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