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 #include <linux/vmalloc.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 = 0;
  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 packets transferred before requeueing the job.
  43  * Using this limit prevents one virtqueue from starving others with small
  44  * pkts.
  45  */
  46 #define VHOST_NET_PKT_WEIGHT 256
  47
  48 /* MAX number of TX used buffers for outstanding zerocopy */
  49 #define VHOST_MAX_PEND 128
  50 #define VHOST_GOODCOPY_LEN 256
  51
  52 /*
  53  * For transmit, used buffer len is unused; we override it to track buffer
  54  * status internally; used for zerocopy tx only.
  55  */
  56 /* Lower device DMA failed */
  57 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  58 /* Lower device DMA done */
  59 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  60 /* Lower device DMA in progress */
  61 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  62 /* Buffer unused */
  63 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  64
  65 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  66
  67 enum {
  68         VHOST_NET_FEATURES = VHOST_FEATURES |
  69                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  70                          (1ULL << VIRTIO_NET_F_MRG_RXBUF)
  71 };
  72
  73 enum {
  74         VHOST_NET_VQ_RX = 0,
  75         VHOST_NET_VQ_TX = 1,
  76         VHOST_NET_VQ_MAX = 2,
  77 };
  78
  79 struct vhost_net_ubuf_ref {
  80         /* refcount follows semantics similar to kref:
  81          *  0: object is released
  82          *  1: no outstanding ubufs
  83          * >1: outstanding ubufs
  84          */
  85         atomic_t refcount;
  86         wait_queue_head_t wait;
  87         struct vhost_virtqueue *vq;
  88 };
  89
  90 struct vhost_net_virtqueue {
  91         struct vhost_virtqueue vq;
  92         size_t vhost_hlen;
  93         size_t sock_hlen;
  94         /* vhost zerocopy support fields below: */
  95         /* last used idx for outstanding DMA zerocopy buffers */
  96         int upend_idx;
  97         /* first used idx for DMA done zerocopy buffers */
  98         int done_idx;
  99         /* an array of userspace buffers info */
 100         struct ubuf_info *ubuf_info;
 101         /* Reference counting for outstanding ubufs.
 102          * Protected by vq mutex. Writers must also take device mutex. */
 103         struct vhost_net_ubuf_ref *ubufs;
 104 };
 105
 106 struct vhost_net {
 107         struct vhost_dev dev;
 108         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 109         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 110         /* Number of TX recently submitted.
 111          * Protected by tx vq lock. */
 112         unsigned tx_packets;
 113         /* Number of times zerocopy TX recently failed.
 114          * Protected by tx vq lock. */
 115         unsigned tx_zcopy_err;
 116         /* Flush in progress. Protected by tx vq lock. */
 117         bool tx_flush;
 118 };
 119
 120 static unsigned vhost_net_zcopy_mask __read_mostly;
 121
 122 static void vhost_net_enable_zcopy(int vq)
 123 {
 124         vhost_net_zcopy_mask |= 0x1 << vq;
 125 }
 126
 127 static struct vhost_net_ubuf_ref *
 128 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 129 {
 130         struct vhost_net_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         atomic_set(&ubufs->refcount, 1);
 138         init_waitqueue_head(&ubufs->wait);
 139         ubufs->vq = vq;
 140         return ubufs;
 141 }
 142
 143 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 144 {
 145         int r = atomic_sub_return(1, &ubufs->refcount);
 146         if (unlikely(!r))
 147                 wake_up(&ubufs->wait);
 148         return r;
 149 }
 150
 151 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 152 {
 153         vhost_net_ubuf_put(ubufs);
 154         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 155 }
 156
 157 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 158 {
 159         vhost_net_ubuf_put_and_wait(ubufs);
 160         kfree(ubufs);
 161 }
 162
 163 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 164 {
 165         int i;
 166
 167         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 168                 kfree(n->vqs[i].ubuf_info);
 169                 n->vqs[i].ubuf_info = NULL;
 170         }
 171 }
 172
 173 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 174 {
 175         bool zcopy;
 176         int i;
 177
 178         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 179                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 180                 if (!zcopy)
 181                         continue;
 182                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 183                                               UIO_MAXIOV, GFP_KERNEL);
 184                 if  (!n->vqs[i].ubuf_info)
 185                         goto err;
 186         }
 187         return 0;
 188
 189 err:
 190         vhost_net_clear_ubuf_info(n);
 191         return -ENOMEM;
 192 }
 193
 194 static void vhost_net_vq_reset(struct vhost_net *n)
 195 {
 196         int i;
 197
 198         vhost_net_clear_ubuf_info(n);
 199
 200         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 201                 n->vqs[i].done_idx = 0;
 202                 n->vqs[i].upend_idx = 0;
 203                 n->vqs[i].ubufs = 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 /* In case of DMA done not in order in lower device driver for some reason.
 240  * upend_idx is used to track end of used idx, done_idx is used to track head
 241  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 242  * guest used idx.
 243  */
 244 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 245                                        struct vhost_virtqueue *vq)
 246 {
 247         struct vhost_net_virtqueue *nvq =
 248                 container_of(vq, struct vhost_net_virtqueue, vq);
 249         int i, add;
 250         int j = 0;
 251
 252         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 253                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 254                         vhost_net_tx_err(net);
 255                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 256                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 257                         ++j;
 258                 } else
 259                         break;
 260         }
 261         while (j) {
 262                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 263                 vhost_add_used_and_signal_n(vq->dev, vq,
 264                                             &vq->heads[nvq->done_idx], add);
 265                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 266                 j -= add;
 267         }
 268 }
 269
 270 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 271 {
 272         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 273         struct vhost_virtqueue *vq = ubufs->vq;
 274         int cnt;
 275
 276         rcu_read_lock_bh();
 277
 278         /* set len to mark this desc buffers done DMA */
 279         vq->heads[ubuf->desc].len = success ?
 280                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 281         cnt = vhost_net_ubuf_put(ubufs);
 282
 283         /*
 284          * Trigger polling thread if guest stopped submitting new buffers:
 285          * in this case, the refcount after decrement will eventually reach 1.
 286          * We also trigger polling periodically after each 16 packets
 287          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 288          * less than 10% of times).
 289          */
 290         if (cnt <= 1 || !(cnt % 16))
 291                 vhost_poll_queue(&vq->poll);
 292
 293         rcu_read_unlock_bh();
 294 }
 295
 296 /* Expects to be always run from workqueue - which acts as
 297  * read-size critical section for our kind of RCU. */
 298 static void handle_tx(struct vhost_net *net)
 299 {
 300         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 301         struct vhost_virtqueue *vq = &nvq->vq;
 302         unsigned out, in;
 303         int head;
 304         struct msghdr msg = {
 305                 .msg_name = NULL,
 306                 .msg_namelen = 0,
 307                 .msg_control = NULL,
 308                 .msg_controllen = 0,
 309                 .msg_flags = MSG_DONTWAIT,
 310         };
 311         size_t len, total_len = 0;
 312         int err;
 313         size_t hdr_size;
 314         struct socket *sock;
 315         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 316         struct ubuf_info *ubuf;
 317         bool zcopy, zcopy_used;
 318         int sent_pkts = 0;
 319
 320         mutex_lock(&vq->mutex);
 321         sock = vq->private_data;
 322         if (!sock)
 323                 goto out;
 324
 325         vhost_disable_notify(&net->dev, vq);
 326
 327         hdr_size = nvq->vhost_hlen;
 328         zcopy = nvq->ubufs;
 329
 330         do {
 331                 /* Release DMAs done buffers first */
 332                 if (zcopy)
 333                         vhost_zerocopy_signal_used(net, vq);
 334
 335                 /* If more outstanding DMAs, queue the work.
 336                  * Handle upend_idx wrap around
 337                  */
 338                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
 339                               % UIO_MAXIOV == nvq->done_idx))
 340                         break;
 341
 342                 head = vhost_get_vq_desc(vq, vq->iov,
 343                                          ARRAY_SIZE(vq->iov),
 344                                          &out, &in,
 345                                          NULL, NULL);
 346                 /* On error, stop handling until the next kick. */
 347                 if (unlikely(head < 0))
 348                         break;
 349                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 350                 if (head == vq->num) {
 351                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 352                                 vhost_disable_notify(&net->dev, vq);
 353                                 continue;
 354                         }
 355                         break;
 356                 }
 357                 if (in) {
 358                         vq_err(vq, "Unexpected descriptor format for TX: "
 359                                "out %d, int %d\n", out, in);
 360                         break;
 361                 }
 362                 /* Skip header. TODO: support TSO. */
 363                 len = iov_length(vq->iov, out);
 364                 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
 365                 iov_iter_advance(&msg.msg_iter, hdr_size);
 366                 /* Sanity check */
 367                 if (!msg_data_left(&msg)) {
 368                         vq_err(vq, "Unexpected header len for TX: "
 369                                "%zd expected %zd\n",
 370                                len, hdr_size);
 371                         break;
 372                 }
 373                 len = msg_data_left(&msg);
 374
 375                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 376                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 377                                       nvq->done_idx
 378                                    && vhost_net_tx_select_zcopy(net);
 379
 380                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 381                 if (zcopy_used) {
 382                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 383                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 384                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 385                         ubuf->callback = vhost_zerocopy_callback;
 386                         ubuf->ctx = nvq->ubufs;
 387                         ubuf->desc = nvq->upend_idx;
 388                         msg.msg_control = ubuf;
 389                         msg.msg_controllen = sizeof(ubuf);
 390                         ubufs = nvq->ubufs;
 391                         atomic_inc(&ubufs->refcount);
 392                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 393                 } else {
 394                         msg.msg_control = NULL;
 395                         ubufs = NULL;
 396                 }
 397                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 398                 err = sock->ops->sendmsg(sock, &msg, len);
 399                 if (unlikely(err < 0)) {
 400                         if (zcopy_used) {
 401                                 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
 402                                         vhost_net_ubuf_put(ubufs);
 403                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 404                                         % UIO_MAXIOV;
 405                         }
 406                         vhost_discard_vq_desc(vq, 1);
 407                         break;
 408                 }
 409                 if (err != len)
 410                         pr_debug("Truncated TX packet: "
 411                                  " len %d != %zd\n", err, len);
 412                 if (!zcopy_used)
 413                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 414                 else
 415                         vhost_zerocopy_signal_used(net, vq);
 416                 total_len += len;
 417                 vhost_net_tx_packet(net);
 418         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 419 out:
 420         mutex_unlock(&vq->mutex);
 421 }
 422
 423 static int peek_head_len(struct sock *sk)
 424 {
 425         struct sk_buff *head;
 426         int len = 0;
 427         unsigned long flags;
 428
 429         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 430         head = skb_peek(&sk->sk_receive_queue);
 431         if (likely(head)) {
 432                 len = head->len;
 433                 if (skb_vlan_tag_present(head))
 434                         len += VLAN_HLEN;
 435         }
 436
 437         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 438         return len;
 439 }
 440
 441 /* This is a multi-buffer version of vhost_get_desc, that works if
 442  *      vq has read descriptors only.
 443  * @vq          - the relevant virtqueue
 444  * @datalen     - data length we'll be reading
 445  * @iovcount    - returned count of io vectors we fill
 446  * @log         - vhost log
 447  * @log_num     - log offset
 448  * @quota       - headcount quota, 1 for big buffer
 449  *      returns number of buffer heads allocated, negative on error
 450  */
 451 static int get_rx_bufs(struct vhost_virtqueue *vq,
 452                        struct vring_used_elem *heads,
 453                        int datalen,
 454                        unsigned *iovcount,
 455                        struct vhost_log *log,
 456                        unsigned *log_num,
 457                        unsigned int quota)
 458 {
 459         unsigned int out, in;
 460         int seg = 0;
 461         int headcount = 0;
 462         unsigned d;
 463         int r, nlogs = 0;
 464         /* len is always initialized before use since we are always called with
 465          * datalen > 0.
 466          */
 467         u32 uninitialized_var(len);
 468
 469         while (datalen > 0 && headcount < quota) {
 470                 if (unlikely(seg >= UIO_MAXIOV)) {
 471                         r = -ENOBUFS;
 472                         goto err;
 473                 }
 474                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 475                                       ARRAY_SIZE(vq->iov) - seg, &out,
 476                                       &in, log, log_num);
 477                 if (unlikely(r < 0))
 478                         goto err;
 479
 480                 d = r;
 481                 if (d == vq->num) {
 482                         r = 0;
 483                         goto err;
 484                 }
 485                 if (unlikely(out || in <= 0)) {
 486                         vq_err(vq, "unexpected descriptor format for RX: "
 487                                 "out %d, in %d\n", out, in);
 488                         r = -EINVAL;
 489                         goto err;
 490                 }
 491                 if (unlikely(log)) {
 492                         nlogs += *log_num;
 493                         log += *log_num;
 494                 }
 495                 heads[headcount].id = cpu_to_vhost32(vq, d);
 496                 len = iov_length(vq->iov + seg, in);
 497                 heads[headcount].len = cpu_to_vhost32(vq, len);
 498                 datalen -= len;
 499                 ++headcount;
 500                 seg += in;
 501         }
 502         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 503         *iovcount = seg;
 504         if (unlikely(log))
 505                 *log_num = nlogs;
 506
 507         /* Detect overrun */
 508         if (unlikely(datalen > 0)) {
 509                 r = UIO_MAXIOV + 1;
 510                 goto err;
 511         }
 512         return headcount;
 513 err:
 514         vhost_discard_vq_desc(vq, headcount);
 515         return r;
 516 }
 517
 518 /* Expects to be always run from workqueue - which acts as
 519  * read-size critical section for our kind of RCU. */
 520 static void handle_rx(struct vhost_net *net)
 521 {
 522         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 523         struct vhost_virtqueue *vq = &nvq->vq;
 524         unsigned uninitialized_var(in), log;
 525         struct vhost_log *vq_log;
 526         struct msghdr msg = {
 527                 .msg_name = NULL,
 528                 .msg_namelen = 0,
 529                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 530                 .msg_controllen = 0,
 531                 .msg_flags = MSG_DONTWAIT,
 532         };
 533         struct virtio_net_hdr hdr = {
 534                 .flags = 0,
 535                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 536         };
 537         size_t total_len = 0;
 538         int err, mergeable;
 539         s16 headcount;
 540         size_t vhost_hlen, sock_hlen;
 541         size_t vhost_len, sock_len;
 542         struct socket *sock;
 543         struct iov_iter fixup;
 544         __virtio16 num_buffers;
 545         int recv_pkts = 0;
 546
 547         mutex_lock(&vq->mutex);
 548         sock = vq->private_data;
 549         if (!sock)
 550                 goto out;
 551         vhost_disable_notify(&net->dev, vq);
 552
 553         vhost_hlen = nvq->vhost_hlen;
 554         sock_hlen = nvq->sock_hlen;
 555
 556         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 557                 vq->log : NULL;
 558         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 559
 560         do {
 561                 sock_len = peek_head_len(sock->sk);
 562                 if (!sock_len)
 563                         break;
 564                 sock_len += sock_hlen;
 565                 vhost_len = sock_len + vhost_hlen;
 566                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 567                                         &in, vq_log, &log,
 568                                         likely(mergeable) ? UIO_MAXIOV : 1);
 569                 /* On error, stop handling until the next kick. */
 570                 if (unlikely(headcount < 0))
 571                         break;
 572                 /* On overrun, truncate and discard */
 573                 if (unlikely(headcount > UIO_MAXIOV)) {
 574                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 575                         err = sock->ops->recvmsg(sock, &msg,
 576                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 577                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 578                         continue;
 579                 }
 580                 /* OK, now we need to know about added descriptors. */
 581                 if (!headcount) {
 582                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 583                                 /* They have slipped one in as we were
 584                                  * doing that: check again. */
 585                                 vhost_disable_notify(&net->dev, vq);
 586                                 continue;
 587                         }
 588                         /* Nothing new?  Wait for eventfd to tell us
 589                          * they refilled. */
 590                         break;
 591                 }
 592                 /* We don't need to be notified again. */
 593                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
 594                 fixup = msg.msg_iter;
 595                 if (unlikely((vhost_hlen))) {
 596                         /* We will supply the header ourselves
 597                          * TODO: support TSO.
 598                          */
 599                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
 600                 }
 601                 err = sock->ops->recvmsg(sock, &msg,
 602                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 603                 /* Userspace might have consumed the packet meanwhile:
 604                  * it's not supposed to do this usually, but might be hard
 605                  * to prevent. Discard data we got (if any) and keep going. */
 606                 if (unlikely(err != sock_len)) {
 607                         pr_debug("Discarded rx packet: "
 608                                  " len %d, expected %zd\n", err, sock_len);
 609                         vhost_discard_vq_desc(vq, headcount);
 610                         continue;
 611                 }
 612                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
 613                 if (unlikely(vhost_hlen)) {
 614                         if (copy_to_iter(&hdr, sizeof(hdr),
 615                                          &fixup) != sizeof(hdr)) {
 616                                 vq_err(vq, "Unable to write vnet_hdr "
 617                                        "at addr %p\n", vq->iov->iov_base);
 618                                 break;
 619                         }
 620                 } else {
 621                         /* Header came from socket; we'll need to patch
 622                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
 623                          */
 624                         iov_iter_advance(&fixup, sizeof(hdr));
 625                 }
 626                 /* TODO: Should check and handle checksum. */
 627
 628                 num_buffers = cpu_to_vhost16(vq, headcount);
 629                 if (likely(mergeable) &&
 630                     copy_to_iter(&num_buffers, sizeof num_buffers,
 631                                  &fixup) != sizeof num_buffers) {
 632                         vq_err(vq, "Failed num_buffers write");
 633                         vhost_discard_vq_desc(vq, headcount);
 634                         break;
 635                 }
 636                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 637                                             headcount);
 638                 if (unlikely(vq_log))
 639                         vhost_log_write(vq, vq_log, log, vhost_len);
 640                 total_len += vhost_len;
 641         } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
 642
 643 out:
 644         mutex_unlock(&vq->mutex);
 645 }
 646
 647 static void handle_tx_kick(struct vhost_work *work)
 648 {
 649         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 650                                                   poll.work);
 651         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 652
 653         handle_tx(net);
 654 }
 655
 656 static void handle_rx_kick(struct vhost_work *work)
 657 {
 658         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 659                                                   poll.work);
 660         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 661
 662         handle_rx(net);
 663 }
 664
 665 static void handle_tx_net(struct vhost_work *work)
 666 {
 667         struct vhost_net *net = container_of(work, struct vhost_net,
 668                                              poll[VHOST_NET_VQ_TX].work);
 669         handle_tx(net);
 670 }
 671
 672 static void handle_rx_net(struct vhost_work *work)
 673 {
 674         struct vhost_net *net = container_of(work, struct vhost_net,
 675                                              poll[VHOST_NET_VQ_RX].work);
 676         handle_rx(net);
 677 }
 678
 679 static int vhost_net_open(struct inode *inode, struct file *f)
 680 {
 681         struct vhost_net *n;
 682         struct vhost_dev *dev;
 683         struct vhost_virtqueue **vqs;
 684         int i;
 685
 686         n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 687         if (!n) {
 688                 n = vmalloc(sizeof *n);
 689                 if (!n)
 690                         return -ENOMEM;
 691         }
 692         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 693         if (!vqs) {
 694                 kvfree(n);
 695                 return -ENOMEM;
 696         }
 697
 698         dev = &n->dev;
 699         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 700         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 701         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 702         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 703         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 704                 n->vqs[i].ubufs = NULL;
 705                 n->vqs[i].ubuf_info = NULL;
 706                 n->vqs[i].upend_idx = 0;
 707                 n->vqs[i].done_idx = 0;
 708                 n->vqs[i].vhost_hlen = 0;
 709                 n->vqs[i].sock_hlen = 0;
 710         }
 711         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
 712                        VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT);
 713
 714         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 715         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 716
 717         f->private_data = n;
 718
 719         return 0;
 720 }
 721
 722 static void vhost_net_disable_vq(struct vhost_net *n,
 723                                  struct vhost_virtqueue *vq)
 724 {
 725         struct vhost_net_virtqueue *nvq =
 726                 container_of(vq, struct vhost_net_virtqueue, vq);
 727         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 728         if (!vq->private_data)
 729                 return;
 730         vhost_poll_stop(poll);
 731 }
 732
 733 static int vhost_net_enable_vq(struct vhost_net *n,
 734                                 struct vhost_virtqueue *vq)
 735 {
 736         struct vhost_net_virtqueue *nvq =
 737                 container_of(vq, struct vhost_net_virtqueue, vq);
 738         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 739         struct socket *sock;
 740
 741         sock = vq->private_data;
 742         if (!sock)
 743                 return 0;
 744
 745         return vhost_poll_start(poll, sock->file);
 746 }
 747
 748 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 749                                         struct vhost_virtqueue *vq)
 750 {
 751         struct socket *sock;
 752
 753         mutex_lock(&vq->mutex);
 754         sock = vq->private_data;
 755         vhost_net_disable_vq(n, vq);
 756         vq->private_data = NULL;
 757         mutex_unlock(&vq->mutex);
 758         return sock;
 759 }
 760
 761 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 762                            struct socket **rx_sock)
 763 {
 764         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 765         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 766 }
 767
 768 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 769 {
 770         vhost_poll_flush(n->poll + index);
 771         vhost_poll_flush(&n->vqs[index].vq.poll);
 772 }
 773
 774 static void vhost_net_flush(struct vhost_net *n)
 775 {
 776         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 777         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 778         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 779                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 780                 n->tx_flush = true;
 781                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 782                 /* Wait for all lower device DMAs done. */
 783                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 784                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 785                 n->tx_flush = false;
 786                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 787                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 788         }
 789 }
 790
 791 static int vhost_net_release(struct inode *inode, struct file *f)
 792 {
 793         struct vhost_net *n = f->private_data;
 794         struct socket *tx_sock;
 795         struct socket *rx_sock;
 796
 797         vhost_net_stop(n, &tx_sock, &rx_sock);
 798         vhost_net_flush(n);
 799         vhost_dev_stop(&n->dev);
 800         vhost_dev_cleanup(&n->dev, false);
 801         vhost_net_vq_reset(n);
 802         if (tx_sock)
 803                 sockfd_put(tx_sock);
 804         if (rx_sock)
 805                 sockfd_put(rx_sock);
 806         /* Make sure no callbacks are outstanding */
 807         synchronize_rcu_bh();
 808         /* We do an extra flush before freeing memory,
 809          * since jobs can re-queue themselves. */
 810         vhost_net_flush(n);
 811         kfree(n->dev.vqs);
 812         kvfree(n);
 813         return 0;
 814 }
 815
 816 static struct socket *get_raw_socket(int fd)
 817 {
 818         int r;
 819         struct socket *sock = sockfd_lookup(fd, &r);
 820
 821         if (!sock)
 822                 return ERR_PTR(-ENOTSOCK);
 823
 824         /* Parameter checking */
 825         if (sock->sk->sk_type != SOCK_RAW) {
 826                 r = -ESOCKTNOSUPPORT;
 827                 goto err;
 828         }
 829
 830         if (sock->sk->sk_family != AF_PACKET) {
 831                 r = -EPFNOSUPPORT;
 832                 goto err;
 833         }
 834         return sock;
 835 err:
 836         sockfd_put(sock);
 837         return ERR_PTR(r);
 838 }
 839
 840 static struct socket *get_tap_socket(int fd)
 841 {
 842         struct file *file = fget(fd);
 843         struct socket *sock;
 844
 845         if (!file)
 846                 return ERR_PTR(-EBADF);
 847         sock = tun_get_socket(file);
 848         if (!IS_ERR(sock))
 849                 return sock;
 850         sock = macvtap_get_socket(file);
 851         if (IS_ERR(sock))
 852                 fput(file);
 853         return sock;
 854 }
 855
 856 static struct socket *get_socket(int fd)
 857 {
 858         struct socket *sock;
 859
 860         /* special case to disable backend */
 861         if (fd == -1)
 862                 return NULL;
 863         sock = get_raw_socket(fd);
 864         if (!IS_ERR(sock))
 865                 return sock;
 866         sock = get_tap_socket(fd);
 867         if (!IS_ERR(sock))
 868                 return sock;
 869         return ERR_PTR(-ENOTSOCK);
 870 }
 871
 872 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 873 {
 874         struct socket *sock, *oldsock;
 875         struct vhost_virtqueue *vq;
 876         struct vhost_net_virtqueue *nvq;
 877         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 878         int r;
 879
 880         mutex_lock(&n->dev.mutex);
 881         r = vhost_dev_check_owner(&n->dev);
 882         if (r)
 883                 goto err;
 884
 885         if (index >= VHOST_NET_VQ_MAX) {
 886                 r = -ENOBUFS;
 887                 goto err;
 888         }
 889         vq = &n->vqs[index].vq;
 890         nvq = &n->vqs[index];
 891         mutex_lock(&vq->mutex);
 892
 893         /* Verify that ring has been setup correctly. */
 894         if (!vhost_vq_access_ok(vq)) {
 895                 r = -EFAULT;
 896                 goto err_vq;
 897         }
 898         sock = get_socket(fd);
 899         if (IS_ERR(sock)) {
 900                 r = PTR_ERR(sock);
 901                 goto err_vq;
 902         }
 903
 904         /* start polling new socket */
 905         oldsock = vq->private_data;
 906         if (sock != oldsock) {
 907                 ubufs = vhost_net_ubuf_alloc(vq,
 908                                              sock && vhost_sock_zcopy(sock));
 909                 if (IS_ERR(ubufs)) {
 910                         r = PTR_ERR(ubufs);
 911                         goto err_ubufs;
 912                 }
 913
 914                 vhost_net_disable_vq(n, vq);
 915                 vq->private_data = sock;
 916                 r = vhost_init_used(vq);
 917                 if (r)
 918                         goto err_used;
 919                 r = vhost_net_enable_vq(n, vq);
 920                 if (r)
 921                         goto err_used;
 922
 923                 oldubufs = nvq->ubufs;
 924                 nvq->ubufs = ubufs;
 925
 926                 n->tx_packets = 0;
 927                 n->tx_zcopy_err = 0;
 928                 n->tx_flush = false;
 929         }
 930
 931         mutex_unlock(&vq->mutex);
 932
 933         if (oldubufs) {
 934                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 935                 mutex_lock(&vq->mutex);
 936                 vhost_zerocopy_signal_used(n, vq);
 937                 mutex_unlock(&vq->mutex);
 938         }
 939
 940         if (oldsock) {
 941                 vhost_net_flush_vq(n, index);
 942                 sockfd_put(oldsock);
 943         }
 944
 945         mutex_unlock(&n->dev.mutex);
 946         return 0;
 947
 948 err_used:
 949         vq->private_data = oldsock;
 950         vhost_net_enable_vq(n, vq);
 951         if (ubufs)
 952                 vhost_net_ubuf_put_wait_and_free(ubufs);
 953 err_ubufs:
 954         if (sock)
 955                 sockfd_put(sock);
 956 err_vq:
 957         mutex_unlock(&vq->mutex);
 958 err:
 959         mutex_unlock(&n->dev.mutex);
 960         return r;
 961 }
 962
 963 static long vhost_net_reset_owner(struct vhost_net *n)
 964 {
 965         struct socket *tx_sock = NULL;
 966         struct socket *rx_sock = NULL;
 967         long err;
 968         struct vhost_memory *memory;
 969
 970         mutex_lock(&n->dev.mutex);
 971         err = vhost_dev_check_owner(&n->dev);
 972         if (err)
 973                 goto done;
 974         memory = vhost_dev_reset_owner_prepare();
 975         if (!memory) {
 976                 err = -ENOMEM;
 977                 goto done;
 978         }
 979         vhost_net_stop(n, &tx_sock, &rx_sock);
 980         vhost_net_flush(n);
 981         vhost_dev_stop(&n->dev);
 982         vhost_dev_reset_owner(&n->dev, memory);
 983         vhost_net_vq_reset(n);
 984 done:
 985         mutex_unlock(&n->dev.mutex);
 986         if (tx_sock)
 987                 sockfd_put(tx_sock);
 988         if (rx_sock)
 989                 sockfd_put(rx_sock);
 990         return err;
 991 }
 992
 993 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 994 {
 995         size_t vhost_hlen, sock_hlen, hdr_len;
 996         int i;
 997
 998         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
 999                                (1ULL << VIRTIO_F_VERSION_1))) ?
1000                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1001                         sizeof(struct virtio_net_hdr);
1002         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1003                 /* vhost provides vnet_hdr */
1004                 vhost_hlen = hdr_len;
1005                 sock_hlen = 0;
1006         } else {
1007                 /* socket provides vnet_hdr */
1008                 vhost_hlen = 0;
1009                 sock_hlen = hdr_len;
1010         }
1011         mutex_lock(&n->dev.mutex);
1012         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1013             !vhost_log_access_ok(&n->dev)) {
1014                 mutex_unlock(&n->dev.mutex);
1015                 return -EFAULT;
1016         }
1017         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1018                 mutex_lock(&n->vqs[i].vq.mutex);
1019                 n->vqs[i].vq.acked_features = features;
1020                 n->vqs[i].vhost_hlen = vhost_hlen;
1021                 n->vqs[i].sock_hlen = sock_hlen;
1022                 mutex_unlock(&n->vqs[i].vq.mutex);
1023         }
1024         mutex_unlock(&n->dev.mutex);
1025         return 0;
1026 }
1027
1028 static long vhost_net_set_owner(struct vhost_net *n)
1029 {
1030         int r;
1031
1032         mutex_lock(&n->dev.mutex);
1033         if (vhost_dev_has_owner(&n->dev)) {
1034                 r = -EBUSY;
1035                 goto out;
1036         }
1037         r = vhost_net_set_ubuf_info(n);
1038         if (r)
1039                 goto out;
1040         r = vhost_dev_set_owner(&n->dev);
1041         if (r)
1042                 vhost_net_clear_ubuf_info(n);
1043         vhost_net_flush(n);
1044 out:
1045         mutex_unlock(&n->dev.mutex);
1046         return r;
1047 }
1048
1049 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1050                             unsigned long arg)
1051 {
1052         struct vhost_net *n = f->private_data;
1053         void __user *argp = (void __user *)arg;
1054         u64 __user *featurep = argp;
1055         struct vhost_vring_file backend;
1056         u64 features;
1057         int r;
1058
1059         switch (ioctl) {
1060         case VHOST_NET_SET_BACKEND:
1061                 if (copy_from_user(&backend, argp, sizeof backend))
1062                         return -EFAULT;
1063                 return vhost_net_set_backend(n, backend.index, backend.fd);
1064         case VHOST_GET_FEATURES:
1065                 features = VHOST_NET_FEATURES;
1066                 if (copy_to_user(featurep, &features, sizeof features))
1067                         return -EFAULT;
1068                 return 0;
1069         case VHOST_SET_FEATURES:
1070                 if (copy_from_user(&features, featurep, sizeof features))
1071                         return -EFAULT;
1072                 if (features & ~VHOST_NET_FEATURES)
1073                         return -EOPNOTSUPP;
1074                 return vhost_net_set_features(n, features);
1075         case VHOST_RESET_OWNER:
1076                 return vhost_net_reset_owner(n);
1077         case VHOST_SET_OWNER:
1078                 return vhost_net_set_owner(n);
1079         default:
1080                 mutex_lock(&n->dev.mutex);
1081                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1082                 if (r == -ENOIOCTLCMD)
1083                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1084                 else
1085                         vhost_net_flush(n);
1086                 mutex_unlock(&n->dev.mutex);
1087                 return r;
1088         }
1089 }
1090
1091 #ifdef CONFIG_COMPAT
1092 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1093                                    unsigned long arg)
1094 {
1095         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1096 }
1097 #endif
1098
1099 static const struct file_operations vhost_net_fops = {
1100         .owner          = THIS_MODULE,
1101         .release        = vhost_net_release,
1102         .unlocked_ioctl = vhost_net_ioctl,
1103 #ifdef CONFIG_COMPAT
1104         .compat_ioctl   = vhost_net_compat_ioctl,
1105 #endif
1106         .open           = vhost_net_open,
1107         .llseek         = noop_llseek,
1108 };
1109
1110 static struct miscdevice vhost_net_misc = {
1111         .minor = VHOST_NET_MINOR,
1112         .name = "vhost-net",
1113         .fops = &vhost_net_fops,
1114 };
1115
1116 static int vhost_net_init(void)
1117 {
1118         if (experimental_zcopytx)
1119                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1120         return misc_register(&vhost_net_misc);
1121 }
1122 module_init(vhost_net_init);
1123
1124 static void vhost_net_exit(void)
1125 {
1126         misc_deregister(&vhost_net_misc);
1127 }
1128 module_exit(vhost_net_exit);
1129
1130 MODULE_VERSION("0.0.1");
1131 MODULE_LICENSE("GPL v2");
1132 MODULE_AUTHOR("Michael S. Tsirkin");
1133 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1134 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1135 MODULE_ALIAS("devname:vhost-net");