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/rcupdate.h>
  19 #include <linux/file.h>
  20 #include <linux/slab.h>
  21
  22 #include <linux/net.h>
  23 #include <linux/if_packet.h>
  24 #include <linux/if_arp.h>
  25 #include <linux/if_tun.h>
  26 #include <linux/if_macvlan.h>
  27 #include <linux/if_vlan.h>
  28
  29 #include <net/sock.h>
  30
  31 #include "vhost.h"
  32
  33 static int experimental_zcopytx = 1;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  36                                        " 1 -Enable; 0 - Disable");
  37
  38 /* Max number of bytes transferred before requeueing the job.
  39  * Using this limit prevents one virtqueue from starving others. */
  40 #define VHOST_NET_WEIGHT 0x80000
  41
  42 /* MAX number of TX used buffers for outstanding zerocopy */
  43 #define VHOST_MAX_PEND 128
  44 #define VHOST_GOODCOPY_LEN 256
  45
  46 /*
  47  * For transmit, used buffer len is unused; we override it to track buffer
  48  * status internally; used for zerocopy tx only.
  49  */
  50 /* Lower device DMA failed */
  51 #define VHOST_DMA_FAILED_LEN    3
  52 /* Lower device DMA done */
  53 #define VHOST_DMA_DONE_LEN      2
  54 /* Lower device DMA in progress */
  55 #define VHOST_DMA_IN_PROGRESS   1
  56 /* Buffer unused */
  57 #define VHOST_DMA_CLEAR_LEN     0
  58
  59 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
  60
  61 enum {
  62         VHOST_NET_VQ_RX = 0,
  63         VHOST_NET_VQ_TX = 1,
  64         VHOST_NET_VQ_MAX = 2,
  65 };
  66
  67 struct vhost_ubuf_ref {
  68         struct kref kref;
  69         wait_queue_head_t wait;
  70         struct vhost_virtqueue *vq;
  71 };
  72
  73 struct vhost_net_virtqueue {
  74         struct vhost_virtqueue vq;
  75         /* vhost zerocopy support fields below: */
  76         /* last used idx for outstanding DMA zerocopy buffers */
  77         int upend_idx;
  78         /* first used idx for DMA done zerocopy buffers */
  79         int done_idx;
  80         /* an array of userspace buffers info */
  81         struct ubuf_info *ubuf_info;
  82         /* Reference counting for outstanding ubufs.
  83          * Protected by vq mutex. Writers must also take device mutex. */
  84         struct vhost_ubuf_ref *ubufs;
  85 };
  86
  87 struct vhost_net {
  88         struct vhost_dev dev;
  89         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
  90         struct vhost_poll poll[VHOST_NET_VQ_MAX];
  91         /* Number of TX recently submitted.
  92          * Protected by tx vq lock. */
  93         unsigned tx_packets;
  94         /* Number of times zerocopy TX recently failed.
  95          * Protected by tx vq lock. */
  96         unsigned tx_zcopy_err;
  97         /* Flush in progress. Protected by tx vq lock. */
  98         bool tx_flush;
  99 };
 100
 101 static unsigned vhost_zcopy_mask __read_mostly;
 102
 103 void vhost_enable_zcopy(int vq)
 104 {
 105         vhost_zcopy_mask |= 0x1 << vq;
 106 }
 107
 108 static void vhost_zerocopy_done_signal(struct kref *kref)
 109 {
 110         struct vhost_ubuf_ref *ubufs = container_of(kref, struct vhost_ubuf_ref,
 111                                                     kref);
 112         wake_up(&ubufs->wait);
 113 }
 114
 115 struct vhost_ubuf_ref *vhost_ubuf_alloc(struct vhost_virtqueue *vq,
 116                                         bool zcopy)
 117 {
 118         struct vhost_ubuf_ref *ubufs;
 119         /* No zero copy backend? Nothing to count. */
 120         if (!zcopy)
 121                 return NULL;
 122         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 123         if (!ubufs)
 124                 return ERR_PTR(-ENOMEM);
 125         kref_init(&ubufs->kref);
 126         init_waitqueue_head(&ubufs->wait);
 127         ubufs->vq = vq;
 128         return ubufs;
 129 }
 130
 131 void vhost_ubuf_put(struct vhost_ubuf_ref *ubufs)
 132 {
 133         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
 134 }
 135
 136 void vhost_ubuf_put_and_wait(struct vhost_ubuf_ref *ubufs)
 137 {
 138         kref_put(&ubufs->kref, vhost_zerocopy_done_signal);
 139         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
 140         kfree(ubufs);
 141 }
 142
 143 int vhost_net_set_ubuf_info(struct vhost_net *n)
 144 {
 145         bool zcopy;
 146         int i;
 147
 148         for (i = 0; i < n->dev.nvqs; ++i) {
 149                 zcopy = vhost_zcopy_mask & (0x1 << i);
 150                 if (!zcopy)
 151                         continue;
 152                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 153                                               UIO_MAXIOV, GFP_KERNEL);
 154                 if  (!n->vqs[i].ubuf_info)
 155                         goto err;
 156         }
 157         return 0;
 158
 159 err:
 160         while (i--) {
 161                 zcopy = vhost_zcopy_mask & (0x1 << i);
 162                 if (!zcopy)
 163                         continue;
 164                 kfree(n->vqs[i].ubuf_info);
 165         }
 166         return -ENOMEM;
 167 }
 168
 169 void vhost_net_reset_ubuf_info(struct vhost_net *n)
 170 {
 171         int i;
 172
 173         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 174                 n->vqs[i].done_idx = 0;
 175                 n->vqs[i].upend_idx = 0;
 176                 n->vqs[i].ubufs = NULL;
 177                 kfree(n->vqs[i].ubuf_info);
 178                 n->vqs[i].ubuf_info = NULL;
 179         }
 180
 181 }
 182
 183 static void vhost_net_tx_packet(struct vhost_net *net)
 184 {
 185         ++net->tx_packets;
 186         if (net->tx_packets < 1024)
 187                 return;
 188         net->tx_packets = 0;
 189         net->tx_zcopy_err = 0;
 190 }
 191
 192 static void vhost_net_tx_err(struct vhost_net *net)
 193 {
 194         ++net->tx_zcopy_err;
 195 }
 196
 197 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 198 {
 199         /* TX flush waits for outstanding DMAs to be done.
 200          * Don't start new DMAs.
 201          */
 202         return !net->tx_flush &&
 203                 net->tx_packets / 64 >= net->tx_zcopy_err;
 204 }
 205
 206 static bool vhost_sock_zcopy(struct socket *sock)
 207 {
 208         return unlikely(experimental_zcopytx) &&
 209                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 210 }
 211
 212 /* Pop first len bytes from iovec. Return number of segments used. */
 213 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 214                           size_t len, int iov_count)
 215 {
 216         int seg = 0;
 217         size_t size;
 218
 219         while (len && seg < iov_count) {
 220                 size = min(from->iov_len, len);
 221                 to->iov_base = from->iov_base;
 222                 to->iov_len = size;
 223                 from->iov_len -= size;
 224                 from->iov_base += size;
 225                 len -= size;
 226                 ++from;
 227                 ++to;
 228                 ++seg;
 229         }
 230         return seg;
 231 }
 232 /* Copy iovec entries for len bytes from iovec. */
 233 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 234                            size_t len, int iovcount)
 235 {
 236         int seg = 0;
 237         size_t size;
 238
 239         while (len && seg < iovcount) {
 240                 size = min(from->iov_len, len);
 241                 to->iov_base = from->iov_base;
 242                 to->iov_len = size;
 243                 len -= size;
 244                 ++from;
 245                 ++to;
 246                 ++seg;
 247         }
 248 }
 249
 250 /* In case of DMA done not in order in lower device driver for some reason.
 251  * upend_idx is used to track end of used idx, done_idx is used to track head
 252  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 253  * guest used idx.
 254  */
 255 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 256                                       struct vhost_virtqueue *vq)
 257 {
 258         struct vhost_net_virtqueue *nvq =
 259                 container_of(vq, struct vhost_net_virtqueue, vq);
 260         int i;
 261         int j = 0;
 262
 263         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 264                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 265                         vhost_net_tx_err(net);
 266                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 267                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 268                         vhost_add_used_and_signal(vq->dev, vq,
 269                                                   vq->heads[i].id, 0);
 270                         ++j;
 271                 } else
 272                         break;
 273         }
 274         if (j)
 275                 nvq->done_idx = i;
 276         return j;
 277 }
 278
 279 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 280 {
 281         struct vhost_ubuf_ref *ubufs = ubuf->ctx;
 282         struct vhost_virtqueue *vq = ubufs->vq;
 283         int cnt = atomic_read(&ubufs->kref.refcount);
 284
 285         /*
 286          * Trigger polling thread if guest stopped submitting new buffers:
 287          * in this case, the refcount after decrement will eventually reach 1
 288          * so here it is 2.
 289          * We also trigger polling periodically after each 16 packets
 290          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 291          * less than 10% of times).
 292          */
 293         if (cnt <= 2 || !(cnt % 16))
 294                 vhost_poll_queue(&vq->poll);
 295         /* set len to mark this desc buffers done DMA */
 296         vq->heads[ubuf->desc].len = success ?
 297                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 298         vhost_ubuf_put(ubufs);
 299 }
 300
 301 /* Expects to be always run from workqueue - which acts as
 302  * read-size critical section for our kind of RCU. */
 303 static void handle_tx(struct vhost_net *net)
 304 {
 305         struct vhost_virtqueue *vq = &net->vqs[VHOST_NET_VQ_TX].vq;
 306         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 307         unsigned out, in, s;
 308         int head;
 309         struct msghdr msg = {
 310                 .msg_name = NULL,
 311                 .msg_namelen = 0,
 312                 .msg_control = NULL,
 313                 .msg_controllen = 0,
 314                 .msg_iov = vq->iov,
 315                 .msg_flags = MSG_DONTWAIT,
 316         };
 317         size_t len, total_len = 0;
 318         int err;
 319         size_t hdr_size;
 320         struct socket *sock;
 321         struct vhost_ubuf_ref *uninitialized_var(ubufs);
 322         bool zcopy, zcopy_used;
 323
 324         /* TODO: check that we are running from vhost_worker? */
 325         sock = rcu_dereference_check(vq->private_data, 1);
 326         if (!sock)
 327                 return;
 328
 329         mutex_lock(&vq->mutex);
 330         vhost_disable_notify(&net->dev, vq);
 331
 332         hdr_size = vq->vhost_hlen;
 333         zcopy = nvq->ubufs;
 334
 335         for (;;) {
 336                 /* Release DMAs done buffers first */
 337                 if (zcopy)
 338                         vhost_zerocopy_signal_used(net, vq);
 339
 340                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 341                                          ARRAY_SIZE(vq->iov),
 342                                          &out, &in,
 343                                          NULL, NULL);
 344                 /* On error, stop handling until the next kick. */
 345                 if (unlikely(head < 0))
 346                         break;
 347                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 348                 if (head == vq->num) {
 349                         int num_pends;
 350
 351                         /* If more outstanding DMAs, queue the work.
 352                          * Handle upend_idx wrap around
 353                          */
 354                         num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
 355                                     (nvq->upend_idx - nvq->done_idx) :
 356                                     (nvq->upend_idx + UIO_MAXIOV -
 357                                      nvq->done_idx);
 358                         if (unlikely(num_pends > VHOST_MAX_PEND))
 359                                 break;
 360                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 361                                 vhost_disable_notify(&net->dev, vq);
 362                                 continue;
 363                         }
 364                         break;
 365                 }
 366                 if (in) {
 367                         vq_err(vq, "Unexpected descriptor format for TX: "
 368                                "out %d, int %d\n", out, in);
 369                         break;
 370                 }
 371                 /* Skip header. TODO: support TSO. */
 372                 s = move_iovec_hdr(vq->iov, vq->hdr, hdr_size, out);
 373                 msg.msg_iovlen = out;
 374                 len = iov_length(vq->iov, out);
 375                 /* Sanity check */
 376                 if (!len) {
 377                         vq_err(vq, "Unexpected header len for TX: "
 378                                "%zd expected %zd\n",
 379                                iov_length(vq->hdr, s), hdr_size);
 380                         break;
 381                 }
 382                 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
 383                                        nvq->upend_idx != nvq->done_idx);
 384
 385                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 386                 if (zcopy_used) {
 387                         vq->heads[nvq->upend_idx].id = head;
 388                         if (!vhost_net_tx_select_zcopy(net) ||
 389                             len < VHOST_GOODCOPY_LEN) {
 390                                 /* copy don't need to wait for DMA done */
 391                                 vq->heads[nvq->upend_idx].len =
 392                                                         VHOST_DMA_DONE_LEN;
 393                                 msg.msg_control = NULL;
 394                                 msg.msg_controllen = 0;
 395                                 ubufs = NULL;
 396                         } else {
 397                                 struct ubuf_info *ubuf;
 398                                 ubuf = nvq->ubuf_info + nvq->upend_idx;
 399
 400                                 vq->heads[nvq->upend_idx].len =
 401                                         VHOST_DMA_IN_PROGRESS;
 402                                 ubuf->callback = vhost_zerocopy_callback;
 403                                 ubuf->ctx = nvq->ubufs;
 404                                 ubuf->desc = nvq->upend_idx;
 405                                 msg.msg_control = ubuf;
 406                                 msg.msg_controllen = sizeof(ubuf);
 407                                 ubufs = nvq->ubufs;
 408                                 kref_get(&ubufs->kref);
 409                         }
 410                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 411                 }
 412                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 413                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 414                 if (unlikely(err < 0)) {
 415                         if (zcopy_used) {
 416                                 if (ubufs)
 417                                         vhost_ubuf_put(ubufs);
 418                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 419                                         % UIO_MAXIOV;
 420                         }
 421                         vhost_discard_vq_desc(vq, 1);
 422                         break;
 423                 }
 424                 if (err != len)
 425                         pr_debug("Truncated TX packet: "
 426                                  " len %d != %zd\n", err, len);
 427                 if (!zcopy_used)
 428                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 429                 else
 430                         vhost_zerocopy_signal_used(net, vq);
 431                 total_len += len;
 432                 vhost_net_tx_packet(net);
 433                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 434                         vhost_poll_queue(&vq->poll);
 435                         break;
 436                 }
 437         }
 438
 439         mutex_unlock(&vq->mutex);
 440 }
 441
 442 static int peek_head_len(struct sock *sk)
 443 {
 444         struct sk_buff *head;
 445         int len = 0;
 446         unsigned long flags;
 447
 448         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 449         head = skb_peek(&sk->sk_receive_queue);
 450         if (likely(head)) {
 451                 len = head->len;
 452                 if (vlan_tx_tag_present(head))
 453                         len += VLAN_HLEN;
 454         }
 455
 456         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 457         return len;
 458 }
 459
 460 /* This is a multi-buffer version of vhost_get_desc, that works if
 461  *      vq has read descriptors only.
 462  * @vq          - the relevant virtqueue
 463  * @datalen     - data length we'll be reading
 464  * @iovcount    - returned count of io vectors we fill
 465  * @log         - vhost log
 466  * @log_num     - log offset
 467  * @quota       - headcount quota, 1 for big buffer
 468  *      returns number of buffer heads allocated, negative on error
 469  */
 470 static int get_rx_bufs(struct vhost_virtqueue *vq,
 471                        struct vring_used_elem *heads,
 472                        int datalen,
 473                        unsigned *iovcount,
 474                        struct vhost_log *log,
 475                        unsigned *log_num,
 476                        unsigned int quota)
 477 {
 478         unsigned int out, in;
 479         int seg = 0;
 480         int headcount = 0;
 481         unsigned d;
 482         int r, nlogs = 0;
 483
 484         while (datalen > 0 && headcount < quota) {
 485                 if (unlikely(seg >= UIO_MAXIOV)) {
 486                         r = -ENOBUFS;
 487                         goto err;
 488                 }
 489                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 490                                       ARRAY_SIZE(vq->iov) - seg, &out,
 491                                       &in, log, log_num);
 492                 if (d == vq->num) {
 493                         r = 0;
 494                         goto err;
 495                 }
 496                 if (unlikely(out || in <= 0)) {
 497                         vq_err(vq, "unexpected descriptor format for RX: "
 498                                 "out %d, in %d\n", out, in);
 499                         r = -EINVAL;
 500                         goto err;
 501                 }
 502                 if (unlikely(log)) {
 503                         nlogs += *log_num;
 504                         log += *log_num;
 505                 }
 506                 heads[headcount].id = d;
 507                 heads[headcount].len = iov_length(vq->iov + seg, in);
 508                 datalen -= heads[headcount].len;
 509                 ++headcount;
 510                 seg += in;
 511         }
 512         heads[headcount - 1].len += datalen;
 513         *iovcount = seg;
 514         if (unlikely(log))
 515                 *log_num = nlogs;
 516         return headcount;
 517 err:
 518         vhost_discard_vq_desc(vq, headcount);
 519         return r;
 520 }
 521
 522 /* Expects to be always run from workqueue - which acts as
 523  * read-size critical section for our kind of RCU. */
 524 static void handle_rx(struct vhost_net *net)
 525 {
 526         struct vhost_virtqueue *vq = &net->vqs[VHOST_NET_VQ_RX].vq;
 527         unsigned uninitialized_var(in), log;
 528         struct vhost_log *vq_log;
 529         struct msghdr msg = {
 530                 .msg_name = NULL,
 531                 .msg_namelen = 0,
 532                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 533                 .msg_controllen = 0,
 534                 .msg_iov = vq->iov,
 535                 .msg_flags = MSG_DONTWAIT,
 536         };
 537         struct virtio_net_hdr_mrg_rxbuf hdr = {
 538                 .hdr.flags = 0,
 539                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 540         };
 541         size_t total_len = 0;
 542         int err, mergeable;
 543         s16 headcount;
 544         size_t vhost_hlen, sock_hlen;
 545         size_t vhost_len, sock_len;
 546         /* TODO: check that we are running from vhost_worker? */
 547         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 548
 549         if (!sock)
 550                 return;
 551
 552         mutex_lock(&vq->mutex);
 553         vhost_disable_notify(&net->dev, vq);
 554         vhost_hlen = vq->vhost_hlen;
 555         sock_hlen = vq->sock_hlen;
 556
 557         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 558                 vq->log : NULL;
 559         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 560
 561         while ((sock_len = peek_head_len(sock->sk))) {
 562                 sock_len += sock_hlen;
 563                 vhost_len = sock_len + vhost_hlen;
 564                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 565                                         &in, vq_log, &log,
 566                                         likely(mergeable) ? UIO_MAXIOV : 1);
 567                 /* On error, stop handling until the next kick. */
 568                 if (unlikely(headcount < 0))
 569                         break;
 570                 /* OK, now we need to know about added descriptors. */
 571                 if (!headcount) {
 572                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 573                                 /* They have slipped one in as we were
 574                                  * doing that: check again. */
 575                                 vhost_disable_notify(&net->dev, vq);
 576                                 continue;
 577                         }
 578                         /* Nothing new?  Wait for eventfd to tell us
 579                          * they refilled. */
 580                         break;
 581                 }
 582                 /* We don't need to be notified again. */
 583                 if (unlikely((vhost_hlen)))
 584                         /* Skip header. TODO: support TSO. */
 585                         move_iovec_hdr(vq->iov, vq->hdr, vhost_hlen, in);
 586                 else
 587                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 588                          * needed because recvmsg can modify msg_iov. */
 589                         copy_iovec_hdr(vq->iov, vq->hdr, sock_hlen, in);
 590                 msg.msg_iovlen = in;
 591                 err = sock->ops->recvmsg(NULL, sock, &msg,
 592                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 593                 /* Userspace might have consumed the packet meanwhile:
 594                  * it's not supposed to do this usually, but might be hard
 595                  * to prevent. Discard data we got (if any) and keep going. */
 596                 if (unlikely(err != sock_len)) {
 597                         pr_debug("Discarded rx packet: "
 598                                  " len %d, expected %zd\n", err, sock_len);
 599                         vhost_discard_vq_desc(vq, headcount);
 600                         continue;
 601                 }
 602                 if (unlikely(vhost_hlen) &&
 603                     memcpy_toiovecend(vq->hdr, (unsigned char *)&hdr, 0,
 604                                       vhost_hlen)) {
 605                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 606                                vq->iov->iov_base);
 607                         break;
 608                 }
 609                 /* TODO: Should check and handle checksum. */
 610                 if (likely(mergeable) &&
 611                     memcpy_toiovecend(vq->hdr, (unsigned char *)&headcount,
 612                                       offsetof(typeof(hdr), num_buffers),
 613                                       sizeof hdr.num_buffers)) {
 614                         vq_err(vq, "Failed num_buffers write");
 615                         vhost_discard_vq_desc(vq, headcount);
 616                         break;
 617                 }
 618                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 619                                             headcount);
 620                 if (unlikely(vq_log))
 621                         vhost_log_write(vq, vq_log, log, vhost_len);
 622                 total_len += vhost_len;
 623                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 624                         vhost_poll_queue(&vq->poll);
 625                         break;
 626                 }
 627         }
 628
 629         mutex_unlock(&vq->mutex);
 630 }
 631
 632 static void handle_tx_kick(struct vhost_work *work)
 633 {
 634         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 635                                                   poll.work);
 636         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 637
 638         handle_tx(net);
 639 }
 640
 641 static void handle_rx_kick(struct vhost_work *work)
 642 {
 643         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 644                                                   poll.work);
 645         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 646
 647         handle_rx(net);
 648 }
 649
 650 static void handle_tx_net(struct vhost_work *work)
 651 {
 652         struct vhost_net *net = container_of(work, struct vhost_net,
 653                                              poll[VHOST_NET_VQ_TX].work);
 654         handle_tx(net);
 655 }
 656
 657 static void handle_rx_net(struct vhost_work *work)
 658 {
 659         struct vhost_net *net = container_of(work, struct vhost_net,
 660                                              poll[VHOST_NET_VQ_RX].work);
 661         handle_rx(net);
 662 }
 663
 664 static int vhost_net_open(struct inode *inode, struct file *f)
 665 {
 666         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 667         struct vhost_dev *dev;
 668         struct vhost_virtqueue **vqs;
 669         int r, i;
 670
 671         if (!n)
 672                 return -ENOMEM;
 673         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 674         if (!vqs) {
 675                 kfree(n);
 676                 return -ENOMEM;
 677         }
 678
 679         dev = &n->dev;
 680         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 681         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 682         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 683         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 684         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 685                 n->vqs[i].ubufs = NULL;
 686                 n->vqs[i].ubuf_info = NULL;
 687                 n->vqs[i].upend_idx = 0;
 688                 n->vqs[i].done_idx = 0;
 689         }
 690         r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 691         if (r < 0) {
 692                 kfree(n);
 693                 kfree(vqs);
 694                 return r;
 695         }
 696
 697         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 698         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 699
 700         f->private_data = n;
 701
 702         return 0;
 703 }
 704
 705 static void vhost_net_disable_vq(struct vhost_net *n,
 706                                  struct vhost_virtqueue *vq)
 707 {
 708         struct vhost_net_virtqueue *nvq =
 709                 container_of(vq, struct vhost_net_virtqueue, vq);
 710         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 711         if (!vq->private_data)
 712                 return;
 713         vhost_poll_stop(poll);
 714 }
 715
 716 static int vhost_net_enable_vq(struct vhost_net *n,
 717                                 struct vhost_virtqueue *vq)
 718 {
 719         struct vhost_net_virtqueue *nvq =
 720                 container_of(vq, struct vhost_net_virtqueue, vq);
 721         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 722         struct socket *sock;
 723
 724         sock = rcu_dereference_protected(vq->private_data,
 725                                          lockdep_is_held(&vq->mutex));
 726         if (!sock)
 727                 return 0;
 728
 729         return vhost_poll_start(poll, sock->file);
 730 }
 731
 732 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 733                                         struct vhost_virtqueue *vq)
 734 {
 735         struct socket *sock;
 736
 737         mutex_lock(&vq->mutex);
 738         sock = rcu_dereference_protected(vq->private_data,
 739                                          lockdep_is_held(&vq->mutex));
 740         vhost_net_disable_vq(n, vq);
 741         rcu_assign_pointer(vq->private_data, NULL);
 742         mutex_unlock(&vq->mutex);
 743         return sock;
 744 }
 745
 746 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 747                            struct socket **rx_sock)
 748 {
 749         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 750         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 751 }
 752
 753 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 754 {
 755         vhost_poll_flush(n->poll + index);
 756         vhost_poll_flush(&n->vqs[index].vq.poll);
 757 }
 758
 759 static void vhost_net_flush(struct vhost_net *n)
 760 {
 761         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 762         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 763         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 764                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 765                 n->tx_flush = true;
 766                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 767                 /* Wait for all lower device DMAs done. */
 768                 vhost_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 769                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 770                 n->tx_flush = false;
 771                 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
 772                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 773         }
 774 }
 775
 776 static int vhost_net_release(struct inode *inode, struct file *f)
 777 {
 778         struct vhost_net *n = f->private_data;
 779         struct socket *tx_sock;
 780         struct socket *rx_sock;
 781
 782         vhost_net_stop(n, &tx_sock, &rx_sock);
 783         vhost_net_flush(n);
 784         vhost_dev_stop(&n->dev);
 785         vhost_dev_cleanup(&n->dev, false);
 786         vhost_net_reset_ubuf_info(n);
 787         if (tx_sock)
 788                 fput(tx_sock->file);
 789         if (rx_sock)
 790                 fput(rx_sock->file);
 791         /* We do an extra flush before freeing memory,
 792          * since jobs can re-queue themselves. */
 793         vhost_net_flush(n);
 794         kfree(n->dev.vqs);
 795         kfree(n);
 796         return 0;
 797 }
 798
 799 static struct socket *get_raw_socket(int fd)
 800 {
 801         struct {
 802                 struct sockaddr_ll sa;
 803                 char  buf[MAX_ADDR_LEN];
 804         } uaddr;
 805         int uaddr_len = sizeof uaddr, r;
 806         struct socket *sock = sockfd_lookup(fd, &r);
 807
 808         if (!sock)
 809                 return ERR_PTR(-ENOTSOCK);
 810
 811         /* Parameter checking */
 812         if (sock->sk->sk_type != SOCK_RAW) {
 813                 r = -ESOCKTNOSUPPORT;
 814                 goto err;
 815         }
 816
 817         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 818                                &uaddr_len, 0);
 819         if (r)
 820                 goto err;
 821
 822         if (uaddr.sa.sll_family != AF_PACKET) {
 823                 r = -EPFNOSUPPORT;
 824                 goto err;
 825         }
 826         return sock;
 827 err:
 828         fput(sock->file);
 829         return ERR_PTR(r);
 830 }
 831
 832 static struct socket *get_tap_socket(int fd)
 833 {
 834         struct file *file = fget(fd);
 835         struct socket *sock;
 836
 837         if (!file)
 838                 return ERR_PTR(-EBADF);
 839         sock = tun_get_socket(file);
 840         if (!IS_ERR(sock))
 841                 return sock;
 842         sock = macvtap_get_socket(file);
 843         if (IS_ERR(sock))
 844                 fput(file);
 845         return sock;
 846 }
 847
 848 static struct socket *get_socket(int fd)
 849 {
 850         struct socket *sock;
 851
 852         /* special case to disable backend */
 853         if (fd == -1)
 854                 return NULL;
 855         sock = get_raw_socket(fd);
 856         if (!IS_ERR(sock))
 857                 return sock;
 858         sock = get_tap_socket(fd);
 859         if (!IS_ERR(sock))
 860                 return sock;
 861         return ERR_PTR(-ENOTSOCK);
 862 }
 863
 864 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 865 {
 866         struct socket *sock, *oldsock;
 867         struct vhost_virtqueue *vq;
 868         struct vhost_net_virtqueue *nvq;
 869         struct vhost_ubuf_ref *ubufs, *oldubufs = NULL;
 870         int r;
 871
 872         mutex_lock(&n->dev.mutex);
 873         r = vhost_dev_check_owner(&n->dev);
 874         if (r)
 875                 goto err;
 876
 877         if (index >= VHOST_NET_VQ_MAX) {
 878                 r = -ENOBUFS;
 879                 goto err;
 880         }
 881         vq = &n->vqs[index].vq;
 882         nvq = &n->vqs[index];
 883         mutex_lock(&vq->mutex);
 884
 885         /* Verify that ring has been setup correctly. */
 886         if (!vhost_vq_access_ok(vq)) {
 887                 r = -EFAULT;
 888                 goto err_vq;
 889         }
 890         sock = get_socket(fd);
 891         if (IS_ERR(sock)) {
 892                 r = PTR_ERR(sock);
 893                 goto err_vq;
 894         }
 895
 896         /* start polling new socket */
 897         oldsock = rcu_dereference_protected(vq->private_data,
 898                                             lockdep_is_held(&vq->mutex));
 899         if (sock != oldsock) {
 900                 ubufs = vhost_ubuf_alloc(vq, sock && vhost_sock_zcopy(sock));
 901                 if (IS_ERR(ubufs)) {
 902                         r = PTR_ERR(ubufs);
 903                         goto err_ubufs;
 904                 }
 905
 906                 vhost_net_disable_vq(n, vq);
 907                 rcu_assign_pointer(vq->private_data, sock);
 908                 r = vhost_init_used(vq);
 909                 if (r)
 910                         goto err_used;
 911                 r = vhost_net_enable_vq(n, vq);
 912                 if (r)
 913                         goto err_used;
 914
 915                 oldubufs = nvq->ubufs;
 916                 nvq->ubufs = ubufs;
 917
 918                 n->tx_packets = 0;
 919                 n->tx_zcopy_err = 0;
 920                 n->tx_flush = false;
 921         }
 922
 923         mutex_unlock(&vq->mutex);
 924
 925         if (oldubufs) {
 926                 vhost_ubuf_put_and_wait(oldubufs);
 927                 mutex_lock(&vq->mutex);
 928                 vhost_zerocopy_signal_used(n, vq);
 929                 mutex_unlock(&vq->mutex);
 930         }
 931
 932         if (oldsock) {
 933                 vhost_net_flush_vq(n, index);
 934                 fput(oldsock->file);
 935         }
 936
 937         mutex_unlock(&n->dev.mutex);
 938         return 0;
 939
 940 err_used:
 941         rcu_assign_pointer(vq->private_data, oldsock);
 942         vhost_net_enable_vq(n, vq);
 943         if (ubufs)
 944                 vhost_ubuf_put_and_wait(ubufs);
 945 err_ubufs:
 946         fput(sock->file);
 947 err_vq:
 948         mutex_unlock(&vq->mutex);
 949 err:
 950         mutex_unlock(&n->dev.mutex);
 951         return r;
 952 }
 953
 954 static long vhost_net_reset_owner(struct vhost_net *n)
 955 {
 956         struct socket *tx_sock = NULL;
 957         struct socket *rx_sock = NULL;
 958         long err;
 959
 960         mutex_lock(&n->dev.mutex);
 961         err = vhost_dev_check_owner(&n->dev);
 962         if (err)
 963                 goto done;
 964         vhost_net_stop(n, &tx_sock, &rx_sock);
 965         vhost_net_flush(n);
 966         err = vhost_dev_reset_owner(&n->dev);
 967         vhost_net_reset_ubuf_info(n);
 968 done:
 969         mutex_unlock(&n->dev.mutex);
 970         if (tx_sock)
 971                 fput(tx_sock->file);
 972         if (rx_sock)
 973                 fput(rx_sock->file);
 974         return err;
 975 }
 976
 977 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 978 {
 979         size_t vhost_hlen, sock_hlen, hdr_len;
 980         int i;
 981
 982         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
 983                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
 984                         sizeof(struct virtio_net_hdr);
 985         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
 986                 /* vhost provides vnet_hdr */
 987                 vhost_hlen = hdr_len;
 988                 sock_hlen = 0;
 989         } else {
 990                 /* socket provides vnet_hdr */
 991                 vhost_hlen = 0;
 992                 sock_hlen = hdr_len;
 993         }
 994         mutex_lock(&n->dev.mutex);
 995         if ((features & (1 << VHOST_F_LOG_ALL)) &&
 996             !vhost_log_access_ok(&n->dev)) {
 997                 mutex_unlock(&n->dev.mutex);
 998                 return -EFAULT;
 999         }
1000         n->dev.acked_features = features;
1001         smp_wmb();
1002         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1003                 mutex_lock(&n->vqs[i].vq.mutex);
1004                 n->vqs[i].vq.vhost_hlen = vhost_hlen;
1005                 n->vqs[i].vq.sock_hlen = sock_hlen;
1006                 mutex_unlock(&n->vqs[i].vq.mutex);
1007         }
1008         vhost_net_flush(n);
1009         mutex_unlock(&n->dev.mutex);
1010         return 0;
1011 }
1012
1013 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1014                             unsigned long arg)
1015 {
1016         struct vhost_net *n = f->private_data;
1017         void __user *argp = (void __user *)arg;
1018         u64 __user *featurep = argp;
1019         struct vhost_vring_file backend;
1020         u64 features;
1021         int r;
1022
1023         switch (ioctl) {
1024         case VHOST_NET_SET_BACKEND:
1025                 if (copy_from_user(&backend, argp, sizeof backend))
1026                         return -EFAULT;
1027                 return vhost_net_set_backend(n, backend.index, backend.fd);
1028         case VHOST_GET_FEATURES:
1029                 features = VHOST_NET_FEATURES;
1030                 if (copy_to_user(featurep, &features, sizeof features))
1031                         return -EFAULT;
1032                 return 0;
1033         case VHOST_SET_FEATURES:
1034                 if (copy_from_user(&features, featurep, sizeof features))
1035                         return -EFAULT;
1036                 if (features & ~VHOST_NET_FEATURES)
1037                         return -EOPNOTSUPP;
1038                 return vhost_net_set_features(n, features);
1039         case VHOST_RESET_OWNER:
1040                 return vhost_net_reset_owner(n);
1041         default:
1042                 mutex_lock(&n->dev.mutex);
1043                 if (ioctl == VHOST_SET_OWNER) {
1044                         r = vhost_net_set_ubuf_info(n);
1045                         if (r)
1046                                 goto out;
1047                 }
1048                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1049                 if (r == -ENOIOCTLCMD)
1050                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1051                 else
1052                         vhost_net_flush(n);
1053 out:
1054                 mutex_unlock(&n->dev.mutex);
1055                 return r;
1056         }
1057 }
1058
1059 #ifdef CONFIG_COMPAT
1060 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1061                                    unsigned long arg)
1062 {
1063         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1064 }
1065 #endif
1066
1067 static const struct file_operations vhost_net_fops = {
1068         .owner          = THIS_MODULE,
1069         .release        = vhost_net_release,
1070         .unlocked_ioctl = vhost_net_ioctl,
1071 #ifdef CONFIG_COMPAT
1072         .compat_ioctl   = vhost_net_compat_ioctl,
1073 #endif
1074         .open           = vhost_net_open,
1075         .llseek         = noop_llseek,
1076 };
1077
1078 static struct miscdevice vhost_net_misc = {
1079         .minor = VHOST_NET_MINOR,
1080         .name = "vhost-net",
1081         .fops = &vhost_net_fops,
1082 };
1083
1084 static int vhost_net_init(void)
1085 {
1086         if (experimental_zcopytx)
1087                 vhost_enable_zcopy(VHOST_NET_VQ_TX);
1088         return misc_register(&vhost_net_misc);
1089 }
1090 module_init(vhost_net_init);
1091
1092 static void vhost_net_exit(void)
1093 {
1094         misc_deregister(&vhost_net_misc);
1095 }
1096 module_exit(vhost_net_exit);
1097
1098 MODULE_VERSION("0.0.1");
1099 MODULE_LICENSE("GPL v2");
1100 MODULE_AUTHOR("Michael S. Tsirkin");
1101 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1102 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1103 MODULE_ALIAS("devname:vhost-net");