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 #include <linux/vmalloc.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         /* refcount follows semantics similar to kref:
  75          *  0: object is released
  76          *  1: no outstanding ubufs
  77          * >1: outstanding ubufs
  78          */
  79         atomic_t refcount;
  80         wait_queue_head_t wait;
  81         struct vhost_virtqueue *vq;
  82 };
  83
  84 struct vhost_net_virtqueue {
  85         struct vhost_virtqueue vq;
  86         /* hdr is used to store the virtio header.
  87          * Since each iovec has >= 1 byte length, we never need more than
  88          * header length entries to store the header. */
  89         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
  90         size_t vhost_hlen;
  91         size_t sock_hlen;
  92         /* vhost zerocopy support fields below: */
  93         /* last used idx for outstanding DMA zerocopy buffers */
  94         int upend_idx;
  95         /* first used idx for DMA done zerocopy buffers */
  96         int done_idx;
  97         /* an array of userspace buffers info */
  98         struct ubuf_info *ubuf_info;
  99         /* Reference counting for outstanding ubufs.
 100          * Protected by vq mutex. Writers must also take device mutex. */
 101         struct vhost_net_ubuf_ref *ubufs;
 102 };
 103
 104 struct vhost_net {
 105         struct vhost_dev dev;
 106         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 107         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 108         /* Number of TX recently submitted.
 109          * Protected by tx vq lock. */
 110         unsigned tx_packets;
 111         /* Number of times zerocopy TX recently failed.
 112          * Protected by tx vq lock. */
 113         unsigned tx_zcopy_err;
 114         /* Flush in progress. Protected by tx vq lock. */
 115         bool tx_flush;
 116 };
 117
 118 static unsigned vhost_net_zcopy_mask __read_mostly;
 119
 120 static void vhost_net_enable_zcopy(int vq)
 121 {
 122         vhost_net_zcopy_mask |= 0x1 << vq;
 123 }
 124
 125 static struct vhost_net_ubuf_ref *
 126 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 127 {
 128         struct vhost_net_ubuf_ref *ubufs;
 129         /* No zero copy backend? Nothing to count. */
 130         if (!zcopy)
 131                 return NULL;
 132         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 133         if (!ubufs)
 134                 return ERR_PTR(-ENOMEM);
 135         atomic_set(&ubufs->refcount, 1);
 136         init_waitqueue_head(&ubufs->wait);
 137         ubufs->vq = vq;
 138         return ubufs;
 139 }
 140
 141 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 142 {
 143         int r = atomic_sub_return(1, &ubufs->refcount);
 144         if (unlikely(!r))
 145                 wake_up(&ubufs->wait);
 146         return r;
 147 }
 148
 149 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 150 {
 151         vhost_net_ubuf_put(ubufs);
 152         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 153 }
 154
 155 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 156 {
 157         vhost_net_ubuf_put_and_wait(ubufs);
 158         kfree(ubufs);
 159 }
 160
 161 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 162 {
 163         int i;
 164
 165         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 166                 kfree(n->vqs[i].ubuf_info);
 167                 n->vqs[i].ubuf_info = NULL;
 168         }
 169 }
 170
 171 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 172 {
 173         bool zcopy;
 174         int i;
 175
 176         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 177                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 178                 if (!zcopy)
 179                         continue;
 180                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 181                                               UIO_MAXIOV, GFP_KERNEL);
 182                 if  (!n->vqs[i].ubuf_info)
 183                         goto err;
 184         }
 185         return 0;
 186
 187 err:
 188         vhost_net_clear_ubuf_info(n);
 189         return -ENOMEM;
 190 }
 191
 192 static void vhost_net_vq_reset(struct vhost_net *n)
 193 {
 194         int i;
 195
 196         vhost_net_clear_ubuf_info(n);
 197
 198         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 199                 n->vqs[i].done_idx = 0;
 200                 n->vqs[i].upend_idx = 0;
 201                 n->vqs[i].ubufs = NULL;
 202                 n->vqs[i].vhost_hlen = 0;
 203                 n->vqs[i].sock_hlen = 0;
 204         }
 205
 206 }
 207
 208 static void vhost_net_tx_packet(struct vhost_net *net)
 209 {
 210         ++net->tx_packets;
 211         if (net->tx_packets < 1024)
 212                 return;
 213         net->tx_packets = 0;
 214         net->tx_zcopy_err = 0;
 215 }
 216
 217 static void vhost_net_tx_err(struct vhost_net *net)
 218 {
 219         ++net->tx_zcopy_err;
 220 }
 221
 222 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 223 {
 224         /* TX flush waits for outstanding DMAs to be done.
 225          * Don't start new DMAs.
 226          */
 227         return !net->tx_flush &&
 228                 net->tx_packets / 64 >= net->tx_zcopy_err;
 229 }
 230
 231 static bool vhost_sock_zcopy(struct socket *sock)
 232 {
 233         return unlikely(experimental_zcopytx) &&
 234                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 235 }
 236
 237 /* Pop first len bytes from iovec. Return number of segments used. */
 238 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 239                           size_t len, int iov_count)
 240 {
 241         int seg = 0;
 242         size_t size;
 243
 244         while (len && seg < iov_count) {
 245                 size = min(from->iov_len, len);
 246                 to->iov_base = from->iov_base;
 247                 to->iov_len = size;
 248                 from->iov_len -= size;
 249                 from->iov_base += size;
 250                 len -= size;
 251                 ++from;
 252                 ++to;
 253                 ++seg;
 254         }
 255         return seg;
 256 }
 257 /* Copy iovec entries for len bytes from iovec. */
 258 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 259                            size_t len, int iovcount)
 260 {
 261         int seg = 0;
 262         size_t size;
 263
 264         while (len && seg < iovcount) {
 265                 size = min(from->iov_len, len);
 266                 to->iov_base = from->iov_base;
 267                 to->iov_len = size;
 268                 len -= size;
 269                 ++from;
 270                 ++to;
 271                 ++seg;
 272         }
 273 }
 274
 275 /* In case of DMA done not in order in lower device driver for some reason.
 276  * upend_idx is used to track end of used idx, done_idx is used to track head
 277  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 278  * guest used idx.
 279  */
 280 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 281                                        struct vhost_virtqueue *vq)
 282 {
 283         struct vhost_net_virtqueue *nvq =
 284                 container_of(vq, struct vhost_net_virtqueue, vq);
 285         int i, add;
 286         int j = 0;
 287
 288         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 289                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 290                         vhost_net_tx_err(net);
 291                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 292                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 293                         ++j;
 294                 } else
 295                         break;
 296         }
 297         while (j) {
 298                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 299                 vhost_add_used_and_signal_n(vq->dev, vq,
 300                                             &vq->heads[nvq->done_idx], add);
 301                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 302                 j -= add;
 303         }
 304 }
 305
 306 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 307 {
 308         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 309         struct vhost_virtqueue *vq = ubufs->vq;
 310         int cnt;
 311
 312         rcu_read_lock_bh();
 313
 314         /* set len to mark this desc buffers done DMA */
 315         vq->heads[ubuf->desc].len = success ?
 316                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 317         cnt = vhost_net_ubuf_put(ubufs);
 318
 319         /*
 320          * Trigger polling thread if guest stopped submitting new buffers:
 321          * in this case, the refcount after decrement will eventually reach 1.
 322          * We also trigger polling periodically after each 16 packets
 323          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 324          * less than 10% of times).
 325          */
 326         if (cnt <= 1 || !(cnt % 16))
 327                 vhost_poll_queue(&vq->poll);
 328
 329         rcu_read_unlock_bh();
 330 }
 331
 332 /* Expects to be always run from workqueue - which acts as
 333  * read-size critical section for our kind of RCU. */
 334 static void handle_tx(struct vhost_net *net)
 335 {
 336         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 337         struct vhost_virtqueue *vq = &nvq->vq;
 338         unsigned out, in, s;
 339         int head;
 340         struct msghdr msg = {
 341                 .msg_name = NULL,
 342                 .msg_namelen = 0,
 343                 .msg_control = NULL,
 344                 .msg_controllen = 0,
 345                 .msg_iov = vq->iov,
 346                 .msg_flags = MSG_DONTWAIT,
 347         };
 348         size_t len, total_len = 0;
 349         int err;
 350         size_t hdr_size;
 351         struct socket *sock;
 352         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 353         bool zcopy, zcopy_used;
 354
 355         mutex_lock(&vq->mutex);
 356         sock = vq->private_data;
 357         if (!sock)
 358                 goto out;
 359
 360         vhost_disable_notify(&net->dev, vq);
 361
 362         hdr_size = nvq->vhost_hlen;
 363         zcopy = nvq->ubufs;
 364
 365         for (;;) {
 366                 /* Release DMAs done buffers first */
 367                 if (zcopy)
 368                         vhost_zerocopy_signal_used(net, vq);
 369
 370                 /* If more outstanding DMAs, queue the work.
 371                  * Handle upend_idx wrap around
 372                  */
 373                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
 374                               % UIO_MAXIOV == nvq->done_idx))
 375                         break;
 376
 377                 head = vhost_get_vq_desc(vq, vq->iov,
 378                                          ARRAY_SIZE(vq->iov),
 379                                          &out, &in,
 380                                          NULL, NULL);
 381                 /* On error, stop handling until the next kick. */
 382                 if (unlikely(head < 0))
 383                         break;
 384                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 385                 if (head == vq->num) {
 386                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 387                                 vhost_disable_notify(&net->dev, vq);
 388                                 continue;
 389                         }
 390                         break;
 391                 }
 392                 if (in) {
 393                         vq_err(vq, "Unexpected descriptor format for TX: "
 394                                "out %d, int %d\n", out, in);
 395                         break;
 396                 }
 397                 /* Skip header. TODO: support TSO. */
 398                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 399                 msg.msg_iovlen = out;
 400                 len = iov_length(vq->iov, out);
 401                 /* Sanity check */
 402                 if (!len) {
 403                         vq_err(vq, "Unexpected header len for TX: "
 404                                "%zd expected %zd\n",
 405                                iov_length(nvq->hdr, s), hdr_size);
 406                         break;
 407                 }
 408
 409                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 410                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 411                                       nvq->done_idx
 412                                    && vhost_net_tx_select_zcopy(net);
 413
 414                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 415                 if (zcopy_used) {
 416                         struct ubuf_info *ubuf;
 417                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 418
 419                         vq->heads[nvq->upend_idx].id = head;
 420                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 421                         ubuf->callback = vhost_zerocopy_callback;
 422                         ubuf->ctx = nvq->ubufs;
 423                         ubuf->desc = nvq->upend_idx;
 424                         msg.msg_control = ubuf;
 425                         msg.msg_controllen = sizeof(ubuf);
 426                         ubufs = nvq->ubufs;
 427                         atomic_inc(&ubufs->refcount);
 428                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 429                 } else {
 430                         msg.msg_control = NULL;
 431                         ubufs = NULL;
 432                 }
 433                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 434                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 435                 if (unlikely(err < 0)) {
 436                         if (zcopy_used) {
 437                                 vhost_net_ubuf_put(ubufs);
 438                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 439                                         % UIO_MAXIOV;
 440                         }
 441                         vhost_discard_vq_desc(vq, 1);
 442                         break;
 443                 }
 444                 if (err != len)
 445                         pr_debug("Truncated TX packet: "
 446                                  " len %d != %zd\n", err, len);
 447                 if (!zcopy_used)
 448                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 449                 else
 450                         vhost_zerocopy_signal_used(net, vq);
 451                 total_len += len;
 452                 vhost_net_tx_packet(net);
 453                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 454                         vhost_poll_queue(&vq->poll);
 455                         break;
 456                 }
 457         }
 458 out:
 459         mutex_unlock(&vq->mutex);
 460 }
 461
 462 static int peek_head_len(struct sock *sk)
 463 {
 464         struct sk_buff *head;
 465         int len = 0;
 466         unsigned long flags;
 467
 468         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 469         head = skb_peek(&sk->sk_receive_queue);
 470         if (likely(head)) {
 471                 len = head->len;
 472                 if (vlan_tx_tag_present(head))
 473                         len += VLAN_HLEN;
 474         }
 475
 476         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 477         return len;
 478 }
 479
 480 /* This is a multi-buffer version of vhost_get_desc, that works if
 481  *      vq has read descriptors only.
 482  * @vq          - the relevant virtqueue
 483  * @datalen     - data length we'll be reading
 484  * @iovcount    - returned count of io vectors we fill
 485  * @log         - vhost log
 486  * @log_num     - log offset
 487  * @quota       - headcount quota, 1 for big buffer
 488  *      returns number of buffer heads allocated, negative on error
 489  */
 490 static int get_rx_bufs(struct vhost_virtqueue *vq,
 491                        struct vring_used_elem *heads,
 492                        int datalen,
 493                        unsigned *iovcount,
 494                        struct vhost_log *log,
 495                        unsigned *log_num,
 496                        unsigned int quota)
 497 {
 498         unsigned int out, in;
 499         int seg = 0;
 500         int headcount = 0;
 501         unsigned d;
 502         int r, nlogs = 0;
 503
 504         while (datalen > 0 && headcount < quota) {
 505                 if (unlikely(seg >= UIO_MAXIOV)) {
 506                         r = -ENOBUFS;
 507                         goto err;
 508                 }
 509                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 510                                       ARRAY_SIZE(vq->iov) - seg, &out,
 511                                       &in, log, log_num);
 512                 if (unlikely(r < 0))
 513                         goto err;
 514
 515                 d = r;
 516                 if (d == vq->num) {
 517                         r = 0;
 518                         goto err;
 519                 }
 520                 if (unlikely(out || in <= 0)) {
 521                         vq_err(vq, "unexpected descriptor format for RX: "
 522                                 "out %d, in %d\n", out, in);
 523                         r = -EINVAL;
 524                         goto err;
 525                 }
 526                 if (unlikely(log)) {
 527                         nlogs += *log_num;
 528                         log += *log_num;
 529                 }
 530                 heads[headcount].id = d;
 531                 heads[headcount].len = iov_length(vq->iov + seg, in);
 532                 datalen -= heads[headcount].len;
 533                 ++headcount;
 534                 seg += in;
 535         }
 536         heads[headcount - 1].len += datalen;
 537         *iovcount = seg;
 538         if (unlikely(log))
 539                 *log_num = nlogs;
 540
 541         /* Detect overrun */
 542         if (unlikely(datalen > 0)) {
 543                 r = UIO_MAXIOV + 1;
 544                 goto err;
 545         }
 546         return headcount;
 547 err:
 548         vhost_discard_vq_desc(vq, headcount);
 549         return r;
 550 }
 551
 552 /* Expects to be always run from workqueue - which acts as
 553  * read-size critical section for our kind of RCU. */
 554 static void handle_rx(struct vhost_net *net)
 555 {
 556         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 557         struct vhost_virtqueue *vq = &nvq->vq;
 558         unsigned uninitialized_var(in), log;
 559         struct vhost_log *vq_log;
 560         struct msghdr msg = {
 561                 .msg_name = NULL,
 562                 .msg_namelen = 0,
 563                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 564                 .msg_controllen = 0,
 565                 .msg_iov = vq->iov,
 566                 .msg_flags = MSG_DONTWAIT,
 567         };
 568         struct virtio_net_hdr_mrg_rxbuf hdr = {
 569                 .hdr.flags = 0,
 570                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 571         };
 572         size_t total_len = 0;
 573         int err, mergeable;
 574         s16 headcount;
 575         size_t vhost_hlen, sock_hlen;
 576         size_t vhost_len, sock_len;
 577         struct socket *sock;
 578
 579         mutex_lock(&vq->mutex);
 580         sock = vq->private_data;
 581         if (!sock)
 582                 goto out;
 583         vhost_disable_notify(&net->dev, vq);
 584
 585         vhost_hlen = nvq->vhost_hlen;
 586         sock_hlen = nvq->sock_hlen;
 587
 588         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 589                 vq->log : NULL;
 590         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 591
 592         while ((sock_len = peek_head_len(sock->sk))) {
 593                 sock_len += sock_hlen;
 594                 vhost_len = sock_len + vhost_hlen;
 595                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 596                                         &in, vq_log, &log,
 597                                         likely(mergeable) ? UIO_MAXIOV : 1);
 598                 /* On error, stop handling until the next kick. */
 599                 if (unlikely(headcount < 0))
 600                         break;
 601                 /* On overrun, truncate and discard */
 602                 if (unlikely(headcount > UIO_MAXIOV)) {
 603                         msg.msg_iovlen = 1;
 604                         err = sock->ops->recvmsg(NULL, sock, &msg,
 605                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 606                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 607                         continue;
 608                 }
 609                 /* OK, now we need to know about added descriptors. */
 610                 if (!headcount) {
 611                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 612                                 /* They have slipped one in as we were
 613                                  * doing that: check again. */
 614                                 vhost_disable_notify(&net->dev, vq);
 615                                 continue;
 616                         }
 617                         /* Nothing new?  Wait for eventfd to tell us
 618                          * they refilled. */
 619                         break;
 620                 }
 621                 /* We don't need to be notified again. */
 622                 if (unlikely((vhost_hlen)))
 623                         /* Skip header. TODO: support TSO. */
 624                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 625                 else
 626                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 627                          * needed because recvmsg can modify msg_iov. */
 628                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 629                 msg.msg_iovlen = in;
 630                 err = sock->ops->recvmsg(NULL, sock, &msg,
 631                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 632                 /* Userspace might have consumed the packet meanwhile:
 633                  * it's not supposed to do this usually, but might be hard
 634                  * to prevent. Discard data we got (if any) and keep going. */
 635                 if (unlikely(err != sock_len)) {
 636                         pr_debug("Discarded rx packet: "
 637                                  " len %d, expected %zd\n", err, sock_len);
 638                         vhost_discard_vq_desc(vq, headcount);
 639                         continue;
 640                 }
 641                 if (unlikely(vhost_hlen) &&
 642                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 643                                       vhost_hlen)) {
 644                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 645                                vq->iov->iov_base);
 646                         break;
 647                 }
 648                 /* TODO: Should check and handle checksum. */
 649                 if (likely(mergeable) &&
 650                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 651                                       offsetof(typeof(hdr), num_buffers),
 652                                       sizeof hdr.num_buffers)) {
 653                         vq_err(vq, "Failed num_buffers write");
 654                         vhost_discard_vq_desc(vq, headcount);
 655                         break;
 656                 }
 657                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 658                                             headcount);
 659                 if (unlikely(vq_log))
 660                         vhost_log_write(vq, vq_log, log, vhost_len);
 661                 total_len += vhost_len;
 662                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 663                         vhost_poll_queue(&vq->poll);
 664                         break;
 665                 }
 666         }
 667 out:
 668         mutex_unlock(&vq->mutex);
 669 }
 670
 671 static void handle_tx_kick(struct vhost_work *work)
 672 {
 673         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 674                                                   poll.work);
 675         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 676
 677         handle_tx(net);
 678 }
 679
 680 static void handle_rx_kick(struct vhost_work *work)
 681 {
 682         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 683                                                   poll.work);
 684         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 685
 686         handle_rx(net);
 687 }
 688
 689 static void handle_tx_net(struct vhost_work *work)
 690 {
 691         struct vhost_net *net = container_of(work, struct vhost_net,
 692                                              poll[VHOST_NET_VQ_TX].work);
 693         handle_tx(net);
 694 }
 695
 696 static void handle_rx_net(struct vhost_work *work)
 697 {
 698         struct vhost_net *net = container_of(work, struct vhost_net,
 699                                              poll[VHOST_NET_VQ_RX].work);
 700         handle_rx(net);
 701 }
 702
 703 static int vhost_net_open(struct inode *inode, struct file *f)
 704 {
 705         struct vhost_net *n;
 706         struct vhost_dev *dev;
 707         struct vhost_virtqueue **vqs;
 708         int i;
 709
 710         n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 711         if (!n) {
 712                 n = vmalloc(sizeof *n);
 713                 if (!n)
 714                         return -ENOMEM;
 715         }
 716         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 717         if (!vqs) {
 718                 kvfree(n);
 719                 return -ENOMEM;
 720         }
 721
 722         dev = &n->dev;
 723         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 724         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 725         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 726         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 727         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 728                 n->vqs[i].ubufs = NULL;
 729                 n->vqs[i].ubuf_info = NULL;
 730                 n->vqs[i].upend_idx = 0;
 731                 n->vqs[i].done_idx = 0;
 732                 n->vqs[i].vhost_hlen = 0;
 733                 n->vqs[i].sock_hlen = 0;
 734         }
 735         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 736
 737         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 738         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 739
 740         f->private_data = n;
 741
 742         return 0;
 743 }
 744
 745 static void vhost_net_disable_vq(struct vhost_net *n,
 746                                  struct vhost_virtqueue *vq)
 747 {
 748         struct vhost_net_virtqueue *nvq =
 749                 container_of(vq, struct vhost_net_virtqueue, vq);
 750         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 751         if (!vq->private_data)
 752                 return;
 753         vhost_poll_stop(poll);
 754 }
 755
 756 static int vhost_net_enable_vq(struct vhost_net *n,
 757                                 struct vhost_virtqueue *vq)
 758 {
 759         struct vhost_net_virtqueue *nvq =
 760                 container_of(vq, struct vhost_net_virtqueue, vq);
 761         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 762         struct socket *sock;
 763
 764         sock = vq->private_data;
 765         if (!sock)
 766                 return 0;
 767
 768         return vhost_poll_start(poll, sock->file);
 769 }
 770
 771 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 772                                         struct vhost_virtqueue *vq)
 773 {
 774         struct socket *sock;
 775
 776         mutex_lock(&vq->mutex);
 777         sock = vq->private_data;
 778         vhost_net_disable_vq(n, vq);
 779         vq->private_data = NULL;
 780         mutex_unlock(&vq->mutex);
 781         return sock;
 782 }
 783
 784 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 785                            struct socket **rx_sock)
 786 {
 787         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 788         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 789 }
 790
 791 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 792 {
 793         vhost_poll_flush(n->poll + index);
 794         vhost_poll_flush(&n->vqs[index].vq.poll);
 795 }
 796
 797 static void vhost_net_flush(struct vhost_net *n)
 798 {
 799         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 800         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 801         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 802                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 803                 n->tx_flush = true;
 804                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 805                 /* Wait for all lower device DMAs done. */
 806                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 807                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 808                 n->tx_flush = false;
 809                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 810                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 811         }
 812 }
 813
 814 static int vhost_net_release(struct inode *inode, struct file *f)
 815 {
 816         struct vhost_net *n = f->private_data;
 817         struct socket *tx_sock;
 818         struct socket *rx_sock;
 819
 820         vhost_net_stop(n, &tx_sock, &rx_sock);
 821         vhost_net_flush(n);
 822         vhost_dev_stop(&n->dev);
 823         vhost_dev_cleanup(&n->dev, false);
 824         vhost_net_vq_reset(n);
 825         if (tx_sock)
 826                 sockfd_put(tx_sock);
 827         if (rx_sock)
 828                 sockfd_put(rx_sock);
 829         /* Make sure no callbacks are outstanding */
 830         synchronize_rcu_bh();
 831         /* We do an extra flush before freeing memory,
 832          * since jobs can re-queue themselves. */
 833         vhost_net_flush(n);
 834         kfree(n->dev.vqs);
 835         kvfree(n);
 836         return 0;
 837 }
 838
 839 static struct socket *get_raw_socket(int fd)
 840 {
 841         struct {
 842                 struct sockaddr_ll sa;
 843                 char  buf[MAX_ADDR_LEN];
 844         } uaddr;
 845         int uaddr_len = sizeof uaddr, r;
 846         struct socket *sock = sockfd_lookup(fd, &r);
 847
 848         if (!sock)
 849                 return ERR_PTR(-ENOTSOCK);
 850
 851         /* Parameter checking */
 852         if (sock->sk->sk_type != SOCK_RAW) {
 853                 r = -ESOCKTNOSUPPORT;
 854                 goto err;
 855         }
 856
 857         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 858                                &uaddr_len, 0);
 859         if (r)
 860                 goto err;
 861
 862         if (uaddr.sa.sll_family != AF_PACKET) {
 863                 r = -EPFNOSUPPORT;
 864                 goto err;
 865         }
 866         return sock;
 867 err:
 868         sockfd_put(sock);
 869         return ERR_PTR(r);
 870 }
 871
 872 static struct socket *get_tap_socket(int fd)
 873 {
 874         struct file *file = fget(fd);
 875         struct socket *sock;
 876
 877         if (!file)
 878                 return ERR_PTR(-EBADF);
 879         sock = tun_get_socket(file);
 880         if (!IS_ERR(sock))
 881                 return sock;
 882         sock = macvtap_get_socket(file);
 883         if (IS_ERR(sock))
 884                 fput(file);
 885         return sock;
 886 }
 887
 888 static struct socket *get_socket(int fd)
 889 {
 890         struct socket *sock;
 891
 892         /* special case to disable backend */
 893         if (fd == -1)
 894                 return NULL;
 895         sock = get_raw_socket(fd);
 896         if (!IS_ERR(sock))
 897                 return sock;
 898         sock = get_tap_socket(fd);
 899         if (!IS_ERR(sock))
 900                 return sock;
 901         return ERR_PTR(-ENOTSOCK);
 902 }
 903
 904 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 905 {
 906         struct socket *sock, *oldsock;
 907         struct vhost_virtqueue *vq;
 908         struct vhost_net_virtqueue *nvq;
 909         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 910         int r;
 911
 912         mutex_lock(&n->dev.mutex);
 913         r = vhost_dev_check_owner(&n->dev);
 914         if (r)
 915                 goto err;
 916
 917         if (index >= VHOST_NET_VQ_MAX) {
 918                 r = -ENOBUFS;
 919                 goto err;
 920         }
 921         vq = &n->vqs[index].vq;
 922         nvq = &n->vqs[index];
 923         mutex_lock(&vq->mutex);
 924
 925         /* Verify that ring has been setup correctly. */
 926         if (!vhost_vq_access_ok(vq)) {
 927                 r = -EFAULT;
 928                 goto err_vq;
 929         }
 930         sock = get_socket(fd);
 931         if (IS_ERR(sock)) {
 932                 r = PTR_ERR(sock);
 933                 goto err_vq;
 934         }
 935
 936         /* start polling new socket */
 937         oldsock = vq->private_data;
 938         if (sock != oldsock) {
 939                 ubufs = vhost_net_ubuf_alloc(vq,
 940                                              sock && vhost_sock_zcopy(sock));
 941                 if (IS_ERR(ubufs)) {
 942                         r = PTR_ERR(ubufs);
 943                         goto err_ubufs;
 944                 }
 945
 946                 vhost_net_disable_vq(n, vq);
 947                 vq->private_data = sock;
 948                 r = vhost_init_used(vq);
 949                 if (r)
 950                         goto err_used;
 951                 r = vhost_net_enable_vq(n, vq);
 952                 if (r)
 953                         goto err_used;
 954
 955                 oldubufs = nvq->ubufs;
 956                 nvq->ubufs = ubufs;
 957
 958                 n->tx_packets = 0;
 959                 n->tx_zcopy_err = 0;
 960                 n->tx_flush = false;
 961         }
 962
 963         mutex_unlock(&vq->mutex);
 964
 965         if (oldubufs) {
 966                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 967                 mutex_lock(&vq->mutex);
 968                 vhost_zerocopy_signal_used(n, vq);
 969                 mutex_unlock(&vq->mutex);
 970         }
 971
 972         if (oldsock) {
 973                 vhost_net_flush_vq(n, index);
 974                 sockfd_put(oldsock);
 975         }
 976
 977         mutex_unlock(&n->dev.mutex);
 978         return 0;
 979
 980 err_used:
 981         vq->private_data = oldsock;
 982         vhost_net_enable_vq(n, vq);
 983         if (ubufs)
 984                 vhost_net_ubuf_put_wait_and_free(ubufs);
 985 err_ubufs:
 986         sockfd_put(sock);
 987 err_vq:
 988         mutex_unlock(&vq->mutex);
 989 err:
 990         mutex_unlock(&n->dev.mutex);
 991         return r;
 992 }
 993
 994 static long vhost_net_reset_owner(struct vhost_net *n)
 995 {
 996         struct socket *tx_sock = NULL;
 997         struct socket *rx_sock = NULL;
 998         long err;
 999         struct vhost_memory *memory;
1000
1001         mutex_lock(&n->dev.mutex);
1002         err = vhost_dev_check_owner(&n->dev);
1003         if (err)
1004                 goto done;
1005         memory = vhost_dev_reset_owner_prepare();
1006         if (!memory) {
1007                 err = -ENOMEM;
1008                 goto done;
1009         }
1010         vhost_net_stop(n, &tx_sock, &rx_sock);
1011         vhost_net_flush(n);
1012         vhost_dev_reset_owner(&n->dev, memory);
1013         vhost_net_vq_reset(n);
1014 done:
1015         mutex_unlock(&n->dev.mutex);
1016         if (tx_sock)
1017                 sockfd_put(tx_sock);
1018         if (rx_sock)
1019                 sockfd_put(rx_sock);
1020         return err;
1021 }
1022
1023 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1024 {
1025         size_t vhost_hlen, sock_hlen, hdr_len;
1026         int i;
1027
1028         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1029                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1030                         sizeof(struct virtio_net_hdr);
1031         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1032                 /* vhost provides vnet_hdr */
1033                 vhost_hlen = hdr_len;
1034                 sock_hlen = 0;
1035         } else {
1036                 /* socket provides vnet_hdr */
1037                 vhost_hlen = 0;
1038                 sock_hlen = hdr_len;
1039         }
1040         mutex_lock(&n->dev.mutex);
1041         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1042             !vhost_log_access_ok(&n->dev)) {
1043                 mutex_unlock(&n->dev.mutex);
1044                 return -EFAULT;
1045         }
1046         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1047                 mutex_lock(&n->vqs[i].vq.mutex);
1048                 n->vqs[i].vq.acked_features = features;
1049                 n->vqs[i].vhost_hlen = vhost_hlen;
1050                 n->vqs[i].sock_hlen = sock_hlen;
1051                 mutex_unlock(&n->vqs[i].vq.mutex);
1052         }
1053         mutex_unlock(&n->dev.mutex);
1054         return 0;
1055 }
1056
1057 static long vhost_net_set_owner(struct vhost_net *n)
1058 {
1059         int r;
1060
1061         mutex_lock(&n->dev.mutex);
1062         if (vhost_dev_has_owner(&n->dev)) {
1063                 r = -EBUSY;
1064                 goto out;
1065         }
1066         r = vhost_net_set_ubuf_info(n);
1067         if (r)
1068                 goto out;
1069         r = vhost_dev_set_owner(&n->dev);
1070         if (r)
1071                 vhost_net_clear_ubuf_info(n);
1072         vhost_net_flush(n);
1073 out:
1074         mutex_unlock(&n->dev.mutex);
1075         return r;
1076 }
1077
1078 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1079                             unsigned long arg)
1080 {
1081         struct vhost_net *n = f->private_data;
1082         void __user *argp = (void __user *)arg;
1083         u64 __user *featurep = argp;
1084         struct vhost_vring_file backend;
1085         u64 features;
1086         int r;
1087
1088         switch (ioctl) {
1089         case VHOST_NET_SET_BACKEND:
1090                 if (copy_from_user(&backend, argp, sizeof backend))
1091                         return -EFAULT;
1092                 return vhost_net_set_backend(n, backend.index, backend.fd);
1093         case VHOST_GET_FEATURES:
1094                 features = VHOST_NET_FEATURES;
1095                 if (copy_to_user(featurep, &features, sizeof features))
1096                         return -EFAULT;
1097                 return 0;
1098         case VHOST_SET_FEATURES:
1099                 if (copy_from_user(&features, featurep, sizeof features))
1100                         return -EFAULT;
1101                 if (features & ~VHOST_NET_FEATURES)
1102                         return -EOPNOTSUPP;
1103                 return vhost_net_set_features(n, features);
1104         case VHOST_RESET_OWNER:
1105                 return vhost_net_reset_owner(n);
1106         case VHOST_SET_OWNER:
1107                 return vhost_net_set_owner(n);
1108         default:
1109                 mutex_lock(&n->dev.mutex);
1110                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1111                 if (r == -ENOIOCTLCMD)
1112                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1113                 else
1114                         vhost_net_flush(n);
1115                 mutex_unlock(&n->dev.mutex);
1116                 return r;
1117         }
1118 }
1119
1120 #ifdef CONFIG_COMPAT
1121 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1122                                    unsigned long arg)
1123 {
1124         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1125 }
1126 #endif
1127
1128 static const struct file_operations vhost_net_fops = {
1129         .owner          = THIS_MODULE,
1130         .release        = vhost_net_release,
1131         .unlocked_ioctl = vhost_net_ioctl,
1132 #ifdef CONFIG_COMPAT
1133         .compat_ioctl   = vhost_net_compat_ioctl,
1134 #endif
1135         .open           = vhost_net_open,
1136         .llseek         = noop_llseek,
1137 };
1138
1139 static struct miscdevice vhost_net_misc = {
1140         .minor = VHOST_NET_MINOR,
1141         .name = "vhost-net",
1142         .fops = &vhost_net_fops,
1143 };
1144
1145 static int vhost_net_init(void)
1146 {
1147         if (experimental_zcopytx)
1148                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1149         return misc_register(&vhost_net_misc);
1150 }
1151 module_init(vhost_net_init);
1152
1153 static void vhost_net_exit(void)
1154 {
1155         misc_deregister(&vhost_net_misc);
1156 }
1157 module_exit(vhost_net_exit);
1158
1159 MODULE_VERSION("0.0.1");
1160 MODULE_LICENSE("GPL v2");
1161 MODULE_AUTHOR("Michael S. Tsirkin");
1162 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1163 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1164 MODULE_ALIAS("devname:vhost-net");