Pileus Git - ~andy/linux/blob - drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/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         struct kref kref;
  74         wait_queue_head_t wait;
  75         struct vhost_virtqueue *vq;
  76 };
  77
  78 struct vhost_net_virtqueue {
  79         struct vhost_virtqueue vq;
  80         /* hdr is used to store the virtio header.
  81          * Since each iovec has >= 1 byte length, we never need more than
  82          * header length entries to store the header. */
  83         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
  84         size_t vhost_hlen;
  85         size_t sock_hlen;
  86         /* vhost zerocopy support fields below: */
  87         /* last used idx for outstanding DMA zerocopy buffers */
  88         int upend_idx;
  89         /* first used idx for DMA done zerocopy buffers */
  90         int done_idx;
  91         /* an array of userspace buffers info */
  92         struct ubuf_info *ubuf_info;
  93         /* Reference counting for outstanding ubufs.
  94          * Protected by vq mutex. Writers must also take device mutex. */
  95         struct vhost_net_ubuf_ref *ubufs;
  96 };
  97
  98 struct vhost_net {
  99         struct vhost_dev dev;
 100         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 101         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 102         /* Number of TX recently submitted.
 103          * Protected by tx vq lock. */
 104         unsigned tx_packets;
 105         /* Number of times zerocopy TX recently failed.
 106          * Protected by tx vq lock. */
 107         unsigned tx_zcopy_err;
 108         /* Flush in progress. Protected by tx vq lock. */
 109         bool tx_flush;
 110 };
 111
 112 static unsigned vhost_net_zcopy_mask __read_mostly;
 113
 114 static void vhost_net_enable_zcopy(int vq)
 115 {
 116         vhost_net_zcopy_mask |= 0x1 << vq;
 117 }
 118
 119 static void vhost_net_zerocopy_done_signal(struct kref *kref)
 120 {
 121         struct vhost_net_ubuf_ref *ubufs;
 122
 123         ubufs = container_of(kref, struct vhost_net_ubuf_ref, kref);
 124         wake_up(&ubufs->wait);
 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         kref_init(&ubufs->kref);
 138         init_waitqueue_head(&ubufs->wait);
 139         ubufs->vq = vq;
 140         return ubufs;
 141 }
 142
 143 static void vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 144 {
 145         kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
 146 }
 147
 148 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 149 {
 150         kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
 151         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.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 = atomic_read(&ubufs->kref.refcount);
 310
 311         /* set len to mark this desc buffers done DMA */
 312         vq->heads[ubuf->desc].len = success ?
 313                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 314         vhost_net_ubuf_put(ubufs);
 315
 316         /*
 317          * Trigger polling thread if guest stopped submitting new buffers:
 318          * in this case, the refcount after decrement will eventually reach 1
 319          * so here it is 2.
 320          * We also trigger polling periodically after each 16 packets
 321          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 322          * less than 10% of times).
 323          */
 324         if (cnt <= 2 || !(cnt % 16))
 325                 vhost_poll_queue(&vq->poll);
 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_net_ubuf_ref *uninitialized_var(ubufs);
 349         bool zcopy, zcopy_used;
 350
 351         mutex_lock(&vq->mutex);
 352         sock = vq->private_data;
 353         if (!sock)
 354                 goto out;
 355
 356         vhost_disable_notify(&net->dev, vq);
 357
 358         hdr_size = nvq->vhost_hlen;
 359         zcopy = nvq->ubufs;
 360
 361         for (;;) {
 362                 /* Release DMAs done buffers first */
 363                 if (zcopy)
 364                         vhost_zerocopy_signal_used(net, vq);
 365
 366                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 367                                          ARRAY_SIZE(vq->iov),
 368                                          &out, &in,
 369                                          NULL, NULL);
 370                 /* On error, stop handling until the next kick. */
 371                 if (unlikely(head < 0))
 372                         break;
 373                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 374                 if (head == vq->num) {
 375                         int num_pends;
 376
 377                         /* If more outstanding DMAs, queue the work.
 378                          * Handle upend_idx wrap around
 379                          */
 380                         num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
 381                                     (nvq->upend_idx - nvq->done_idx) :
 382                                     (nvq->upend_idx + UIO_MAXIOV -
 383                                      nvq->done_idx);
 384                         if (unlikely(num_pends > VHOST_MAX_PEND))
 385                                 break;
 386                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 387                                 vhost_disable_notify(&net->dev, vq);
 388                                 continue;
 389                         }
 390                         break;
 391                 }
 392                 if (in) {
 393                         vq_err(vq, "Unexpected descriptor format for TX: "
 394                                "out %d, int %d\n", out, in);
 395                         break;
 396                 }
 397                 /* Skip header. TODO: support TSO. */
 398                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 399                 msg.msg_iovlen = out;
 400                 len = iov_length(vq->iov, out);
 401                 /* Sanity check */
 402                 if (!len) {
 403                         vq_err(vq, "Unexpected header len for TX: "
 404                                "%zd expected %zd\n",
 405                                iov_length(nvq->hdr, s), hdr_size);
 406                         break;
 407                 }
 408
 409                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 410                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 411                                       nvq->done_idx
 412                                    && vhost_net_tx_select_zcopy(net);
 413
 414                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 415                 if (zcopy_used) {
 416                         struct ubuf_info *ubuf;
 417                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 418
 419                         vq->heads[nvq->upend_idx].id = head;
 420                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 421                         ubuf->callback = vhost_zerocopy_callback;
 422                         ubuf->ctx = nvq->ubufs;
 423                         ubuf->desc = nvq->upend_idx;
 424                         msg.msg_control = ubuf;
 425                         msg.msg_controllen = sizeof(ubuf);
 426                         ubufs = nvq->ubufs;
 427                         kref_get(&ubufs->kref);
 428                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 429                 } else {
 430                         msg.msg_control = NULL;
 431                         ubufs = NULL;
 432                 }
 433                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 434                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 435                 if (unlikely(err < 0)) {
 436                         if (zcopy_used) {
 437                                 vhost_net_ubuf_put(ubufs);
 438                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 439                                         % UIO_MAXIOV;
 440                         }
 441                         vhost_discard_vq_desc(vq, 1);
 442                         break;
 443                 }
 444                 if (err != len)
 445                         pr_debug("Truncated TX packet: "
 446                                  " len %d != %zd\n", err, len);
 447                 if (!zcopy_used)
 448                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 449                 else
 450                         vhost_zerocopy_signal_used(net, vq);
 451                 total_len += len;
 452                 vhost_net_tx_packet(net);
 453                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 454                         vhost_poll_queue(&vq->poll);
 455                         break;
 456                 }
 457         }
 458 out:
 459         mutex_unlock(&vq->mutex);
 460 }
 461
 462 static int peek_head_len(struct sock *sk)
 463 {
 464         struct sk_buff *head;
 465         int len = 0;
 466         unsigned long flags;
 467
 468         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 469         head = skb_peek(&sk->sk_receive_queue);
 470         if (likely(head)) {
 471                 len = head->len;
 472                 if (vlan_tx_tag_present(head))
 473                         len += VLAN_HLEN;
 474         }
 475
 476         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 477         return len;
 478 }
 479
 480 /* This is a multi-buffer version of vhost_get_desc, that works if
 481  *      vq has read descriptors only.
 482  * @vq          - the relevant virtqueue
 483  * @datalen     - data length we'll be reading
 484  * @iovcount    - returned count of io vectors we fill
 485  * @log         - vhost log
 486  * @log_num     - log offset
 487  * @quota       - headcount quota, 1 for big buffer
 488  *      returns number of buffer heads allocated, negative on error
 489  */
 490 static int get_rx_bufs(struct vhost_virtqueue *vq,
 491                        struct vring_used_elem *heads,
 492                        int datalen,
 493                        unsigned *iovcount,
 494                        struct vhost_log *log,
 495                        unsigned *log_num,
 496                        unsigned int quota)
 497 {
 498         unsigned int out, in;
 499         int seg = 0;
 500         int headcount = 0;
 501         unsigned d;
 502         int r, nlogs = 0;
 503
 504         while (datalen > 0 && headcount < quota) {
 505                 if (unlikely(seg >= UIO_MAXIOV)) {
 506                         r = -ENOBUFS;
 507                         goto err;
 508                 }
 509                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 510                                       ARRAY_SIZE(vq->iov) - seg, &out,
 511                                       &in, log, log_num);
 512                 if (d == vq->num) {
 513                         r = 0;
 514                         goto err;
 515                 }
 516                 if (unlikely(out || in <= 0)) {
 517                         vq_err(vq, "unexpected descriptor format for RX: "
 518                                 "out %d, in %d\n", out, in);
 519                         r = -EINVAL;
 520                         goto err;
 521                 }
 522                 if (unlikely(log)) {
 523                         nlogs += *log_num;
 524                         log += *log_num;
 525                 }
 526                 heads[headcount].id = d;
 527                 heads[headcount].len = iov_length(vq->iov + seg, in);
 528                 datalen -= heads[headcount].len;
 529                 ++headcount;
 530                 seg += in;
 531         }
 532         heads[headcount - 1].len += datalen;
 533         *iovcount = seg;
 534         if (unlikely(log))
 535                 *log_num = nlogs;
 536         return headcount;
 537 err:
 538         vhost_discard_vq_desc(vq, headcount);
 539         return r;
 540 }
 541
 542 /* Expects to be always run from workqueue - which acts as
 543  * read-size critical section for our kind of RCU. */
 544 static void handle_rx(struct vhost_net *net)
 545 {
 546         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 547         struct vhost_virtqueue *vq = &nvq->vq;
 548         unsigned uninitialized_var(in), log;
 549         struct vhost_log *vq_log;
 550         struct msghdr msg = {
 551                 .msg_name = NULL,
 552                 .msg_namelen = 0,
 553                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 554                 .msg_controllen = 0,
 555                 .msg_iov = vq->iov,
 556                 .msg_flags = MSG_DONTWAIT,
 557         };
 558         struct virtio_net_hdr_mrg_rxbuf hdr = {
 559                 .hdr.flags = 0,
 560                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 561         };
 562         size_t total_len = 0;
 563         int err, mergeable;
 564         s16 headcount;
 565         size_t vhost_hlen, sock_hlen;
 566         size_t vhost_len, sock_len;
 567         struct socket *sock;
 568
 569         mutex_lock(&vq->mutex);
 570         sock = vq->private_data;
 571         if (!sock)
 572                 goto out;
 573         vhost_disable_notify(&net->dev, vq);
 574
 575         vhost_hlen = nvq->vhost_hlen;
 576         sock_hlen = nvq->sock_hlen;
 577
 578         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 579                 vq->log : NULL;
 580         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 581
 582         while ((sock_len = peek_head_len(sock->sk))) {
 583                 sock_len += sock_hlen;
 584                 vhost_len = sock_len + vhost_hlen;
 585                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 586                                         &in, vq_log, &log,
 587                                         likely(mergeable) ? UIO_MAXIOV : 1);
 588                 /* On error, stop handling until the next kick. */
 589                 if (unlikely(headcount < 0))
 590                         break;
 591                 /* OK, now we need to know about added descriptors. */
 592                 if (!headcount) {
 593                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 594                                 /* They have slipped one in as we were
 595                                  * doing that: check again. */
 596                                 vhost_disable_notify(&net->dev, vq);
 597                                 continue;
 598                         }
 599                         /* Nothing new?  Wait for eventfd to tell us
 600                          * they refilled. */
 601                         break;
 602                 }
 603                 /* We don't need to be notified again. */
 604                 if (unlikely((vhost_hlen)))
 605                         /* Skip header. TODO: support TSO. */
 606                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 607                 else
 608                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 609                          * needed because recvmsg can modify msg_iov. */
 610                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 611                 msg.msg_iovlen = in;
 612                 err = sock->ops->recvmsg(NULL, sock, &msg,
 613                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 614                 /* Userspace might have consumed the packet meanwhile:
 615                  * it's not supposed to do this usually, but might be hard
 616                  * to prevent. Discard data we got (if any) and keep going. */
 617                 if (unlikely(err != sock_len)) {
 618                         pr_debug("Discarded rx packet: "
 619                                  " len %d, expected %zd\n", err, sock_len);
 620                         vhost_discard_vq_desc(vq, headcount);
 621                         continue;
 622                 }
 623                 if (unlikely(vhost_hlen) &&
 624                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 625                                       vhost_hlen)) {
 626                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 627                                vq->iov->iov_base);
 628                         break;
 629                 }
 630                 /* TODO: Should check and handle checksum. */
 631                 if (likely(mergeable) &&
 632                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 633                                       offsetof(typeof(hdr), num_buffers),
 634                                       sizeof hdr.num_buffers)) {
 635                         vq_err(vq, "Failed num_buffers write");
 636                         vhost_discard_vq_desc(vq, headcount);
 637                         break;
 638                 }
 639                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 640                                             headcount);
 641                 if (unlikely(vq_log))
 642                         vhost_log_write(vq, vq_log, log, vhost_len);
 643                 total_len += vhost_len;
 644                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 645                         vhost_poll_queue(&vq->poll);
 646                         break;
 647                 }
 648         }
 649 out:
 650         mutex_unlock(&vq->mutex);
 651 }
 652
 653 static void handle_tx_kick(struct vhost_work *work)
 654 {
 655         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 656                                                   poll.work);
 657         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 658
 659         handle_tx(net);
 660 }
 661
 662 static void handle_rx_kick(struct vhost_work *work)
 663 {
 664         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 665                                                   poll.work);
 666         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 667
 668         handle_rx(net);
 669 }
 670
 671 static void handle_tx_net(struct vhost_work *work)
 672 {
 673         struct vhost_net *net = container_of(work, struct vhost_net,
 674                                              poll[VHOST_NET_VQ_TX].work);
 675         handle_tx(net);
 676 }
 677
 678 static void handle_rx_net(struct vhost_work *work)
 679 {
 680         struct vhost_net *net = container_of(work, struct vhost_net,
 681                                              poll[VHOST_NET_VQ_RX].work);
 682         handle_rx(net);
 683 }
 684
 685 static int vhost_net_open(struct inode *inode, struct file *f)
 686 {
 687         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 688         struct vhost_dev *dev;
 689         struct vhost_virtqueue **vqs;
 690         int r, i;
 691
 692         if (!n)
 693                 return -ENOMEM;
 694         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 695         if (!vqs) {
 696                 kfree(n);
 697                 return -ENOMEM;
 698         }
 699
 700         dev = &n->dev;
 701         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 702         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 703         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 704         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 705         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 706                 n->vqs[i].ubufs = NULL;
 707                 n->vqs[i].ubuf_info = NULL;
 708                 n->vqs[i].upend_idx = 0;
 709                 n->vqs[i].done_idx = 0;
 710                 n->vqs[i].vhost_hlen = 0;
 711                 n->vqs[i].sock_hlen = 0;
 712         }
 713         r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 714         if (r < 0) {
 715                 kfree(n);
 716                 kfree(vqs);
 717                 return r;
 718         }
 719
 720         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 721         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 722
 723         f->private_data = n;
 724
 725         return 0;
 726 }
 727
 728 static void vhost_net_disable_vq(struct vhost_net *n,
 729                                  struct vhost_virtqueue *vq)
 730 {
 731         struct vhost_net_virtqueue *nvq =
 732                 container_of(vq, struct vhost_net_virtqueue, vq);
 733         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 734         if (!vq->private_data)
 735                 return;
 736         vhost_poll_stop(poll);
 737 }
 738
 739 static int vhost_net_enable_vq(struct vhost_net *n,
 740                                 struct vhost_virtqueue *vq)
 741 {
 742         struct vhost_net_virtqueue *nvq =
 743                 container_of(vq, struct vhost_net_virtqueue, vq);
 744         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 745         struct socket *sock;
 746
 747         sock = vq->private_data;
 748         if (!sock)
 749                 return 0;
 750
 751         return vhost_poll_start(poll, sock->file);
 752 }
 753
 754 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 755                                         struct vhost_virtqueue *vq)
 756 {
 757         struct socket *sock;
 758
 759         mutex_lock(&vq->mutex);
 760         sock = vq->private_data;
 761         vhost_net_disable_vq(n, vq);
 762         vq->private_data = NULL;
 763         mutex_unlock(&vq->mutex);
 764         return sock;
 765 }
 766
 767 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 768                            struct socket **rx_sock)
 769 {
 770         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 771         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 772 }
 773
 774 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 775 {
 776         vhost_poll_flush(n->poll + index);
 777         vhost_poll_flush(&n->vqs[index].vq.poll);
 778 }
 779
 780 static void vhost_net_flush(struct vhost_net *n)
 781 {
 782         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 783         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 784         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 785                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 786                 n->tx_flush = true;
 787                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 788                 /* Wait for all lower device DMAs done. */
 789                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 790                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 791                 n->tx_flush = false;
 792                 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
 793                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 794         }
 795 }
 796
 797 static int vhost_net_release(struct inode *inode, struct file *f)
 798 {
 799         struct vhost_net *n = f->private_data;
 800         struct socket *tx_sock;
 801         struct socket *rx_sock;
 802
 803         vhost_net_stop(n, &tx_sock, &rx_sock);
 804         vhost_net_flush(n);
 805         vhost_dev_stop(&n->dev);
 806         vhost_dev_cleanup(&n->dev, false);
 807         vhost_net_vq_reset(n);
 808         if (tx_sock)
 809                 fput(tx_sock->file);
 810         if (rx_sock)
 811                 fput(rx_sock->file);
 812         /* We do an extra flush before freeing memory,
 813          * since jobs can re-queue themselves. */
 814         vhost_net_flush(n);
 815         kfree(n->dev.vqs);
 816         kfree(n);
 817         return 0;
 818 }
 819
 820 static struct socket *get_raw_socket(int fd)
 821 {
 822         struct {
 823                 struct sockaddr_ll sa;
 824                 char  buf[MAX_ADDR_LEN];
 825         } uaddr;
 826         int uaddr_len = sizeof uaddr, r;
 827         struct socket *sock = sockfd_lookup(fd, &r);
 828
 829         if (!sock)
 830                 return ERR_PTR(-ENOTSOCK);
 831
 832         /* Parameter checking */
 833         if (sock->sk->sk_type != SOCK_RAW) {
 834                 r = -ESOCKTNOSUPPORT;
 835                 goto err;
 836         }
 837
 838         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 839                                &uaddr_len, 0);
 840         if (r)
 841                 goto err;
 842
 843         if (uaddr.sa.sll_family != AF_PACKET) {
 844                 r = -EPFNOSUPPORT;
 845                 goto err;
 846         }
 847         return sock;
 848 err:
 849         fput(sock->file);
 850         return ERR_PTR(r);
 851 }
 852
 853 static struct socket *get_tap_socket(int fd)
 854 {
 855         struct file *file = fget(fd);
 856         struct socket *sock;
 857
 858         if (!file)
 859                 return ERR_PTR(-EBADF);
 860         sock = tun_get_socket(file);
 861         if (!IS_ERR(sock))
 862                 return sock;
 863         sock = macvtap_get_socket(file);
 864         if (IS_ERR(sock))
 865                 fput(file);
 866         return sock;
 867 }
 868
 869 static struct socket *get_socket(int fd)
 870 {
 871         struct socket *sock;
 872
 873         /* special case to disable backend */
 874         if (fd == -1)
 875                 return NULL;
 876         sock = get_raw_socket(fd);
 877         if (!IS_ERR(sock))
 878                 return sock;
 879         sock = get_tap_socket(fd);
 880         if (!IS_ERR(sock))
 881                 return sock;
 882         return ERR_PTR(-ENOTSOCK);
 883 }
 884
 885 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 886 {
 887         struct socket *sock, *oldsock;
 888         struct vhost_virtqueue *vq;
 889         struct vhost_net_virtqueue *nvq;
 890         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 891         int r;
 892
 893         mutex_lock(&n->dev.mutex);
 894         r = vhost_dev_check_owner(&n->dev);
 895         if (r)
 896                 goto err;
 897
 898         if (index >= VHOST_NET_VQ_MAX) {
 899                 r = -ENOBUFS;
 900                 goto err;
 901         }
 902         vq = &n->vqs[index].vq;
 903         nvq = &n->vqs[index];
 904         mutex_lock(&vq->mutex);
 905
 906         /* Verify that ring has been setup correctly. */
 907         if (!vhost_vq_access_ok(vq)) {
 908                 r = -EFAULT;
 909                 goto err_vq;
 910         }
 911         sock = get_socket(fd);
 912         if (IS_ERR(sock)) {
 913                 r = PTR_ERR(sock);
 914                 goto err_vq;
 915         }
 916
 917         /* start polling new socket */
 918         oldsock = vq->private_data;
 919         if (sock != oldsock) {
 920                 ubufs = vhost_net_ubuf_alloc(vq,
 921                                              sock && vhost_sock_zcopy(sock));
 922                 if (IS_ERR(ubufs)) {
 923                         r = PTR_ERR(ubufs);
 924                         goto err_ubufs;
 925                 }
 926
 927                 vhost_net_disable_vq(n, vq);
 928                 vq->private_data = sock;
 929                 r = vhost_init_used(vq);
 930                 if (r)
 931                         goto err_used;
 932                 r = vhost_net_enable_vq(n, vq);
 933                 if (r)
 934                         goto err_used;
 935
 936                 oldubufs = nvq->ubufs;
 937                 nvq->ubufs = ubufs;
 938
 939                 n->tx_packets = 0;
 940                 n->tx_zcopy_err = 0;
 941                 n->tx_flush = false;
 942         }
 943
 944         mutex_unlock(&vq->mutex);
 945
 946         if (oldubufs) {
 947                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 948                 mutex_lock(&vq->mutex);
 949                 vhost_zerocopy_signal_used(n, vq);
 950                 mutex_unlock(&vq->mutex);
 951         }
 952
 953         if (oldsock) {
 954                 vhost_net_flush_vq(n, index);
 955                 fput(oldsock->file);
 956         }
 957
 958         mutex_unlock(&n->dev.mutex);
 959         return 0;
 960
 961 err_used:
 962         vq->private_data = oldsock;
 963         vhost_net_enable_vq(n, vq);
 964         if (ubufs)
 965                 vhost_net_ubuf_put_wait_and_free(ubufs);
 966 err_ubufs:
 967         fput(sock->file);
 968 err_vq:
 969         mutex_unlock(&vq->mutex);
 970 err:
 971         mutex_unlock(&n->dev.mutex);
 972         return r;
 973 }
 974
 975 static long vhost_net_reset_owner(struct vhost_net *n)
 976 {
 977         struct socket *tx_sock = NULL;
 978         struct socket *rx_sock = NULL;
 979         long err;
 980         struct vhost_memory *memory;
 981
 982         mutex_lock(&n->dev.mutex);
 983         err = vhost_dev_check_owner(&n->dev);
 984         if (err)
 985                 goto done;
 986         memory = vhost_dev_reset_owner_prepare();
 987         if (!memory) {
 988                 err = -ENOMEM;
 989                 goto done;
 990         }
 991         vhost_net_stop(n, &tx_sock, &rx_sock);
 992         vhost_net_flush(n);
 993         vhost_dev_reset_owner(&n->dev, memory);
 994         vhost_net_vq_reset(n);
 995 done:
 996         mutex_unlock(&n->dev.mutex);
 997         if (tx_sock)
 998                 fput(tx_sock->file);
 999         if (rx_sock)
1000                 fput(rx_sock->file);
1001         return err;
1002 }
1003
1004 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1005 {
1006         size_t vhost_hlen, sock_hlen, hdr_len;
1007         int i;
1008
1009         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1010                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1011                         sizeof(struct virtio_net_hdr);
1012         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1013                 /* vhost provides vnet_hdr */
1014                 vhost_hlen = hdr_len;
1015                 sock_hlen = 0;
1016         } else {
1017                 /* socket provides vnet_hdr */
1018                 vhost_hlen = 0;
1019                 sock_hlen = hdr_len;
1020         }
1021         mutex_lock(&n->dev.mutex);
1022         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1023             !vhost_log_access_ok(&n->dev)) {
1024                 mutex_unlock(&n->dev.mutex);
1025                 return -EFAULT;
1026         }
1027         n->dev.acked_features = features;
1028         smp_wmb();
1029         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1030                 mutex_lock(&n->vqs[i].vq.mutex);
1031                 n->vqs[i].vhost_hlen = vhost_hlen;
1032                 n->vqs[i].sock_hlen = sock_hlen;
1033                 mutex_unlock(&n->vqs[i].vq.mutex);
1034         }
1035         vhost_net_flush(n);
1036         mutex_unlock(&n->dev.mutex);
1037         return 0;
1038 }
1039
1040 static long vhost_net_set_owner(struct vhost_net *n)
1041 {
1042         int r;
1043
1044         mutex_lock(&n->dev.mutex);
1045         if (vhost_dev_has_owner(&n->dev)) {
1046                 r = -EBUSY;
1047                 goto out;
1048         }
1049         r = vhost_net_set_ubuf_info(n);
1050         if (r)
1051                 goto out;
1052         r = vhost_dev_set_owner(&n->dev);
1053         if (r)
1054                 vhost_net_clear_ubuf_info(n);
1055         vhost_net_flush(n);
1056 out:
1057         mutex_unlock(&n->dev.mutex);
1058         return r;
1059 }
1060
1061 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1062                             unsigned long arg)
1063 {
1064         struct vhost_net *n = f->private_data;
1065         void __user *argp = (void __user *)arg;
1066         u64 __user *featurep = argp;
1067         struct vhost_vring_file backend;
1068         u64 features;
1069         int r;
1070
1071         switch (ioctl) {
1072         case VHOST_NET_SET_BACKEND:
1073                 if (copy_from_user(&backend, argp, sizeof backend))
1074                         return -EFAULT;
1075                 return vhost_net_set_backend(n, backend.index, backend.fd);
1076         case VHOST_GET_FEATURES:
1077                 features = VHOST_NET_FEATURES;
1078                 if (copy_to_user(featurep, &features, sizeof features))
1079                         return -EFAULT;
1080                 return 0;
1081         case VHOST_SET_FEATURES:
1082                 if (copy_from_user(&features, featurep, sizeof features))
1083                         return -EFAULT;
1084                 if (features & ~VHOST_NET_FEATURES)
1085                         return -EOPNOTSUPP;
1086                 return vhost_net_set_features(n, features);
1087         case VHOST_RESET_OWNER:
1088                 return vhost_net_reset_owner(n);
1089         case VHOST_SET_OWNER:
1090                 return vhost_net_set_owner(n);
1091         default:
1092                 mutex_lock(&n->dev.mutex);
1093                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1094                 if (r == -ENOIOCTLCMD)
1095                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1096                 else
1097                         vhost_net_flush(n);
1098                 mutex_unlock(&n->dev.mutex);
1099                 return r;
1100         }
1101 }
1102
1103 #ifdef CONFIG_COMPAT
1104 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1105                                    unsigned long arg)
1106 {
1107         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1108 }
1109 #endif
1110
1111 static const struct file_operations vhost_net_fops = {
1112         .owner          = THIS_MODULE,
1113         .release        = vhost_net_release,
1114         .unlocked_ioctl = vhost_net_ioctl,
1115 #ifdef CONFIG_COMPAT
1116         .compat_ioctl   = vhost_net_compat_ioctl,
1117 #endif
1118         .open           = vhost_net_open,
1119         .llseek         = noop_llseek,
1120 };
1121
1122 static struct miscdevice vhost_net_misc = {
1123         .minor = VHOST_NET_MINOR,
1124         .name = "vhost-net",
1125         .fops = &vhost_net_fops,
1126 };
1127
1128 static int vhost_net_init(void)
1129 {
1130         if (experimental_zcopytx)
1131                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1132         return misc_register(&vhost_net_misc);
1133 }
1134 module_init(vhost_net_init);
1135
1136 static void vhost_net_exit(void)
1137 {
1138         misc_deregister(&vhost_net_misc);
1139 }
1140 module_exit(vhost_net_exit);
1141
1142 MODULE_VERSION("0.0.1");
1143 MODULE_LICENSE("GPL v2");
1144 MODULE_AUTHOR("Michael S. Tsirkin");
1145 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1146 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1147 MODULE_ALIAS("devname:vhost-net");