drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20
  21 #include <linux/net.h>
  22 #include <linux/if_packet.h>
  23 #include <linux/if_arp.h>
  24 #include <linux/if_tun.h>
  25 #include <linux/if_macvlan.h>
  26 #include <linux/if_vlan.h>
  27
  28 #include <net/sock.h>
  29
  30 #include "vhost.h"
  31
  32 static int experimental_zcopytx = 1;
  33 module_param(experimental_zcopytx, int, 0444);
  34 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  35                                        " 1 -Enable; 0 - Disable");
  36
  37 /* Max number of bytes transferred before requeueing the job.
  38  * Using this limit prevents one virtqueue from starving others. */
  39 #define VHOST_NET_WEIGHT 0x80000
  40
  41 /* MAX number of TX used buffers for outstanding zerocopy */
  42 #define VHOST_MAX_PEND 128
  43 #define VHOST_GOODCOPY_LEN 256
  44
  45 /*
  46  * For transmit, used buffer len is unused; we override it to track buffer
  47  * status internally; used for zerocopy tx only.
  48  */
  49 /* Lower device DMA failed */
  50 #define VHOST_DMA_FAILED_LEN    3
  51 /* Lower device DMA done */
  52 #define VHOST_DMA_DONE_LEN      2
  53 /* Lower device DMA in progress */
  54 #define VHOST_DMA_IN_PROGRESS   1
  55 /* Buffer unused */
  56 #define VHOST_DMA_CLEAR_LEN     0
  57
  58 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
  59
  60 enum {
  61         VHOST_NET_FEATURES = VHOST_FEATURES |
  62                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  63                          (1ULL << VIRTIO_NET_F_MRG_RXBUF),
  64 };
  65
  66 enum {
  67         VHOST_NET_VQ_RX = 0,
  68         VHOST_NET_VQ_TX = 1,
  69         VHOST_NET_VQ_MAX = 2,
  70 };
  71
  72 struct vhost_net_ubuf_ref {
  73         /* refcount follows semantics similar to kref:
  74          *  0: object is released
  75          *  1: no outstanding ubufs
  76          * >1: outstanding ubufs
  77          */
  78         atomic_t refcount;
  79         wait_queue_head_t wait;
  80         struct vhost_virtqueue *vq;
  81 };
  82
  83 struct vhost_net_virtqueue {
  84         struct vhost_virtqueue vq;
  85         /* hdr is used to store the virtio header.
  86          * Since each iovec has >= 1 byte length, we never need more than
  87          * header length entries to store the header. */
  88         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
  89         size_t vhost_hlen;
  90         size_t sock_hlen;
  91         /* vhost zerocopy support fields below: */
  92         /* last used idx for outstanding DMA zerocopy buffers */
  93         int upend_idx;
  94         /* first used idx for DMA done zerocopy buffers */
  95         int done_idx;
  96         /* an array of userspace buffers info */
  97         struct ubuf_info *ubuf_info;
  98         /* Reference counting for outstanding ubufs.
  99          * Protected by vq mutex. Writers must also take device mutex. */
 100         struct vhost_net_ubuf_ref *ubufs;
 101 };
 102
 103 struct vhost_net {
 104         struct vhost_dev dev;
 105         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 106         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 107         /* Number of TX recently submitted.
 108          * Protected by tx vq lock. */
 109         unsigned tx_packets;
 110         /* Number of times zerocopy TX recently failed.
 111          * Protected by tx vq lock. */
 112         unsigned tx_zcopy_err;
 113         /* Flush in progress. Protected by tx vq lock. */
 114         bool tx_flush;
 115 };
 116
 117 static unsigned vhost_net_zcopy_mask __read_mostly;
 118
 119 static void vhost_net_enable_zcopy(int vq)
 120 {
 121         vhost_net_zcopy_mask |= 0x1 << vq;
 122 }
 123
 124 static struct vhost_net_ubuf_ref *
 125 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 126 {
 127         struct vhost_net_ubuf_ref *ubufs;
 128         /* No zero copy backend? Nothing to count. */
 129         if (!zcopy)
 130                 return NULL;
 131         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 132         if (!ubufs)
 133                 return ERR_PTR(-ENOMEM);
 134         atomic_set(&ubufs->refcount, 1);
 135         init_waitqueue_head(&ubufs->wait);
 136         ubufs->vq = vq;
 137         return ubufs;
 138 }
 139
 140 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 141 {
 142         int r = atomic_sub_return(1, &ubufs->refcount);
 143         if (unlikely(!r))
 144                 wake_up(&ubufs->wait);
 145         return r;
 146 }
 147
 148 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 149 {
 150         vhost_net_ubuf_put(ubufs);
 151         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 152 }
 153
 154 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 155 {
 156         vhost_net_ubuf_put_and_wait(ubufs);
 157         kfree(ubufs);
 158 }
 159
 160 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 161 {
 162         int i;
 163
 164         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 165                 kfree(n->vqs[i].ubuf_info);
 166                 n->vqs[i].ubuf_info = NULL;
 167         }
 168 }
 169
 170 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 171 {
 172         bool zcopy;
 173         int i;
 174
 175         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 176                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 177                 if (!zcopy)
 178                         continue;
 179                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 180                                               UIO_MAXIOV, GFP_KERNEL);
 181                 if  (!n->vqs[i].ubuf_info)
 182                         goto err;
 183         }
 184         return 0;
 185
 186 err:
 187         vhost_net_clear_ubuf_info(n);
 188         return -ENOMEM;
 189 }
 190
 191 static void vhost_net_vq_reset(struct vhost_net *n)
 192 {
 193         int i;
 194
 195         vhost_net_clear_ubuf_info(n);
 196
 197         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 198                 n->vqs[i].done_idx = 0;
 199                 n->vqs[i].upend_idx = 0;
 200                 n->vqs[i].ubufs = NULL;
 201                 n->vqs[i].vhost_hlen = 0;
 202                 n->vqs[i].sock_hlen = 0;
 203         }
 204
 205 }
 206
 207 static void vhost_net_tx_packet(struct vhost_net *net)
 208 {
 209         ++net->tx_packets;
 210         if (net->tx_packets < 1024)
 211                 return;
 212         net->tx_packets = 0;
 213         net->tx_zcopy_err = 0;
 214 }
 215
 216 static void vhost_net_tx_err(struct vhost_net *net)
 217 {
 218         ++net->tx_zcopy_err;
 219 }
 220
 221 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 222 {
 223         /* TX flush waits for outstanding DMAs to be done.
 224          * Don't start new DMAs.
 225          */
 226         return !net->tx_flush &&
 227                 net->tx_packets / 64 >= net->tx_zcopy_err;
 228 }
 229
 230 static bool vhost_sock_zcopy(struct socket *sock)
 231 {
 232         return unlikely(experimental_zcopytx) &&
 233                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 234 }
 235
 236 /* Pop first len bytes from iovec. Return number of segments used. */
 237 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 238                           size_t len, int iov_count)
 239 {
 240         int seg = 0;
 241         size_t size;
 242
 243         while (len && seg < iov_count) {
 244                 size = min(from->iov_len, len);
 245                 to->iov_base = from->iov_base;
 246                 to->iov_len = size;
 247                 from->iov_len -= size;
 248                 from->iov_base += size;
 249                 len -= size;
 250                 ++from;
 251                 ++to;
 252                 ++seg;
 253         }
 254         return seg;
 255 }
 256 /* Copy iovec entries for len bytes from iovec. */
 257 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 258                            size_t len, int iovcount)
 259 {
 260         int seg = 0;
 261         size_t size;
 262
 263         while (len && seg < iovcount) {
 264                 size = min(from->iov_len, len);
 265                 to->iov_base = from->iov_base;
 266                 to->iov_len = size;
 267                 len -= size;
 268                 ++from;
 269                 ++to;
 270                 ++seg;
 271         }
 272 }
 273
 274 /* In case of DMA done not in order in lower device driver for some reason.
 275  * upend_idx is used to track end of used idx, done_idx is used to track head
 276  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 277  * guest used idx.
 278  */
 279 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 280                                        struct vhost_virtqueue *vq)
 281 {
 282         struct vhost_net_virtqueue *nvq =
 283                 container_of(vq, struct vhost_net_virtqueue, vq);
 284         int i, add;
 285         int j = 0;
 286
 287         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 288                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 289                         vhost_net_tx_err(net);
 290                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 291                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 292                         ++j;
 293                 } else
 294                         break;
 295         }
 296         while (j) {
 297                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 298                 vhost_add_used_and_signal_n(vq->dev, vq,
 299                                             &vq->heads[nvq->done_idx], add);
 300                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 301                 j -= add;
 302         }
 303 }
 304
 305 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 306 {
 307         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 308         struct vhost_virtqueue *vq = ubufs->vq;
 309         int cnt;
 310
 311         /* 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         cnt = 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          * We also trigger polling periodically after each 16 packets
 320          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 321          * less than 10% of times).
 322          */
 323         if (cnt <= 1 || !(cnt % 16))
 324                 vhost_poll_queue(&vq->poll);
 325 }
 326
 327 /* Expects to be always run from workqueue - which acts as
 328  * read-size critical section for our kind of RCU. */
 329 static void handle_tx(struct vhost_net *net)
 330 {
 331         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 332         struct vhost_virtqueue *vq = &nvq->vq;
 333         unsigned out, in, s;
 334         int head;
 335         struct msghdr msg = {
 336                 .msg_name = NULL,
 337                 .msg_namelen = 0,
 338                 .msg_control = NULL,
 339                 .msg_controllen = 0,
 340                 .msg_iov = vq->iov,
 341                 .msg_flags = MSG_DONTWAIT,
 342         };
 343         size_t len, total_len = 0;
 344         int err;
 345         size_t hdr_size;
 346         struct socket *sock;
 347         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 348         bool zcopy, zcopy_used;
 349
 350         mutex_lock(&vq->mutex);
 351         sock = vq->private_data;
 352         if (!sock)
 353                 goto out;
 354
 355         vhost_disable_notify(&net->dev, vq);
 356
 357         hdr_size = nvq->vhost_hlen;
 358         zcopy = nvq->ubufs;
 359
 360         for (;;) {
 361                 /* Release DMAs done buffers first */
 362                 if (zcopy)
 363                         vhost_zerocopy_signal_used(net, vq);
 364
 365                 /* If more outstanding DMAs, queue the work.
 366                  * Handle upend_idx wrap around
 367                  */
 368                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
 369                               % UIO_MAXIOV == nvq->done_idx))
 370                         break;
 371
 372                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 373                                          ARRAY_SIZE(vq->iov),
 374                                          &out, &in,
 375                                          NULL, NULL);
 376                 /* On error, stop handling until the next kick. */
 377                 if (unlikely(head < 0))
 378                         break;
 379                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 380                 if (head == vq->num) {
 381                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 382                                 vhost_disable_notify(&net->dev, vq);
 383                                 continue;
 384                         }
 385                         break;
 386                 }
 387                 if (in) {
 388                         vq_err(vq, "Unexpected descriptor format for TX: "
 389                                "out %d, int %d\n", out, in);
 390                         break;
 391                 }
 392                 /* Skip header. TODO: support TSO. */
 393                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 394                 msg.msg_iovlen = out;
 395                 len = iov_length(vq->iov, out);
 396                 /* Sanity check */
 397                 if (!len) {
 398                         vq_err(vq, "Unexpected header len for TX: "
 399                                "%zd expected %zd\n",
 400                                iov_length(nvq->hdr, s), hdr_size);
 401                         break;
 402                 }
 403
 404                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 405                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 406                                       nvq->done_idx
 407                                    && vhost_net_tx_select_zcopy(net);
 408
 409                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 410                 if (zcopy_used) {
 411                         struct ubuf_info *ubuf;
 412                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 413
 414                         vq->heads[nvq->upend_idx].id = head;
 415                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 416                         ubuf->callback = vhost_zerocopy_callback;
 417                         ubuf->ctx = nvq->ubufs;
 418                         ubuf->desc = nvq->upend_idx;
 419                         msg.msg_control = ubuf;
 420                         msg.msg_controllen = sizeof(ubuf);
 421                         ubufs = nvq->ubufs;
 422                         atomic_inc(&ubufs->refcount);
 423                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 424                 } else {
 425                         msg.msg_control = NULL;
 426                         ubufs = NULL;
 427                 }
 428                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 429                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 430                 if (unlikely(err < 0)) {
 431                         if (zcopy_used) {
 432                                 vhost_net_ubuf_put(ubufs);
 433                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 434                                         % UIO_MAXIOV;
 435                         }
 436                         vhost_discard_vq_desc(vq, 1);
 437                         break;
 438                 }
 439                 if (err != len)
 440                         pr_debug("Truncated TX packet: "
 441                                  " len %d != %zd\n", err, len);
 442                 if (!zcopy_used)
 443                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 444                 else
 445                         vhost_zerocopy_signal_used(net, vq);
 446                 total_len += len;
 447                 vhost_net_tx_packet(net);
 448                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 449                         vhost_poll_queue(&vq->poll);
 450                         break;
 451                 }
 452         }
 453 out:
 454         mutex_unlock(&vq->mutex);
 455 }
 456
 457 static int peek_head_len(struct sock *sk)
 458 {
 459         struct sk_buff *head;
 460         int len = 0;
 461         unsigned long flags;
 462
 463         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 464         head = skb_peek(&sk->sk_receive_queue);
 465         if (likely(head)) {
 466                 len = head->len;
 467                 if (vlan_tx_tag_present(head))
 468                         len += VLAN_HLEN;
 469         }
 470
 471         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 472         return len;
 473 }
 474
 475 /* This is a multi-buffer version of vhost_get_desc, that works if
 476  *      vq has read descriptors only.
 477  * @vq          - the relevant virtqueue
 478  * @datalen     - data length we'll be reading
 479  * @iovcount    - returned count of io vectors we fill
 480  * @log         - vhost log
 481  * @log_num     - log offset
 482  * @quota       - headcount quota, 1 for big buffer
 483  *      returns number of buffer heads allocated, negative on error
 484  */
 485 static int get_rx_bufs(struct vhost_virtqueue *vq,
 486                        struct vring_used_elem *heads,
 487                        int datalen,
 488                        unsigned *iovcount,
 489                        struct vhost_log *log,
 490                        unsigned *log_num,
 491                        unsigned int quota)
 492 {
 493         unsigned int out, in;
 494         int seg = 0;
 495         int headcount = 0;
 496         unsigned d;
 497         int r, nlogs = 0;
 498
 499         while (datalen > 0 && headcount < quota) {
 500                 if (unlikely(seg >= UIO_MAXIOV)) {
 501                         r = -ENOBUFS;
 502                         goto err;
 503                 }
 504                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 505                                       ARRAY_SIZE(vq->iov) - seg, &out,
 506                                       &in, log, log_num);
 507                 if (d == vq->num) {
 508                         r = 0;
 509                         goto err;
 510                 }
 511                 if (unlikely(out || in <= 0)) {
 512                         vq_err(vq, "unexpected descriptor format for RX: "
 513                                 "out %d, in %d\n", out, in);
 514                         r = -EINVAL;
 515                         goto err;
 516                 }
 517                 if (unlikely(log)) {
 518                         nlogs += *log_num;
 519                         log += *log_num;
 520                 }
 521                 heads[headcount].id = d;
 522                 heads[headcount].len = iov_length(vq->iov + seg, in);
 523                 datalen -= heads[headcount].len;
 524                 ++headcount;
 525                 seg += in;
 526         }
 527         heads[headcount - 1].len += datalen;
 528         *iovcount = seg;
 529         if (unlikely(log))
 530                 *log_num = nlogs;
 531         return headcount;
 532 err:
 533         vhost_discard_vq_desc(vq, headcount);
 534         return r;
 535 }
 536
 537 /* Expects to be always run from workqueue - which acts as
 538  * read-size critical section for our kind of RCU. */
 539 static void handle_rx(struct vhost_net *net)
 540 {
 541         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 542         struct vhost_virtqueue *vq = &nvq->vq;
 543         unsigned uninitialized_var(in), log;
 544         struct vhost_log *vq_log;
 545         struct msghdr msg = {
 546                 .msg_name = NULL,
 547                 .msg_namelen = 0,
 548                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 549                 .msg_controllen = 0,
 550                 .msg_iov = vq->iov,
 551                 .msg_flags = MSG_DONTWAIT,
 552         };
 553         struct virtio_net_hdr_mrg_rxbuf hdr = {
 554                 .hdr.flags = 0,
 555                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 556         };
 557         size_t total_len = 0;
 558         int err, mergeable;
 559         s16 headcount;
 560         size_t vhost_hlen, sock_hlen;
 561         size_t vhost_len, sock_len;
 562         struct socket *sock;
 563
 564         mutex_lock(&vq->mutex);
 565         sock = vq->private_data;
 566         if (!sock)
 567                 goto out;
 568         vhost_disable_notify(&net->dev, vq);
 569
 570         vhost_hlen = nvq->vhost_hlen;
 571         sock_hlen = nvq->sock_hlen;
 572
 573         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 574                 vq->log : NULL;
 575         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 576
 577         while ((sock_len = peek_head_len(sock->sk))) {
 578                 sock_len += sock_hlen;
 579                 vhost_len = sock_len + vhost_hlen;
 580                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 581                                         &in, vq_log, &log,
 582                                         likely(mergeable) ? UIO_MAXIOV : 1);
 583                 /* On error, stop handling until the next kick. */
 584                 if (unlikely(headcount < 0))
 585                         break;
 586                 /* OK, now we need to know about added descriptors. */
 587                 if (!headcount) {
 588                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 589                                 /* They have slipped one in as we were
 590                                  * doing that: check again. */
 591                                 vhost_disable_notify(&net->dev, vq);
 592                                 continue;
 593                         }
 594                         /* Nothing new?  Wait for eventfd to tell us
 595                          * they refilled. */
 596                         break;
 597                 }
 598                 /* We don't need to be notified again. */
 599                 if (unlikely((vhost_hlen)))
 600                         /* Skip header. TODO: support TSO. */
 601                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 602                 else
 603                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 604                          * needed because recvmsg can modify msg_iov. */
 605                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 606                 msg.msg_iovlen = in;
 607                 err = sock->ops->recvmsg(NULL, sock, &msg,
 608                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 609                 /* Userspace might have consumed the packet meanwhile:
 610                  * it's not supposed to do this usually, but might be hard
 611                  * to prevent. Discard data we got (if any) and keep going. */
 612                 if (unlikely(err != sock_len)) {
 613                         pr_debug("Discarded rx packet: "
 614                                  " len %d, expected %zd\n", err, sock_len);
 615                         vhost_discard_vq_desc(vq, headcount);
 616                         continue;
 617                 }
 618                 if (unlikely(vhost_hlen) &&
 619                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 620                                       vhost_hlen)) {
 621                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 622                                vq->iov->iov_base);
 623                         break;
 624                 }
 625                 /* TODO: Should check and handle checksum. */
 626                 if (likely(mergeable) &&
 627                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 628                                       offsetof(typeof(hdr), num_buffers),
 629                                       sizeof hdr.num_buffers)) {
 630                         vq_err(vq, "Failed num_buffers write");
 631                         vhost_discard_vq_desc(vq, headcount);
 632                         break;
 633                 }
 634                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 635                                             headcount);
 636                 if (unlikely(vq_log))
 637                         vhost_log_write(vq, vq_log, log, vhost_len);
 638                 total_len += vhost_len;
 639                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 640                         vhost_poll_queue(&vq->poll);
 641                         break;
 642                 }
 643         }
 644 out:
 645         mutex_unlock(&vq->mutex);
 646 }
 647
 648 static void handle_tx_kick(struct vhost_work *work)
 649 {
 650         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 651                                                   poll.work);
 652         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 653
 654         handle_tx(net);
 655 }
 656
 657 static void handle_rx_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_rx(net);
 664 }
 665
 666 static void handle_tx_net(struct vhost_work *work)
 667 {
 668         struct vhost_net *net = container_of(work, struct vhost_net,
 669                                              poll[VHOST_NET_VQ_TX].work);
 670         handle_tx(net);
 671 }
 672
 673 static void handle_rx_net(struct vhost_work *work)
 674 {
 675         struct vhost_net *net = container_of(work, struct vhost_net,
 676                                              poll[VHOST_NET_VQ_RX].work);
 677         handle_rx(net);
 678 }
 679
 680 static int vhost_net_open(struct inode *inode, struct file *f)
 681 {
 682         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 683         struct vhost_dev *dev;
 684         struct vhost_virtqueue **vqs;
 685         int i;
 686
 687         if (!n)
 688                 return -ENOMEM;
 689         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 690         if (!vqs) {
 691                 kfree(n);
 692                 return -ENOMEM;
 693         }
 694
 695         dev = &n->dev;
 696         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 697         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 698         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 699         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 700         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 701                 n->vqs[i].ubufs = NULL;
 702                 n->vqs[i].ubuf_info = NULL;
 703                 n->vqs[i].upend_idx = 0;
 704                 n->vqs[i].done_idx = 0;
 705                 n->vqs[i].vhost_hlen = 0;
 706                 n->vqs[i].sock_hlen = 0;
 707         }
 708         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 709
 710         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 711         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 712
 713         f->private_data = n;
 714
 715         return 0;
 716 }
 717
 718 static void vhost_net_disable_vq(struct vhost_net *n,
 719                                  struct vhost_virtqueue *vq)
 720 {
 721         struct vhost_net_virtqueue *nvq =
 722                 container_of(vq, struct vhost_net_virtqueue, vq);
 723         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 724         if (!vq->private_data)
 725                 return;
 726         vhost_poll_stop(poll);
 727 }
 728
 729 static int vhost_net_enable_vq(struct vhost_net *n,
 730                                 struct vhost_virtqueue *vq)
 731 {
 732         struct vhost_net_virtqueue *nvq =
 733                 container_of(vq, struct vhost_net_virtqueue, vq);
 734         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 735         struct socket *sock;
 736
 737         sock = vq->private_data;
 738         if (!sock)
 739                 return 0;
 740
 741         return vhost_poll_start(poll, sock->file);
 742 }
 743
 744 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 745                                         struct vhost_virtqueue *vq)
 746 {
 747         struct socket *sock;
 748
 749         mutex_lock(&vq->mutex);
 750         sock = vq->private_data;
 751         vhost_net_disable_vq(n, vq);
 752         vq->private_data = NULL;
 753         mutex_unlock(&vq->mutex);
 754         return sock;
 755 }
 756
 757 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 758                            struct socket **rx_sock)
 759 {
 760         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 761         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 762 }
 763
 764 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 765 {
 766         vhost_poll_flush(n->poll + index);
 767         vhost_poll_flush(&n->vqs[index].vq.poll);
 768 }
 769
 770 static void vhost_net_flush(struct vhost_net *n)
 771 {
 772         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 773         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 774         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 775                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 776                 n->tx_flush = true;
 777                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 778                 /* Wait for all lower device DMAs done. */
 779                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 780                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 781                 n->tx_flush = false;
 782                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 783                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 784         }
 785 }
 786
 787 static int vhost_net_release(struct inode *inode, struct file *f)
 788 {
 789         struct vhost_net *n = f->private_data;
 790         struct socket *tx_sock;
 791         struct socket *rx_sock;
 792
 793         vhost_net_stop(n, &tx_sock, &rx_sock);
 794         vhost_net_flush(n);
 795         vhost_dev_stop(&n->dev);
 796         vhost_dev_cleanup(&n->dev, false);
 797         vhost_net_vq_reset(n);
 798         if (tx_sock)
 799                 fput(tx_sock->file);
 800         if (rx_sock)
 801                 fput(rx_sock->file);
 802         /* We do an extra flush before freeing memory,
 803          * since jobs can re-queue themselves. */
 804         vhost_net_flush(n);
 805         kfree(n->dev.vqs);
 806         kfree(n);
 807         return 0;
 808 }
 809
 810 static struct socket *get_raw_socket(int fd)
 811 {
 812         struct {
 813                 struct sockaddr_ll sa;
 814                 char  buf[MAX_ADDR_LEN];
 815         } uaddr;
 816         int uaddr_len = sizeof uaddr, r;
 817         struct socket *sock = sockfd_lookup(fd, &r);
 818
 819         if (!sock)
 820                 return ERR_PTR(-ENOTSOCK);
 821
 822         /* Parameter checking */
 823         if (sock->sk->sk_type != SOCK_RAW) {
 824                 r = -ESOCKTNOSUPPORT;
 825                 goto err;
 826         }
 827
 828         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 829                                &uaddr_len, 0);
 830         if (r)
 831                 goto err;
 832
 833         if (uaddr.sa.sll_family != AF_PACKET) {
 834                 r = -EPFNOSUPPORT;
 835                 goto err;
 836         }
 837         return sock;
 838 err:
 839         fput(sock->file);
 840         return ERR_PTR(r);
 841 }
 842
 843 static struct socket *get_tap_socket(int fd)
 844 {
 845         struct file *file = fget(fd);
 846         struct socket *sock;
 847
 848         if (!file)
 849                 return ERR_PTR(-EBADF);
 850         sock = tun_get_socket(file);
 851         if (!IS_ERR(sock))
 852                 return sock;
 853         sock = macvtap_get_socket(file);
 854         if (IS_ERR(sock))
 855                 fput(file);
 856         return sock;
 857 }
 858
 859 static struct socket *get_socket(int fd)
 860 {
 861         struct socket *sock;
 862
 863         /* special case to disable backend */
 864         if (fd == -1)
 865                 return NULL;
 866         sock = get_raw_socket(fd);
 867         if (!IS_ERR(sock))
 868                 return sock;
 869         sock = get_tap_socket(fd);
 870         if (!IS_ERR(sock))
 871                 return sock;
 872         return ERR_PTR(-ENOTSOCK);
 873 }
 874
 875 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 876 {
 877         struct socket *sock, *oldsock;
 878         struct vhost_virtqueue *vq;
 879         struct vhost_net_virtqueue *nvq;
 880         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 881         int r;
 882
 883         mutex_lock(&n->dev.mutex);
 884         r = vhost_dev_check_owner(&n->dev);
 885         if (r)
 886                 goto err;
 887
 888         if (index >= VHOST_NET_VQ_MAX) {
 889                 r = -ENOBUFS;
 890                 goto err;
 891         }
 892         vq = &n->vqs[index].vq;
 893         nvq = &n->vqs[index];
 894         mutex_lock(&vq->mutex);
 895
 896         /* Verify that ring has been setup correctly. */
 897         if (!vhost_vq_access_ok(vq)) {
 898                 r = -EFAULT;
 899                 goto err_vq;
 900         }
 901         sock = get_socket(fd);
 902         if (IS_ERR(sock)) {
 903                 r = PTR_ERR(sock);
 904                 goto err_vq;
 905         }
 906
 907         /* start polling new socket */
 908         oldsock = vq->private_data;
 909         if (sock != oldsock) {
 910                 ubufs = vhost_net_ubuf_alloc(vq,
 911                                              sock && vhost_sock_zcopy(sock));
 912                 if (IS_ERR(ubufs)) {
 913                         r = PTR_ERR(ubufs);
 914                         goto err_ubufs;
 915                 }
 916
 917                 vhost_net_disable_vq(n, vq);
 918                 vq->private_data = sock;
 919                 r = vhost_init_used(vq);
 920                 if (r)
 921                         goto err_used;
 922                 r = vhost_net_enable_vq(n, vq);
 923                 if (r)
 924                         goto err_used;
 925
 926                 oldubufs = nvq->ubufs;
 927                 nvq->ubufs = ubufs;
 928
 929                 n->tx_packets = 0;
 930                 n->tx_zcopy_err = 0;
 931                 n->tx_flush = false;
 932         }
 933
 934         mutex_unlock(&vq->mutex);
 935
 936         if (oldubufs) {
 937                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 938                 mutex_lock(&vq->mutex);
 939                 vhost_zerocopy_signal_used(n, vq);
 940                 mutex_unlock(&vq->mutex);
 941         }
 942
 943         if (oldsock) {
 944                 vhost_net_flush_vq(n, index);
 945                 fput(oldsock->file);
 946         }
 947
 948         mutex_unlock(&n->dev.mutex);
 949         return 0;
 950
 951 err_used:
 952         vq->private_data = oldsock;
 953         vhost_net_enable_vq(n, vq);
 954         if (ubufs)
 955                 vhost_net_ubuf_put_wait_and_free(ubufs);
 956 err_ubufs:
 957         fput(sock->file);
 958 err_vq:
 959         mutex_unlock(&vq->mutex);
 960 err:
 961         mutex_unlock(&n->dev.mutex);
 962         return r;
 963 }
 964
 965 static long vhost_net_reset_owner(struct vhost_net *n)
 966 {
 967         struct socket *tx_sock = NULL;
 968         struct socket *rx_sock = NULL;
 969         long err;
 970         struct vhost_memory *memory;
 971
 972         mutex_lock(&n->dev.mutex);
 973         err = vhost_dev_check_owner(&n->dev);
 974         if (err)
 975                 goto done;
 976         memory = vhost_dev_reset_owner_prepare();
 977         if (!memory) {
 978                 err = -ENOMEM;
 979                 goto done;
 980         }
 981         vhost_net_stop(n, &tx_sock, &rx_sock);
 982         vhost_net_flush(n);
 983         vhost_dev_reset_owner(&n->dev, memory);
 984         vhost_net_vq_reset(n);
 985 done:
 986         mutex_unlock(&n->dev.mutex);
 987         if (tx_sock)
 988                 fput(tx_sock->file);
 989         if (rx_sock)
 990                 fput(rx_sock->file);
 991         return err;
 992 }
 993
 994 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 995 {
 996         size_t vhost_hlen, sock_hlen, hdr_len;
 997         int i;
 998
 999         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
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         n->dev.acked_features = features;
1018         smp_wmb();
1019         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1020                 mutex_lock(&n->vqs[i].vq.mutex);
1021                 n->vqs[i].vhost_hlen = vhost_hlen;
1022                 n->vqs[i].sock_hlen = sock_hlen;
1023                 mutex_unlock(&n->vqs[i].vq.mutex);
1024         }
1025         vhost_net_flush(n);
1026         mutex_unlock(&n->dev.mutex);
1027         return 0;
1028 }
1029
1030 static long vhost_net_set_owner(struct vhost_net *n)
1031 {
1032         int r;
1033
1034         mutex_lock(&n->dev.mutex);
1035         if (vhost_dev_has_owner(&n->dev)) {
1036                 r = -EBUSY;
1037                 goto out;
1038         }
1039         r = vhost_net_set_ubuf_info(n);
1040         if (r)
1041                 goto out;
1042         r = vhost_dev_set_owner(&n->dev);
1043         if (r)
1044                 vhost_net_clear_ubuf_info(n);
1045         vhost_net_flush(n);
1046 out:
1047         mutex_unlock(&n->dev.mutex);
1048         return r;
1049 }
1050
1051 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1052                             unsigned long arg)
1053 {
1054         struct vhost_net *n = f->private_data;
1055         void __user *argp = (void __user *)arg;
1056         u64 __user *featurep = argp;
1057         struct vhost_vring_file backend;
1058         u64 features;
1059         int r;
1060
1061         switch (ioctl) {
1062         case VHOST_NET_SET_BACKEND:
1063                 if (copy_from_user(&backend, argp, sizeof backend))
1064                         return -EFAULT;
1065                 return vhost_net_set_backend(n, backend.index, backend.fd);
1066         case VHOST_GET_FEATURES:
1067                 features = VHOST_NET_FEATURES;
1068                 if (copy_to_user(featurep, &features, sizeof features))
1069                         return -EFAULT;
1070                 return 0;
1071         case VHOST_SET_FEATURES:
1072                 if (copy_from_user(&features, featurep, sizeof features))
1073                         return -EFAULT;
1074                 if (features & ~VHOST_NET_FEATURES)
1075                         return -EOPNOTSUPP;
1076                 return vhost_net_set_features(n, features);
1077         case VHOST_RESET_OWNER:
1078                 return vhost_net_reset_owner(n);
1079         case VHOST_SET_OWNER:
1080                 return vhost_net_set_owner(n);
1081         default:
1082                 mutex_lock(&n->dev.mutex);
1083                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1084                 if (r == -ENOIOCTLCMD)
1085                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1086                 else
1087                         vhost_net_flush(n);
1088                 mutex_unlock(&n->dev.mutex);
1089                 return r;
1090         }
1091 }
1092
1093 #ifdef CONFIG_COMPAT
1094 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1095                                    unsigned long arg)
1096 {
1097         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1098 }
1099 #endif
1100
1101 static const struct file_operations vhost_net_fops = {
1102         .owner          = THIS_MODULE,
1103         .release        = vhost_net_release,
1104         .unlocked_ioctl = vhost_net_ioctl,
1105 #ifdef CONFIG_COMPAT
1106         .compat_ioctl   = vhost_net_compat_ioctl,
1107 #endif
1108         .open           = vhost_net_open,
1109         .llseek         = noop_llseek,
1110 };
1111
1112 static struct miscdevice vhost_net_misc = {
1113         .minor = VHOST_NET_MINOR,
1114         .name = "vhost-net",
1115         .fops = &vhost_net_fops,
1116 };
1117
1118 static int vhost_net_init(void)
1119 {
1120         if (experimental_zcopytx)
1121                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1122         return misc_register(&vhost_net_misc);
1123 }
1124 module_init(vhost_net_init);
1125
1126 static void vhost_net_exit(void)
1127 {
1128         misc_deregister(&vhost_net_misc);
1129 }
1130 module_exit(vhost_net_exit);
1131
1132 MODULE_VERSION("0.0.1");
1133 MODULE_LICENSE("GPL v2");
1134 MODULE_AUTHOR("Michael S. Tsirkin");
1135 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1136 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1137 MODULE_ALIAS("devname:vhost-net");