy TX recently failed.
 111          * Protected by tx vq lock. */
 112         unsigned tx_zcopy_err;
 113         /* Flush in progress. Protected by tx vq lock. */
 114         bool tx_flush;
 115 };
 116
 117 static unsigned vhost_net_zcopy_mask __read_mostly;
 118
 119 static void vhost_net_enable_zcopy(int vq)
 120 {
 121         vhost_net_zcopy_mask |= 0x1 << vq;
 122 }
 123
 124 static struct vhost_net_ubuf_ref *
 125 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 126 {
 127         struct vhost_net_ubuf_ref *ubufs;
 128         /* No zero copy backend? Nothing to count. */
 129         if (!zcopy)
 130                 return NULL;
 131         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 132         if (!ubufs)
 133                 return ERR_PTR(-ENOMEM);
 134         atomic_set(&ubufs->refcount, 1);
 135         init_waitqueue_head(&ubufs->wait);
 136         ubufs->vq = vq;
 137         return ubufs;
 138 }
 139
 140 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 141 {
 142         int r = atomic_sub_return(1, &ubufs->refcount);
 143         if (unlikely(!r))
 144                 wake_up(&ubufs->wait);
 145         return r;
 146 }
 147
 148 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 149 {
 150         vhost_net_ubuf_put(ubufs);
 151         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 152 }
 153
 154 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 155 {
 156         vhost_net_ubuf_put_and_wait(ubufs);
 157         kfree(ubufs);
 158 }
 159
 160 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 161 {
 162         int i;
 163
 164         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 165                 kfree(n->vqs[i].ubuf_info);
 166                 n->vqs[i].ubuf_info = NULL;
 167         }
 168 }
 169
 170 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 171 {
 172         bool zcopy;
 173         int i;
 174
 175         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 176                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 177                 if (!zcopy)
 178                         continue;
 179                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 180                                               UIO_MAXIOV, GFP_KERNEL);
 181                 if  (!n->vqs[i].ubuf_info)
 182                         goto err;
 183         }
 184         return 0;
 185
 186 err:
 187         vhost_net_clear_ubuf_info(n);
 188         return -ENOMEM;
 189 }
 190
 191 static void vhost_net_vq_reset(struct vhost_net *n)
 192 {
 193         int i;
 194
 195         vhost_net_clear_ubuf_info(n);
 196
 197         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 198                 n->vqs[i].done_idx = 0;
 199                 n->vqs[i].upend_idx = 0;
 200                 n->vqs[i].ubufs = NULL;
 201                 n->vqs[i].vhost_hlen = 0;
 202                 n->vqs[i].sock_hlen = 0;
 203         }
 204
 205 }
 206
 207 static void vhost_net_tx_packet(struct vhost_net *net)
 208 {
 209         ++net->tx_packets;
 210         if (net->tx_packets < 1024)
 211                 return;
 212         net->tx_packets = 0;
 213         net->tx_zcopy_err = 0;
 214 }
 215
 216 static void vhost_net_tx_err(struct vhost_net *net)
 217 {
 218         ++net->tx_zcopy_err;
 219 }
 220
 221 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 222 {
 223         /* TX flush waits for outstanding DMAs to be done.
 224          * Don't start new DMAs.
 225          */
 226         return !net->tx_flush &&
 227                 net->tx_packets / 64 >= net->tx_zcopy_err;
 228 }
 229
 230 static bool vhost_sock_zcopy(struct socket *sock)
 231 {
 232         return unlikely(experimental_zcopytx) &&
 233                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 234 }
 235
 236 /* Pop first len bytes from iovec. Return number of segments used. */
 237 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 238                           size_t len, int iov_count)
 239 {
 240         int seg = 0;
 241         size_t size;
 242
 243         while (len && seg < iov_count) {
 244                 size = min(from->iov_len, len);
 245                 to->iov_base = from->iov_base;
 246                 to->iov_len = size;
 247                 from->iov_len -= size;
 248                 from->iov_base += size;
 249                 len -= size;
 250                 ++from;
 251                 ++to;
 252                 ++seg;
 253         }
 254         return seg;
 255 }
 256 /* Copy iovec entries for len bytes from iovec. */
 257 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 258                            size_t len, int iovcount)
 259 {
 260         int seg = 0;
 261         size_t size;
 262
 263         while (len && seg < iovcount) {
 264                 size = min(from->iov_len, len);
 265                 to->iov_base = from->iov_base;
 266                 to->iov_len = size;
 267                 len -= size;
 268                 ++from;
 269                 ++to;
 270                 ++seg;
 271         }
 272 }
 273
 274 /* In case of DMA done not in order in lower device driver for some reason.
 275  * upend_idx is used to track end of used idx, done_idx is used to track head
 276  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 277  * guest used idx.
 278  */
 279 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 280                                        struct vhost_virtqueue *vq)
 281 {
 282         struct vhost_net_virtqueue *nvq =
 283                 container_of(vq, struct vhost_net_virtqueue, vq);
 284         int i, add;
 285         int j = 0;
 286
 287         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 288                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 289                         vhost_net_tx_err(net);
 290                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 291                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 292                         ++j;
 293                 } else
 294                         break;
 295         }
 296         while (j) {
 297                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 298                 vhost_add_used_and_signal_n(vq->dev, vq,
 299                                             &vq->heads[nvq->done_idx], add);
 300                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 301                 j -= add;
 302         }
 303 }
 304
 305 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 306 {
 307         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 308         struct vhost_virtqueue *vq = ubufs->vq;
 309         int cnt;
 310
 311         rcu_read_lock_bh();
 312
 313         /* set len to mark this desc buffers done DMA */
 314         vq->heads[ubuf->desc].len = success ?
 315                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 316         cnt = vhost_net_ubuf_put(ubufs);
 317
 318         /*
 319          * Trigger polling thread if guest stopped submitting new buffers:
 320          * in this case, the refcount after decrement will eventually reach 1.
 321          * We also trigger polling periodically after each 16 packets
 322          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 323          * less than 10% of times).
 324          */
 325         if (cnt <= 1 || !(cnt % 16))
 326                 vhost_poll_queue(&vq->poll);
 327
 328         rcu_read_unlock_bh();
 329 }
 330
 331 /* Expects to be always run from workqueue - which acts as
 332  * read-size critical section for our kind of RCU. */
 333 static void handle_tx(struct vhost_net *net)
 334 {
 335         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 336         struct vhost_virtqueue *vq = &nvq->vq;
 337         unsigned out, in, s;
 338         int head;
 339         struct msghdr msg = {
 340                 .msg_name = NULL,
 341                 .msg_namelen = 0,
 342                 .msg_control = NULL,
 343                 .msg_controllen = 0,
 344                 .msg_iov = vq->iov,
 345                 .msg_flags = MSG_DONTWAIT,
 346         };
 347         size_t len, total_len = 0;
 348         int err;
 349         size_t hdr_size;
 350         struct socket *sock;
 351         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 352         bool zcopy, zcopy_used;
 353
 354         mutex_lock(&vq->mutex);
 355         sock = vq->private_data;
 356         if (!sock)
 357                 goto out;
 358
 359         vhost_disable_notify(&net->dev, vq);
 360
 361         hdr_size = nvq->vhost_hlen;
 362         zcopy = nvq->ubufs;
 363
 364         for (;;) {
 365                 /* Release DMAs done buffers first */
 366                 if (zcopy)
 367                         vhost_zerocopy_signal_used(net, vq);
 368
 369                 /* If more outstanding DMAs, queue the work.
 370                  * Handle upend_idx wrap around
 371                  */
 372                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
 373                               % UIO_MAXIOV == nvq->done_idx))
 374                         break;
 375
 376                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 377                                          ARRAY_SIZE(vq->iov),
 378                                          &out, &in,
 379                                          NULL, NULL);
 380                 /* On error, stop handling until the next kick. */
 381                 if (unlikely(head < 0))
 382                         break;
 383                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 384                 if (head == vq->num) {
 385                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 386                                 vhost_disable_notify(&net->dev, vq);
 387                                 continue;
 388                         }
 389                         break;
 390                 }
 391                 if (in) {
 392                         vq_err(vq, "Unexpected descriptor format for TX: "
 393                                "out %d, int %d\n", out, in);
 394                         break;
 395                 }
 396                 /* Skip header. TODO: support TSO. */
 397                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 398                 msg.msg_iovlen = out;
 399                 len = iov_length(vq->iov, out);
 400                 /* Sanity check */
 401                 if (!len) {
 402                         vq_err(vq, "Unexpected header len for TX: "
 403                                "%zd expected %zd\n",
 404                                iov_length(nvq->hdr, s), hdr_size);
 405                         break;
 406                 }
 407
 408                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 409                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 410                                       nvq->done_idx
 411                                    && vhost_net_tx_select_zcopy(net);
 412
 413                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 414                 if (zcopy_used) {
 415                         struct ubuf_info *ubuf;
 416                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 417
 418                         vq->heads[nvq->upend_idx].id = head;
 419                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 420                         ubuf->callback = vhost_zerocopy_callback;
 421                         ubuf->ctx = nvq->ubufs;
 422                         ubuf->desc = nvq->upend_idx;
 423                         msg.msg_control = ubuf;
 424                         msg.msg_controllen = sizeof(ubuf);
 425                         ubufs = nvq->ubufs;
 426                         atomic_inc(&ubufs->refcount);
 427                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 428                 } else {
 429                         msg.msg_control = NULL;
 430                         ubufs = NULL;
 431                 }
 432                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 433                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 434                 if (unlikely(err < 0)) {
 435                         if (zcopy_used) {
 436                                 vhost_net_ubuf_put(ubufs);
 437                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 438                                         % UIO_MAXIOV;
 439                         }
 440                         vhost_discard_vq_desc(vq, 1);
 441                         break;
 442                 }
 443                 if (err != len)
 444                         pr_debug("Truncated TX packet: "
 445                                  " len %d != %zd\n", err, len);
 446                 if (!zcopy_used)
 447                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 448                 else
 449                         vhost_zerocopy_signal_used(net, vq);
 450                 total_len += len;
 451                 vhost_net_tx_packet(net);
 452                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 453                         vhost_poll_queue(&vq->poll);
 454                         break;
 455                 }
 456         }
 457 out:
 458         mutex_unlock(&vq->mutex);
 459 }
 460
 461 static int peek_head_len(struct sock *sk)
 462 {
 463         struct sk_buff *head;
 464         int len = 0;
 465         unsigned long flags;
 466
 467         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 468         head = skb_peek(&sk->sk_receive_queue);
 469         if (likely(head)) {
 470                 len = head->len;
 471                 if (vlan_tx_tag_present(head))
 472                         len += VLAN_HLEN;
 473         }
 474
 475         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 476         return len;
 477 }
 478
 479 /* This is a multi-buffer version of vhost_get_desc, that works if
 480  *      vq has read descriptors only.
 481  * @vq          - the relevant virtqueue
 482  * @datalen     - data length we'll be reading
 483  * @iovcount    - returned count of io vectors we fill
 484  * @log         - vhost log
 485  * @log_num     - log offset
 486  * @quota       - headcount quota, 1 for big buffer
 487  *      returns number of buffer heads allocated, negative on error
 488  */
 489 static int get_rx_bufs(struct vhost_virtqueue *vq,
 490                        struct vring_used_elem *heads,
 491                        int datalen,
 492                        unsigned *iovcount,
 493                        struct vhost_log *log,
 494                        unsigned *log_num,
 495                        unsigned int quota)
 496 {
 497         unsigned int out, in;
 498         int seg = 0;
 499         int headcount = 0;
 500         unsigned d;
 501         int r, nlogs = 0;
 502
 503         while (datalen > 0 && headcount < quota) {
 504                 if (unlikely(seg >= UIO_MAXIOV)) {
 505                         r = -ENOBUFS;
 506                         goto err;
 507                 }
 508                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 509                                       ARRAY_SIZE(vq->iov) - seg, &out,
 510                                       &in, log, log_num);
 511                 if (d == vq->num) {
 512                         r = 0;
 513                         goto err;
 514                 }
 515                 if (unlikely(out || in <= 0)) {
 516                         vq_err(vq, "unexpected descriptor format for RX: "
 517                                 "out %d, in %d\n", out, in);
 518                         r = -EINVAL;
 519                         goto err;
 520                 }
 521                 if (unlikely(log)) {
 522                         nlogs += *log_num;
 523                         log += *log_num;
 524                 }
 525                 heads[headcount].id = d;
 526                 heads[headcount].len = iov_length(vq->iov + seg, in);
 527                 datalen -= heads[headcount].len;
 528                 ++headcount;
 529                 seg += in;
 530         }
 531         heads[headcount - 1].len += datalen;
 532         *iovcount = seg;
 533         if (unlikely(log))
 534                 *log_num = nlogs;
 535
 536         /* Detect overrun */
 537         if (unlikely(datalen > 0)) {
 538                 r = UIO_MAXIOV + 1;
 539                 goto err;
 540         }
 541         return headcount;
 542 err:
 543         vhost_discard_vq_desc(vq, headcount);
 544         return r;
 545 }
 546
 547 /* Expects to be always run from workqueue - which acts as
 548  * read-size critical section for our kind of RCU. */
 549 static void handle_rx(struct vhost_net *net)
 550 {
 551         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 552         struct vhost_virtqueue *vq = &nvq->vq;
 553         unsigned uninitialized_var(in), log;
 554         struct vhost_log *vq_log;
 555         struct msghdr msg = {
 556                 .msg_name = NULL,
 557                 .msg_namelen = 0,
 558                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 559                 .msg_controllen = 0,
 560                 .msg_iov = vq->iov,
 561                 .msg_flags = MSG_DONTWAIT,
 562         };
 563         struct virtio_net_hdr_mrg_rxbuf hdr = {
 564                 .hdr.flags = 0,
 565                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 566         };
 567         size_t total_len = 0;
 568         int err, mergeable;
 569         s16 headcount;
 570         size_t vhost_hlen, sock_hlen;
 571         size_t vhost_len, sock_len;
 572         struct socket *sock;
 573
 574         mutex_lock(&vq->mutex);
 575         sock = vq->private_data;
 576         if (!sock)
 577                 goto out;
 578         vhost_disable_notify(&net->dev, vq);
 579
 580         vhost_hlen = nvq->vhost_hlen;
 581         sock_hlen = nvq->sock_hlen;
 582
 583         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 584                 vq->log : NULL;
 585         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 586
 587         while ((sock_len = peek_head_len(sock->sk))) {
 588                 sock_len += sock_hlen;
 589                 vhost_len = sock_len + vhost_hlen;
 590                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 591                                         &in, vq_log, &log,
 592                                         likely(mergeable) ? UIO_MAXIOV : 1);
 593                 /* On error, stop handling until the next kick. */
 594                 if (unlikely(headcount < 0))
 595                         break;
 596                 /* On overrun, truncate and discard */
 597                 if (unlikely(headcount > UIO_MAXIOV)) {
 598                         msg.msg_iovlen = 1;
 599                         err = sock->ops->recvmsg(NULL, sock, &msg,
 600                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 601                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 602                         continue;
 603                 }
 604                 /* OK, now we need to know about added descriptors. */
 605                 if (!headcount) {
 606                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 607                                 /* They have slipped one in as we were
 608                                  * doing that: check again. */
 609                                 vhost_disable_notify(&net->dev, vq);
 610                                 continue;
 611                         }
 612                         /* Nothing new?  Wait for eventfd to tell us
 613                          * they refilled. */
 614                         break;
 615                 }
 616                 /* We don't need to be notified again. */
 617                 if (unlikely((vhost_hlen)))
 618                         /* Skip header. TODO: support TSO. */
 619                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 620                 else
 621                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 622                          * needed because recvmsg can modify msg_iov. */
 623                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 624                 msg.msg_iovlen = in;
 625                 err = sock->ops->recvmsg(NULL, sock, &msg,
 626                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 627                 /* Userspace might have consumed the packet meanwhile:
 628                  * it's not supposed to do this usually, but might be hard
 629                  * to prevent. Discard data we got (if any) and keep going. */
 630                 if (unlikely(err != sock_len)) {
 631                         pr_debug("Discarded rx packet: "
 632                                  " len %d, expected %zd\n", err, sock_len);
 633                         vhost_discard_vq_desc(vq, headcount);
 634                         continue;
 635                 }
 636                 if (unlikely(vhost_hlen) &&
 637                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 638                                       vhost_hlen)) {
 639                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 640                                vq->iov->iov_base);
 641                         break;
 642                 }
 643                 /* TODO: Should check and handle checksum. */
 644                 if (likely(mergeable) &&
 645                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 646                                       offsetof(typeof(hdr), num_buffers),
 647                                       sizeof hdr.num_buffers)) {
 648                         vq_err(vq, "Failed num_buffers write");
 649                         vhost_discard_vq_desc(vq, headcount);
 650                         break;
 651                 }
 652                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 653                                             headcount);
 654                 if (unlikely(vq_log))
 655                         vhost_log_write(vq, vq_log, log, vhost_len);
 656                 total_len += vhost_len;
 657                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 658                         vhost_poll_queue(&vq->poll);
 659                         break;
 660                 }
 661         }
 662 out:
 663         mutex_unlock(&vq->mutex);
 664 }
 665
 666 static void handle_tx_kick(struct vhost_work *work)
 667 {
 668         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 669                                                   poll.work);
 670         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 671
 672         handle_tx(net);
 673 }
 674
 675 static void handle_rx_kick(struct vhost_work *work)
 676 {
 677         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 678                                                   poll.work);
 679         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 680
 681         handle_rx(net);
 682 }
 683
 684 static void handle_tx_net(struct vhost_work *work)
 685 {
 686         struct vhost_net *net = container_of(work, struct vhost_net,
 687                                              poll[VHOST_NET_VQ_TX].work);
 688         handle_tx(net);
 689 }
 690
 691 static void handle_rx_net(struct vhost_work *work)
 692 {
 693         struct vhost_net *net = container_of(work, struct vhost_net,
 694                                              poll[VHOST_NET_VQ_RX].work);
 695         handle_rx(net);
 696 }
 697
 698 static int vhost_net_open(struct inode *inode, struct file *f)
 699 {
 700         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 701         struct vhost_dev *dev;
 702         struct vhost_virtqueue **vqs;
 703         int i;
 704
 705         if (!n)
 706                 return -ENOMEM;
 707         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 708         if (!vqs) {
 709                 kfree(n);
 710                 return -ENOMEM;
 711         }
 712
 713         dev = &n->dev;
 714         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 715         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 716         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 717         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 718         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 719                 n->vqs[i].ubufs = NULL;
 720                 n->vqs[i].ubuf_info = NULL;
 721                 n->vqs[i].upend_idx = 0;
 722                 n->vqs[i].done_idx = 0;
 723                 n->vqs[i].vhost_hlen = 0;
 724                 n->vqs[i].sock_hlen = 0;
 725         }
 726         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 727
 728         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 729         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 730
 731         f->private_data = n;
 732
 733         return 0;
 734 }
 735
 736 static void vhost_net_disable_vq(struct vhost_net *n,
 737                                  struct vhost_virtqueue *vq)
 738 {
 739         struct vhost_net_virtqueue *nvq =
 740                 container_of(vq, struct vhost_net_virtqueue, vq);
 741         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 742         if (!vq->private_data)
 743                 return;
 744         vhost_poll_stop(poll);
 745 }
 746
 747 static int vhost_net_enable_vq(struct vhost_net *n,
 748                                 struct vhost_virtqueue *vq)
 749 {
 750         struct vhost_net_virtqueue *nvq =
 751                 container_of(vq, struct vhost_net_virtqueue, vq);
 752         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 753         struct socket *sock;
 754
 755         sock = vq->private_data;
 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 = vq->private_data;
 769         vhost_net_disable_vq(n, vq);
 770         vq->private_data = NULL;
 771         mutex_unlock(&vq->mutex);
 772         return sock;
 773 }
 774
 775 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 776                            struct socket **rx_sock)
 777 {
 778         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 779         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 780 }
 781
 782 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 783 {
 784         vhost_poll_flush(n->poll + index);
 785         vhost_poll_flush(&n->vqs[index].vq.poll);
 786 }
 787
 788 static void vhost_net_flush(struct vhost_net *n)
 789 {
 790         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 791         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 792         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 793                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 794                 n->tx_flush = true;
 795                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 796                 /* Wait for all lower device DMAs done. */
 797                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 798                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 799                 n->tx_flush = false;
 800                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 801                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 802         }
 803 }
 804
 805 static int vhost_net_release(struct inode *inode, struct file *f)
 806 {
 807         struct vhost_net *n = f->private_data;
 808         struct socket *tx_sock;
 809         struct socket *rx_sock;
 810
 811         vhost_net_stop(n, &tx_sock, &rx_sock);
 812         vhost_net_flush(n);
 813         vhost_dev_stop(&n->dev);
 814         vhost_dev_cleanup(&n->dev, false);
 815         vhost_net_vq_reset(n);
 816         if (tx_sock)
 817                 fput(tx_sock->file);
 818         if (rx_sock)
 819                 fput(rx_sock->file);
 820         /* Make sure no callbacks are outstanding */
 821         synchronize_rcu_bh();
 822         /* We do an extra flush before freeing memory,
 823          * since jobs can re-queue themselves. */
 824         vhost_net_flush(n);
 825         kfree(n->dev.vqs);
 826         kfree(n);
 827         return 0;
 828 }
 829
 830 static struct socket *get_raw_socket(int fd)
 831 {
 832         struct {
 833                 struct sockaddr_ll sa;
 834                 char  buf[MAX_ADDR_LEN];
 835         } uaddr;
 836         int uaddr_len = sizeof uaddr, r;
 837         struct socket *sock = sockfd_lookup(fd, &r);
 838
 839         if (!sock)
 840                 return ERR_PTR(-ENOTSOCK);
 841
 842         /* Parameter checking */
 843         if (sock->sk->sk_type != SOCK_RAW) {
 844                 r = -ESOCKTNOSUPPORT;
 845                 goto err;
 846         }
 847
 848         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 849                                &uaddr_len, 0);
 850         if (r)
 851                 goto err;
 852
 853         if (uaddr.sa.sll_family != AF_PACKET) {
 854                 r = -EPFNOSUPPORT;
 855                 goto err;
 856         }
 857         return sock;
 858 err:
 859         fput(sock->file);
 860         return ERR_PTR(r);
 861 }
 862
 863 static struct socket *get_tap_socket(int fd)
 864 {
 865         struct file *file = fget(fd);
 866         struct socket *sock;
 867
 868         if (!file)
 869                 return ERR_PTR(-EBADF);
 870         sock = tun_get_socket(file);
 871         if (!IS_ERR(sock))
 872                 return sock;
 873         sock = macvtap_get_socket(file);
 874         if (IS_ERR(sock))
 875                 fput(file);
 876         return sock;
 877 }
 878
 879 static struct socket *get_socket(int fd)
 880 {
 881         struct socket *sock;
 882
 883         /* special case to disable backend */
 884         if (fd == -1)
 885                 return NULL;
 886         sock = get_raw_socket(fd);
 887         if (!IS_ERR(sock))
 888                 return sock;
 889         sock = get_tap_socket(fd);
 890         if (!IS_ERR(sock))
 891                 return sock;
 892         return ERR_PTR(-ENOTSOCK);
 893 }
 894
 895 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 896 {
 897         struct socket *sock, *oldsock;
 898         struct vhost_virtqueue *vq;
 899         struct vhost_net_virtqueue *nvq;
 900         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 901         int r;
 902
 903         mutex_lock(&n->dev.mutex);
 904         r = vhost_dev_check_owner(&n->dev);
 905         if (r)
 906                 goto err;
 907
 908         if (index >= VHOST_NET_VQ_MAX) {
 909                 r = -ENOBUFS;
 910                 goto err;
 911         }
 912         vq = &n->vqs[index].vq;
 913         nvq = &n->vqs[index];
 914         mutex_lock(&vq->mutex);
 915
 916         /* Verify that ring has been setup correctly. */
 917         if (!vhost_vq_access_ok(vq)) {
 918                 r = -EFAULT;
 919                 goto err_vq;
 920         }
 921         sock = get_socket(fd);
 922         if (IS_ERR(sock)) {
 923                 r = PTR_ERR(sock);
 924                 goto err_vq;
 925         }
 926
 927         /* start polling new socket */
 928         oldsock = vq->private_data;
 929         if (sock != oldsock) {
 930                 ubufs = vhost_net_ubuf_alloc(vq,
 931                                              sock && vhost_sock_zcopy(sock));
 932                 if (IS_ERR(ubufs)) {
 933                         r = PTR_ERR(ubufs);
 934                         goto err_ubufs;
 935                 }
 936
 937                 vhost_net_disable_vq(n, vq);
 938                 vq->private_data = sock;
 939                 r = vhost_init_used(vq);
 940                 if (r)
 941                         goto err_used;
 942                 r = vhost_net_enable_vq(n, vq);
 943                 if (r)
 944                         goto err_used;
 945
 946                 oldubufs = nvq->ubufs;
 947                 nvq->ubufs = ubufs;
 948
 949                 n->tx_packets = 0;
 950                 n->tx_zcopy_err = 0;
 951                 n->tx_flush = false;
 952         }
 953
 954         mutex_unlock(&vq->mutex);
 955
 956         if (oldubufs) {
 957                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 958                 mutex_lock(&vq->mutex);
 959                 vhost_zerocopy_signal_used(n, vq);
 960                 mutex_unlock(&vq->mutex);
 961         }
 962
 963         if (oldsock) {
 964                 vhost_net_flush_vq(n, index);
 965                 fput(oldsock->file);
 966         }
 967
 968         mutex_unlock(&n->dev.mutex);
 969         return 0;
 970
 971 err_used:
 972         vq->private_data = oldsock;
 973         vhost_net_enable_vq(n, vq);
 974         if (ubufs)
 975                 vhost_net_ubuf_put_wait_and_free(ubufs);
 976 err_ubufs:
 977         fput(sock->file);
 978 err_vq:
 979         mutex_unlock(&vq->mutex);
 980 err:
 981         mutex_unlock(&n->dev.mutex);
 982         return r;
 983 }
 984
 985 static long vhost_net_reset_owner(struct vhost_net *n)
 986 {
 987         struct socket *tx_sock = NULL;
 988         struct socket *rx_sock = NULL;
 989         long err;
 990         struct vhost_memory *memory;
 991
 992         mutex_lock(&n->dev.mutex);
 993         err = vhost_dev_check_owner(&n->dev);
 994         if (err)
 995                 goto done;
 996         memory = vhost_dev_reset_owner_prepare();
 997         if (!memory) {
 998                 err = -ENOMEM;
 999                 goto done;
1000         }
1001         vhost_net_stop(n, &tx_sock, &rx_sock);
1002         vhost_net_flush(n);
1003         vhost_dev_reset_owner(&n->dev, memory);
1004         vhost_net_vq_reset(n);
1005 done:
1006         mutex_unlock(&n->dev.mutex);
1007         if (tx_sock)
1008                 fput(tx_sock->file);
1009         if (rx_sock)
1010                 fput(rx_sock->file);
1011         return err;
1012 }
1013
1014 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1015 {
1016         size_t vhost_hlen, sock_hlen, hdr_len;
1017         int i;
1018
1019         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1020                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1021                         sizeof(struct virtio_net_hdr);
1022         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1023                 /* vhost provides vnet_hdr */
1024                 vhost_hlen = hdr_len;
1025                 sock_hlen = 0;
1026         } else {
1027                 /* socket provides vnet_hdr */
1028                 vhost_hlen = 0;
1029                 sock_hlen = hdr_len;
1030         }
1031         mutex_lock(&n->dev.mutex);
1032         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1033             !vhost_log_access_ok(&n->dev)) {
1034                 mutex_unlock(&n->dev.mutex);
1035                 return -EFAULT;
1036         }
1037         n->dev.acked_features = features;
1038         smp_wmb();
1039         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1040                 mutex_lock(&n->vqs[i].vq.mutex);
1041                 n->vqs[i].vhost_hlen = vhost_hlen;
1042                 n->vqs[i].sock_hlen = sock_hlen;
1043                 mutex_unlock(&n->vqs[i].vq.mutex);
1044         }
1045         vhost_net_flush(n);
1046         mutex_unlock(&n->dev.mutex);
1047         return 0;
1048 }
1049
1050 static long vhost_net_set_owner(struct vhost_net *n)
1051 {
1052         int r;
1053
1054         mutex_lock(&n->dev.mutex);
1055         if (vhost_dev_has_owner(&n->dev)) {
1056                 r = -EBUSY;
1057                 goto out;
1058         }
1059         r = vhost_net_set_ubuf_info(n);
1060         if (r)
1061                 goto out;
1062         r = vhost_dev_set_owner(&n->dev);
1063         if (r)
1064                 vhost_net_clear_ubuf_info(n);
1065         vhost_net_flush(n);
1066 out:
1067         mutex_unlock(&n->dev.mutex);
1068         return r;
1069 }
1070
1071 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1072                             unsigned long arg)
1073 {
1074         struct vhost_net *n = f->private_data;
1075         void __user *argp = (void __user *)arg;
1076         u64 __user *featurep = argp;
1077         struct vhost_vring_file backend;
1078         u64 features;
1079         int r;
1080
1081         switch (ioctl) {
1082         case VHOST_NET_SET_BACKEND:
1083                 if (copy_from_user(&backend, argp, sizeof backend))
1084                         return -EFAULT;
1085                 return vhost_net_set_backend(n, backend.index, backend.fd);
1086         case VHOST_GET_FEATURES:
1087                 features = VHOST_NET_FEATURES;
1088                 if (copy_to_user(featurep, &features, sizeof features))
1089                         return -EFAULT;
1090                 return 0;
1091         case VHOST_SET_FEATURES:
1092                 if (copy_from_user(&features, featurep, sizeof features))
1093                         return -EFAULT;
1094                 if (features & ~VHOST_NET_FEATURES)
1095                         return -EOPNOTSUPP;
1096                 return vhost_net_set_features(n, features);
1097         case VHOST_RESET_OWNER:
1098                 return vhost_net_reset_owner(n);
1099         case VHOST_SET_OWNER:
1100                 return vhost_net_set_owner(n);
1101         default:
1102                 mutex_lock(&n->dev.mutex);
1103                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1104                 if (r == -ENOIOCTLCMD)
1105                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1106                 else
1107                         vhost_net_flush(n);
1108                 mutex_unlock(&n->dev.mutex);
1109                 return r;
1110         }
1111 }
1112
1113 #ifdef CONFIG_COMPAT
1114 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1115                                    unsigned long arg)
1116 {
1117         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1118 }
1119 #endif
1120
1121 static const struct file_operations vhost_net_fops = {
1122         .owner          = THIS_MODULE,
1123         .release        = vhost_net_release,
1124         .unlocked_ioctl = vhost_net_ioctl,
1125 #ifdef CONFIG_COMPAT
1126         .compat_ioctl   = vhost_net_compat_ioctl,
1127 #endif
1128         .open           = vhost_net_open,
1129         .llseek         = noop_llseek,
1130 };
1131
1132 static struct miscdevice vhost_net_misc = {
1133         .minor = VHOST_NET_MINOR,
1134         .name = "vhost-net",
1135         .fops = &vhost_net_fops,
1136 };
1137
1138 static int vhost_net_init(void)
1139 {
1140         if (experimental_zcopytx)
1141                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1142         return misc_register(&vhost_net_misc);
1143 }
1144 module_init(vhost_net_init);
1145
1146 static void vhost_net_exit(void)
1147 {
1148         misc_deregister(&vhost_net_misc);
1149 }
1150 module_exit(vhost_net_exit);
1151
1152 MODULE_VERSION("0.0.1");
1153 MODULE_LICENSE("GPL v2");
1154 MODULE_AUTHOR("Michael S. Tsirkin");
1155 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1156 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1157 MODULE_ALIAS("devname:vhost-net");