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_FEATURES = VHOST_FEATURES |
  63                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  64                          (1ULL << VIRTIO_NET_F_MRG_RXBUF),
  65 };
  66
  67 enum {
  68         VHOST_NET_VQ_RX = 0,
  69         VHOST_NET_VQ_TX = 1,
  70         VHOST_NET_VQ_MAX = 2,
  71 };
  72
  73 struct vhost_net_ubuf_ref {
  74         struct kref kref;
  75         wait_queue_head_t wait;
  76         struct vhost_virtqueue *vq;
  77 };
  78
  79 struct vhost_net_virtqueue {
  80         struct vhost_virtqueue vq;
  81         /* hdr is used to store the virtio header.
  82          * Since each iovec has >= 1 byte length, we never need more than
  83          * header length entries to store the header. */
  84         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
  85         size_t vhost_hlen;
  86         size_t sock_hlen;
  87         /* vhost zerocopy support fields below: */
  88         /* last used idx for outstanding DMA zerocopy buffers */
  89         int upend_idx;
  90         /* first used idx for DMA done zerocopy buffers */
  91         int done_idx;
  92         /* an array of userspace buffers info */
  93         struct ubuf_info *ubuf_info;
  94         /* Reference counting for outstanding ubufs.
  95          * Protected by vq mutex. Writers must also take device mutex. */
  96         struct vhost_net_ubuf_ref *ubufs;
  97 };
  98
  99 struct vhost_net {
 100         struct vhost_dev dev;
 101         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 102         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 103         /* Number of TX recently submitted.
 104          * Protected by tx vq lock. */
 105         unsigned tx_packets;
 106         /* Number of times zerocopy TX recently failed.
 107          * Protected by tx vq lock. */
 108         unsigned tx_zcopy_err;
 109         /* Flush in progress. Protected by tx vq lock. */
 110         bool tx_flush;
 111 };
 112
 113 static unsigned vhost_net_zcopy_mask __read_mostly;
 114
 115 static void vhost_net_enable_zcopy(int vq)
 116 {
 117         vhost_net_zcopy_mask |= 0x1 << vq;
 118 }
 119
 120 static void vhost_net_zerocopy_done_signal(struct kref *kref)
 121 {
 122         struct vhost_net_ubuf_ref *ubufs;
 123
 124         ubufs = container_of(kref, struct vhost_net_ubuf_ref, kref);
 125         wake_up(&ubufs->wait);
 126 }
 127
 128 static struct vhost_net_ubuf_ref *
 129 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 130 {
 131         struct vhost_net_ubuf_ref *ubufs;
 132         /* No zero copy backend? Nothing to count. */
 133         if (!zcopy)
 134                 return NULL;
 135         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 136         if (!ubufs)
 137                 return ERR_PTR(-ENOMEM);
 138         kref_init(&ubufs->kref);
 139         init_waitqueue_head(&ubufs->wait);
 140         ubufs->vq = vq;
 141         return ubufs;
 142 }
 143
 144 static void vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 145 {
 146         kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
 147 }
 148
 149 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 150 {
 151         kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
 152         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
 153         kfree(ubufs);
 154 }
 155
 156 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 157 {
 158         int i;
 159
 160         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 161                 kfree(n->vqs[i].ubuf_info);
 162                 n->vqs[i].ubuf_info = NULL;
 163         }
 164 }
 165
 166 int vhost_net_set_ubuf_info(struct vhost_net *n)
 167 {
 168         bool zcopy;
 169         int i;
 170
 171         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 172                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 173                 if (!zcopy)
 174                         continue;
 175                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 176                                               UIO_MAXIOV, GFP_KERNEL);
 177                 if  (!n->vqs[i].ubuf_info)
 178                         goto err;
 179         }
 180         return 0;
 181
 182 err:
 183         vhost_net_clear_ubuf_info(n);
 184         return -ENOMEM;
 185 }
 186
 187 void vhost_net_vq_reset(struct vhost_net *n)
 188 {
 189         int i;
 190
 191         vhost_net_clear_ubuf_info(n);
 192
 193         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 194                 n->vqs[i].done_idx = 0;
 195                 n->vqs[i].upend_idx = 0;
 196                 n->vqs[i].ubufs = NULL;
 197                 n->vqs[i].vhost_hlen = 0;
 198                 n->vqs[i].sock_hlen = 0;
 199         }
 200
 201 }
 202
 203 static void vhost_net_tx_packet(struct vhost_net *net)
 204 {
 205         ++net->tx_packets;
 206         if (net->tx_packets < 1024)
 207                 return;
 208         net->tx_packets = 0;
 209         net->tx_zcopy_err = 0;
 210 }
 211
 212 static void vhost_net_tx_err(struct vhost_net *net)
 213 {
 214         ++net->tx_zcopy_err;
 215 }
 216
 217 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 218 {
 219         /* TX flush waits for outstanding DMAs to be done.
 220          * Don't start new DMAs.
 221          */
 222         return !net->tx_flush &&
 223                 net->tx_packets / 64 >= net->tx_zcopy_err;
 224 }
 225
 226 static bool vhost_sock_zcopy(struct socket *sock)
 227 {
 228         return unlikely(experimental_zcopytx) &&
 229                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 230 }
 231
 232 /* Pop first len bytes from iovec. Return number of segments used. */
 233 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 234                           size_t len, int iov_count)
 235 {
 236         int seg = 0;
 237         size_t size;
 238
 239         while (len && seg < iov_count) {
 240                 size = min(from->iov_len, len);
 241                 to->iov_base = from->iov_base;
 242                 to->iov_len = size;
 243                 from->iov_len -= size;
 244                 from->iov_base += size;
 245                 len -= size;
 246                 ++from;
 247                 ++to;
 248                 ++seg;
 249         }
 250         return seg;
 251 }
 252 /* Copy iovec entries for len bytes from iovec. */
 253 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 254                            size_t len, int iovcount)
 255 {
 256         int seg = 0;
 257         size_t size;
 258
 259         while (len && seg < iovcount) {
 260                 size = min(from->iov_len, len);
 261                 to->iov_base = from->iov_base;
 262                 to->iov_len = size;
 263                 len -= size;
 264                 ++from;
 265                 ++to;
 266                 ++seg;
 267         }
 268 }
 269
 270 /* In case of DMA done not in order in lower device driver for some reason.
 271  * upend_idx is used to track end of used idx, done_idx is used to track head
 272  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 273  * guest used idx.
 274  */
 275 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 276                                       struct vhost_virtqueue *vq)
 277 {
 278         struct vhost_net_virtqueue *nvq =
 279                 container_of(vq, struct vhost_net_virtqueue, vq);
 280         int i;
 281         int j = 0;
 282
 283         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 284                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 285                         vhost_net_tx_err(net);
 286                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 287                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 288                         vhost_add_used_and_signal(vq->dev, vq,
 289                                                   vq->heads[i].id, 0);
 290                         ++j;
 291                 } else
 292                         break;
 293         }
 294         if (j)
 295                 nvq->done_idx = i;
 296         return j;
 297 }
 298
 299 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 300 {
 301         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 302         struct vhost_virtqueue *vq = ubufs->vq;
 303         int cnt = atomic_read(&ubufs->kref.refcount);
 304
 305         /*
 306          * Trigger polling thread if guest stopped submitting new buffers:
 307          * in this case, the refcount after decrement will eventually reach 1
 308          * so here it is 2.
 309          * We also trigger polling periodically after each 16 packets
 310          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 311          * less than 10% of times).
 312          */
 313         if (cnt <= 2 || !(cnt % 16))
 314                 vhost_poll_queue(&vq->poll);
 315         /* set len to mark this desc buffers done DMA */
 316         vq->heads[ubuf->desc].len = success ?
 317                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 318         vhost_net_ubuf_put(ubufs);
 319 }
 320
 321 /* Expects to be always run from workqueue - which acts as
 322  * read-size critical section for our kind of RCU. */
 323 static void handle_tx(struct vhost_net *net)
 324 {
 325         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 326         struct vhost_virtqueue *vq = &nvq->vq;
 327         unsigned out, in, s;
 328         int head;
 329         struct msghdr msg = {
 330                 .msg_name = NULL,
 331                 .msg_namelen = 0,
 332                 .msg_control = NULL,
 333                 .msg_controllen = 0,
 334                 .msg_iov = vq->iov,
 335                 .msg_flags = MSG_DONTWAIT,
 336         };
 337         size_t len, total_len = 0;
 338         int err;
 339         size_t hdr_size;
 340         struct socket *sock;
 341         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 342         bool zcopy, zcopy_used;
 343
 344         /* TODO: check that we are running from vhost_worker? */
 345         sock = rcu_dereference_check(vq->private_data, 1);
 346         if (!sock)
 347                 return;
 348
 349         mutex_lock(&vq->mutex);
 350         vhost_disable_notify(&net->dev, vq);
 351
 352         hdr_size = nvq->vhost_hlen;
 353         zcopy = nvq->ubufs;
 354
 355         for (;;) {
 356                 /* Release DMAs done buffers first */
 357                 if (zcopy)
 358                         vhost_zerocopy_signal_used(net, vq);
 359
 360                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 361                                          ARRAY_SIZE(vq->iov),
 362                                          &out, &in,
 363                                          NULL, NULL);
 364                 /* On error, stop handling until the next kick. */
 365                 if (unlikely(head < 0))
 366                         break;
 367                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 368                 if (head == vq->num) {
 369                         int num_pends;
 370
 371                         /* If more outstanding DMAs, queue the work.
 372                          * Handle upend_idx wrap around
 373                          */
 374                         num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
 375                                     (nvq->upend_idx - nvq->done_idx) :
 376                                     (nvq->upend_idx + UIO_MAXIOV -
 377                                      nvq->done_idx);
 378                         if (unlikely(num_pends > VHOST_MAX_PEND))
 379                                 break;
 380                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 381                                 vhost_disable_notify(&net->dev, vq);
 382                                 continue;
 383                         }
 384                         break;
 385                 }
 386                 if (in) {
 387                         vq_err(vq, "Unexpected descriptor format for TX: "
 388                                "out %d, int %d\n", out, in);
 389                         break;
 390                 }
 391                 /* Skip header. TODO: support TSO. */
 392                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 393                 msg.msg_iovlen = out;
 394                 len = iov_length(vq->iov, out);
 395                 /* Sanity check */
 396                 if (!len) {
 397                         vq_err(vq, "Unexpected header len for TX: "
 398                                "%zd expected %zd\n",
 399                                iov_length(nvq->hdr, s), hdr_size);
 400                         break;
 401                 }
 402                 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
 403                                        nvq->upend_idx != nvq->done_idx);
 404
 405                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 406                 if (zcopy_used) {
 407                         vq->heads[nvq->upend_idx].id = head;
 408                         if (!vhost_net_tx_select_zcopy(net) ||
 409                             len < VHOST_GOODCOPY_LEN) {
 410                                 /* copy don't need to wait for DMA done */
 411                                 vq->heads[nvq->upend_idx].len =
 412                                                         VHOST_DMA_DONE_LEN;
 413                                 msg.msg_control = NULL;
 414                                 msg.msg_controllen = 0;
 415                                 ubufs = NULL;
 416                         } else {
 417                                 struct ubuf_info *ubuf;
 418                                 ubuf = nvq->ubuf_info + nvq->upend_idx;
 419
 420                                 vq->heads[nvq->upend_idx].len =
 421                                         VHOST_DMA_IN_PROGRESS;
 422                                 ubuf->callback = vhost_zerocopy_callback;
 423                                 ubuf->ctx = nvq->ubufs;
 424                                 ubuf->desc = nvq->upend_idx;
 425                                 msg.msg_control = ubuf;
 426                                 msg.msg_controllen = sizeof(ubuf);
 427                                 ubufs = nvq->ubufs;
 428                                 kref_get(&ubufs->kref);
 429                         }
 430                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 431                 } else
 432                         msg.msg_control = NULL;
 433                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 434                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 435                 if (unlikely(err < 0)) {
 436                         if (zcopy_used) {
 437                                 if (ubufs)
 438                                         vhost_net_ubuf_put(ubufs);
 439                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 440                                         % UIO_MAXIOV;
 441                         }
 442                         vhost_discard_vq_desc(vq, 1);
 443                         break;
 444                 }
 445                 if (err != len)
 446                         pr_debug("Truncated TX packet: "
 447                                  " len %d != %zd\n", err, len);
 448                 if (!zcopy_used)
 449                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 450                 else
 451                         vhost_zerocopy_signal_used(net, vq);
 452                 total_len += len;
 453                 vhost_net_tx_packet(net);
 454                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 455                         vhost_poll_queue(&vq->poll);
 456                         break;
 457                 }
 458         }
 459
 460         mutex_unlock(&vq->mutex);
 461 }
 462
 463 static int peek_head_len(struct sock *sk)
 464 {
 465         struct sk_buff *head;
 466         int len = 0;
 467         unsigned long flags;
 468
 469         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 470         head = skb_peek(&sk->sk_receive_queue);
 471         if (likely(head)) {
 472                 len = head->len;
 473                 if (vlan_tx_tag_present(head))
 474                         len += VLAN_HLEN;
 475         }
 476
 477         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 478         return len;
 479 }
 480
 481 /* This is a multi-buffer version of vhost_get_desc, that works if
 482  *      vq has read descriptors only.
 483  * @vq          - the relevant virtqueue
 484  * @datalen     - data length we'll be reading
 485  * @iovcount    - returned count of io vectors we fill
 486  * @log         - vhost log
 487  * @log_num     - log offset
 488  * @quota       - headcount quota, 1 for big buffer
 489  *      returns number of buffer heads allocated, negative on error
 490  */
 491 static int get_rx_bufs(struct vhost_virtqueue *vq,
 492                        struct vring_used_elem *heads,
 493                        int datalen,
 494                        unsigned *iovcount,
 495                        struct vhost_log *log,
 496                        unsigned *log_num,
 497                        unsigned int quota)
 498 {
 499         unsigned int out, in;
 500         int seg = 0;
 501         int headcount = 0;
 502         unsigned d;
 503         int r, nlogs = 0;
 504
 505         while (datalen > 0 && headcount < quota) {
 506                 if (unlikely(seg >= UIO_MAXIOV)) {
 507                         r = -ENOBUFS;
 508                         goto err;
 509                 }
 510                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 511                                       ARRAY_SIZE(vq->iov) - seg, &out,
 512                                       &in, log, log_num);
 513                 if (d == vq->num) {
 514                         r = 0;
 515                         goto err;
 516                 }
 517                 if (unlikely(out || in <= 0)) {
 518                         vq_err(vq, "unexpected descriptor format for RX: "
 519                                 "out %d, in %d\n", out, in);
 520                         r = -EINVAL;
 521                         goto err;
 522                 }
 523                 if (unlikely(log)) {
 524                         nlogs += *log_num;
 525                         log += *log_num;
 526                 }
 527                 heads[headcount].id = d;
 528                 heads[headcount].len = iov_length(vq->iov + seg, in);
 529                 datalen -= heads[headcount].len;
 530                 ++headcount;
 531                 seg += in;
 532         }
 533         heads[headcount - 1].len += datalen;
 534         *iovcount = seg;
 535         if (unlikely(log))
 536                 *log_num = nlogs;
 537         return headcount;
 538 err:
 539         vhost_discard_vq_desc(vq, headcount);
 540         return r;
 541 }
 542
 543 /* Expects to be always run from workqueue - which acts as
 544  * read-size critical section for our kind of RCU. */
 545 static void handle_rx(struct vhost_net *net)
 546 {
 547         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 548         struct vhost_virtqueue *vq = &nvq->vq;
 549         unsigned uninitialized_var(in), log;
 550         struct vhost_log *vq_log;
 551         struct msghdr msg = {
 552                 .msg_name = NULL,
 553                 .msg_namelen = 0,
 554                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 555                 .msg_controllen = 0,
 556                 .msg_iov = vq->iov,
 557                 .msg_flags = MSG_DONTWAIT,
 558         };
 559         struct virtio_net_hdr_mrg_rxbuf hdr = {
 560                 .hdr.flags = 0,
 561                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 562         };
 563         size_t total_len = 0;
 564         int err, mergeable;
 565         s16 headcount;
 566         size_t vhost_hlen, sock_hlen;
 567         size_t vhost_len, sock_len;
 568         /* TODO: check that we are running from vhost_worker? */
 569         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 570
 571         if (!sock)
 572                 return;
 573
 574         mutex_lock(&vq->mutex);
 575         vhost_disable_notify(&net->dev, vq);
 576         vhost_hlen = nvq->vhost_hlen;
 577         sock_hlen = nvq->sock_hlen;
 578
 579         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 580                 vq->log : NULL;
 581         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 582
 583         while ((sock_len = peek_head_len(sock->sk))) {
 584                 sock_len += sock_hlen;
 585                 vhost_len = sock_len + vhost_hlen;
 586                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 587                                         &in, vq_log, &log,
 588                                         likely(mergeable) ? UIO_MAXIOV : 1);
 589                 /* On error, stop handling until the next kick. */
 590                 if (unlikely(headcount < 0))
 591                         break;
 592                 /* OK, now we need to know about added descriptors. */
 593                 if (!headcount) {
 594                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 595                                 /* They have slipped one in as we were
 596                                  * doing that: check again. */
 597                                 vhost_disable_notify(&net->dev, vq);
 598                                 continue;
 599                         }
 600                         /* Nothing new?  Wait for eventfd to tell us
 601                          * they refilled. */
 602                         break;
 603                 }
 604                 /* We don't need to be notified again. */
 605                 if (unlikely((vhost_hlen)))
 606                         /* Skip header. TODO: support TSO. */
 607                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 608                 else
 609                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 610                          * needed because recvmsg can modify msg_iov. */
 611                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 612                 msg.msg_iovlen = in;
 613                 err = sock->ops->recvmsg(NULL, sock, &msg,
 614                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 615                 /* Userspace might have consumed the packet meanwhile:
 616                  * it's not supposed to do this usually, but might be hard
 617                  * to prevent. Discard data we got (if any) and keep going. */
 618                 if (unlikely(err != sock_len)) {
 619                         pr_debug("Discarded rx packet: "
 620                                  " len %d, expected %zd\n", err, sock_len);
 621                         vhost_discard_vq_desc(vq, headcount);
 622                         continue;
 623                 }
 624                 if (unlikely(vhost_hlen) &&
 625                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 626                                       vhost_hlen)) {
 627                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 628                                vq->iov->iov_base);
 629                         break;
 630                 }
 631                 /* TODO: Should check and handle checksum. */
 632                 if (likely(mergeable) &&
 633                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 634                                       offsetof(typeof(hdr), num_buffers),
 635                                       sizeof hdr.num_buffers)) {
 636                         vq_err(vq, "Failed num_buffers write");
 637                         vhost_discard_vq_desc(vq, headcount);
 638                         break;
 639                 }
 640                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 641                                             headcount);
 642                 if (unlikely(vq_log))
 643                         vhost_log_write(vq, vq_log, log, vhost_len);
 644                 total_len += vhost_len;
 645                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 646                         vhost_poll_queue(&vq->poll);
 647                         break;
 648                 }
 649         }
 650
 651         mutex_unlock(&vq->mutex);
 652 }
 653
 654 static void handle_tx_kick(struct vhost_work *work)
 655 {
 656         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 657                                                   poll.work);
 658         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 659
 660         handle_tx(net);
 661 }
 662
 663 static void handle_rx_kick(struct vhost_work *work)
 664 {
 665         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 666                                                   poll.work);
 667         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 668
 669         handle_rx(net);
 670 }
 671
 672 static void handle_tx_net(struct vhost_work *work)
 673 {
 674         struct vhost_net *net = container_of(work, struct vhost_net,
 675                                              poll[VHOST_NET_VQ_TX].work);
 676         handle_tx(net);
 677 }
 678
 679 static void handle_rx_net(struct vhost_work *work)
 680 {
 681         struct vhost_net *net = container_of(work, struct vhost_net,
 682                                              poll[VHOST_NET_VQ_RX].work);
 683         handle_rx(net);
 684 }
 685
 686 static int vhost_net_open(struct inode *inode, struct file *f)
 687 {
 688         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 689         struct vhost_dev *dev;
 690         struct vhost_virtqueue **vqs;
 691         int r, i;
 692
 693         if (!n)
 694                 return -ENOMEM;
 695         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 696         if (!vqs) {
 697                 kfree(n);
 698                 return -ENOMEM;
 699         }
 700
 701         dev = &n->dev;
 702         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 703         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 704         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 705         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 706         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 707                 n->vqs[i].ubufs = NULL;
 708                 n->vqs[i].ubuf_info = NULL;
 709                 n->vqs[i].upend_idx = 0;
 710                 n->vqs[i].done_idx = 0;
 711                 n->vqs[i].vhost_hlen = 0;
 712                 n->vqs[i].sock_hlen = 0;
 713         }
 714         r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 715         if (r < 0) {
 716                 kfree(n);
 717                 kfree(vqs);
 718                 return r;
 719         }
 720
 721         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 722         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 723
 724         f->private_data = n;
 725
 726         return 0;
 727 }
 728
 729 static void vhost_net_disable_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         if (!vq->private_data)
 736                 return;
 737         vhost_poll_stop(poll);
 738 }
 739
 740 static int vhost_net_enable_vq(struct vhost_net *n,
 741                                 struct vhost_virtqueue *vq)
 742 {
 743         struct vhost_net_virtqueue *nvq =
 744                 container_of(vq, struct vhost_net_virtqueue, vq);
 745         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 746         struct socket *sock;
 747
 748         sock = rcu_dereference_protected(vq->private_data,
 749                                          lockdep_is_held(&vq->mutex));
 750         if (!sock)
 751                 return 0;
 752
 753         return vhost_poll_start(poll, sock->file);
 754 }
 755
 756 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 757                                         struct vhost_virtqueue *vq)
 758 {
 759         struct socket *sock;
 760
 761         mutex_lock(&vq->mutex);
 762         sock = rcu_dereference_protected(vq->private_data,
 763                                          lockdep_is_held(&vq->mutex));
 764         vhost_net_disable_vq(n, vq);
 765         rcu_assign_pointer(vq->private_data, NULL);
 766         mutex_unlock(&vq->mutex);
 767         return sock;
 768 }
 769
 770 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 771                            struct socket **rx_sock)
 772 {
 773         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 774         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 775 }
 776
 777 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 778 {
 779         vhost_poll_flush(n->poll + index);
 780         vhost_poll_flush(&n->vqs[index].vq.poll);
 781 }
 782
 783 static void vhost_net_flush(struct vhost_net *n)
 784 {
 785         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 786         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 787         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 788                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 789                 n->tx_flush = true;
 790                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 791                 /* Wait for all lower device DMAs done. */
 792                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 793                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 794                 n->tx_flush = false;
 795                 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
 796                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 797         }
 798 }
 799
 800 static int vhost_net_release(struct inode *inode, struct file *f)
 801 {
 802         struct vhost_net *n = f->private_data;
 803         struct socket *tx_sock;
 804         struct socket *rx_sock;
 805
 806         vhost_net_stop(n, &tx_sock, &rx_sock);
 807         vhost_net_flush(n);
 808         vhost_dev_stop(&n->dev);
 809         vhost_dev_cleanup(&n->dev, false);
 810         vhost_net_vq_reset(n);
 811         if (tx_sock)
 812                 fput(tx_sock->file);
 813         if (rx_sock)
 814                 fput(rx_sock->file);
 815         /* We do an extra flush before freeing memory,
 816          * since jobs can re-queue themselves. */
 817         vhost_net_flush(n);
 818         kfree(n->dev.vqs);
 819         kfree(n);
 820         return 0;
 821 }
 822
 823 static struct socket *get_raw_socket(int fd)
 824 {
 825         struct {
 826                 struct sockaddr_ll sa;
 827                 char  buf[MAX_ADDR_LEN];
 828         } uaddr;
 829         int uaddr_len = sizeof uaddr, r;
 830         struct socket *sock = sockfd_lookup(fd, &r);
 831
 832         if (!sock)
 833                 return ERR_PTR(-ENOTSOCK);
 834
 835         /* Parameter checking */
 836         if (sock->sk->sk_type != SOCK_RAW) {
 837                 r = -ESOCKTNOSUPPORT;
 838                 goto err;
 839         }
 840
 841         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 842                                &uaddr_len, 0);
 843         if (r)
 844                 goto err;
 845
 846         if (uaddr.sa.sll_family != AF_PACKET) {
 847                 r = -EPFNOSUPPORT;
 848                 goto err;
 849         }
 850         return sock;
 851 err:
 852         fput(sock->file);
 853         return ERR_PTR(r);
 854 }
 855
 856 static struct socket *get_tap_socket(int fd)
 857 {
 858         struct file *file = fget(fd);
 859         struct socket *sock;
 860
 861         if (!file)
 862                 return ERR_PTR(-EBADF);
 863         sock = tun_get_socket(file);
 864         if (!IS_ERR(sock))
 865                 return sock;
 866         sock = macvtap_get_socket(file);
 867         if (IS_ERR(sock))
 868                 fput(file);
 869         return sock;
 870 }
 871
 872 static struct socket *get_socket(int fd)
 873 {
 874         struct socket *sock;
 875
 876         /* special case to disable backend */
 877         if (fd == -1)
 878                 return NULL;
 879         sock = get_raw_socket(fd);
 880         if (!IS_ERR(sock))
 881                 return sock;
 882         sock = get_tap_socket(fd);
 883         if (!IS_ERR(sock))
 884                 return sock;
 885         return ERR_PTR(-ENOTSOCK);
 886 }
 887
 888 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 889 {
 890         struct socket *sock, *oldsock;
 891         struct vhost_virtqueue *vq;
 892         struct vhost_net_virtqueue *nvq;
 893         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 894         int r;
 895
 896         mutex_lock(&n->dev.mutex);
 897         r = vhost_dev_check_owner(&n->dev);
 898         if (r)
 899                 goto err;
 900
 901         if (index >= VHOST_NET_VQ_MAX) {
 902                 r = -ENOBUFS;
 903                 goto err;
 904         }
 905         vq = &n->vqs[index].vq;
 906         nvq = &n->vqs[index];
 907         mutex_lock(&vq->mutex);
 908
 909         /* Verify that ring has been setup correctly. */
 910         if (!vhost_vq_access_ok(vq)) {
 911                 r = -EFAULT;
 912                 goto err_vq;
 913         }
 914         sock = get_socket(fd);
 915         if (IS_ERR(sock)) {
 916                 r = PTR_ERR(sock);
 917                 goto err_vq;
 918         }
 919
 920         /* start polling new socket */
 921         oldsock = rcu_dereference_protected(vq->private_data,
 922                                             lockdep_is_held(&vq->mutex));
 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                 rcu_assign_pointer(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_and_wait(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         rcu_assign_pointer(vq->private_data, oldsock);
 967         vhost_net_enable_vq(n, vq);
 968         if (ubufs)
 969                 vhost_net_ubuf_put_and_wait(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");