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 void vhost_net_free(void *addr)
 704 {
 705         if (is_vmalloc_addr(addr))
 706                 vfree(addr);
 707         else
 708                 kfree(addr);
 709 }
 710
 711 static int vhost_net_open(struct inode *inode, struct file *f)
 712 {
 713         struct vhost_net *n;
 714         struct vhost_dev *dev;
 715         struct vhost_virtqueue **vqs;
 716         int i;
 717
 718         n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 719         if (!n) {
 720                 n = vmalloc(sizeof *n);
 721                 if (!n)
 722                         return -ENOMEM;
 723         }
 724         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 725         if (!vqs) {
 726                 vhost_net_free(n);
 727                 return -ENOMEM;
 728         }
 729
 730         dev = &n->dev;
 731         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 732         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 733         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 734         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 735         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 736                 n->vqs[i].ubufs = NULL;
 737                 n->vqs[i].ubuf_info = NULL;
 738                 n->vqs[i].upend_idx = 0;
 739                 n->vqs[i].done_idx = 0;
 740                 n->vqs[i].vhost_hlen = 0;
 741                 n->vqs[i].sock_hlen = 0;
 742         }
 743         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 744
 745         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 746         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 747
 748         f->private_data = n;
 749
 750         return 0;
 751 }
 752
 753 static void vhost_net_disable_vq(struct vhost_net *n,
 754                                  struct vhost_virtqueue *vq)
 755 {
 756         struct vhost_net_virtqueue *nvq =
 757                 container_of(vq, struct vhost_net_virtqueue, vq);
 758         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 759         if (!vq->private_data)
 760                 return;
 761         vhost_poll_stop(poll);
 762 }
 763
 764 static int vhost_net_enable_vq(struct vhost_net *n,
 765                                 struct vhost_virtqueue *vq)
 766 {
 767         struct vhost_net_virtqueue *nvq =
 768                 container_of(vq, struct vhost_net_virtqueue, vq);
 769         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 770         struct socket *sock;
 771
 772         sock = vq->private_data;
 773         if (!sock)
 774                 return 0;
 775
 776         return vhost_poll_start(poll, sock->file);
 777 }
 778
 779 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 780                                         struct vhost_virtqueue *vq)
 781 {
 782         struct socket *sock;
 783
 784         mutex_lock(&vq->mutex);
 785         sock = vq->private_data;
 786         vhost_net_disable_vq(n, vq);
 787         vq->private_data = NULL;
 788         mutex_unlock(&vq->mutex);
 789         return sock;
 790 }
 791
 792 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 793                            struct socket **rx_sock)
 794 {
 795         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 796         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 797 }
 798
 799 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 800 {
 801         vhost_poll_flush(n->poll + index);
 802         vhost_poll_flush(&n->vqs[index].vq.poll);
 803 }
 804
 805 static void vhost_net_flush(struct vhost_net *n)
 806 {
 807         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 808         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 809         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 810                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 811                 n->tx_flush = true;
 812                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 813                 /* Wait for all lower device DMAs done. */
 814                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 815                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 816                 n->tx_flush = false;
 817                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 818                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 819         }
 820 }
 821
 822 static int vhost_net_release(struct inode *inode, struct file *f)
 823 {
 824         struct vhost_net *n = f->private_data;
 825         struct socket *tx_sock;
 826         struct socket *rx_sock;
 827
 828         vhost_net_stop(n, &tx_sock, &rx_sock);
 829         vhost_net_flush(n);
 830         vhost_dev_stop(&n->dev);
 831         vhost_dev_cleanup(&n->dev, false);
 832         vhost_net_vq_reset(n);
 833         if (tx_sock)
 834                 sockfd_put(tx_sock);
 835         if (rx_sock)
 836                 sockfd_put(rx_sock);
 837         /* Make sure no callbacks are outstanding */
 838         synchronize_rcu_bh();
 839         /* We do an extra flush before freeing memory,
 840          * since jobs can re-queue themselves. */
 841         vhost_net_flush(n);
 842         kfree(n->dev.vqs);
 843         vhost_net_free(n);
 844         return 0;
 845 }
 846
 847 static struct socket *get_raw_socket(int fd)
 848 {
 849         struct {
 850                 struct sockaddr_ll sa;
 851                 char  buf[MAX_ADDR_LEN];
 852         } uaddr;
 853         int uaddr_len = sizeof uaddr, r;
 854         struct socket *sock = sockfd_lookup(fd, &r);
 855
 856         if (!sock)
 857                 return ERR_PTR(-ENOTSOCK);
 858
 859         /* Parameter checking */
 860         if (sock->sk->sk_type != SOCK_RAW) {
 861                 r = -ESOCKTNOSUPPORT;
 862                 goto err;
 863         }
 864
 865         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 866                                &uaddr_len, 0);
 867         if (r)
 868                 goto err;
 869
 870         if (uaddr.sa.sll_family != AF_PACKET) {
 871                 r = -EPFNOSUPPORT;
 872                 goto err;
 873         }
 874         return sock;
 875 err:
 876         sockfd_put(sock);
 877         return ERR_PTR(r);
 878 }
 879
 880 static struct socket *get_tap_socket(int fd)
 881 {
 882         struct file *file = fget(fd);
 883         struct socket *sock;
 884
 885         if (!file)
 886                 return ERR_PTR(-EBADF);
 887         sock = tun_get_socket(file);
 888         if (!IS_ERR(sock))
 889                 return sock;
 890         sock = macvtap_get_socket(file);
 891         if (IS_ERR(sock))
 892                 fput(file);
 893         return sock;
 894 }
 895
 896 static struct socket *get_socket(int fd)
 897 {
 898         struct socket *sock;
 899
 900         /* special case to disable backend */
 901         if (fd == -1)
 902                 return NULL;
 903         sock = get_raw_socket(fd);
 904         if (!IS_ERR(sock))
 905                 return sock;
 906         sock = get_tap_socket(fd);
 907         if (!IS_ERR(sock))
 908                 return sock;
 909         return ERR_PTR(-ENOTSOCK);
 910 }
 911
 912 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 913 {
 914         struct socket *sock, *oldsock;
 915         struct vhost_virtqueue *vq;
 916         struct vhost_net_virtqueue *nvq;
 917         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 918         int r;
 919
 920         mutex_lock(&n->dev.mutex);
 921         r = vhost_dev_check_owner(&n->dev);
 922         if (r)
 923                 goto err;
 924
 925         if (index >= VHOST_NET_VQ_MAX) {
 926                 r = -ENOBUFS;
 927                 goto err;
 928         }
 929         vq = &n->vqs[index].vq;
 930         nvq = &n->vqs[index];
 931         mutex_lock(&vq->mutex);
 932
 933         /* Verify that ring has been setup correctly. */
 934         if (!vhost_vq_access_ok(vq)) {
 935                 r = -EFAULT;
 936                 goto err_vq;
 937         }
 938         sock = get_socket(fd);
 939         if (IS_ERR(sock)) {
 940                 r = PTR_ERR(sock);
 941                 goto err_vq;
 942         }
 943
 944         /* start polling new socket */
 945         oldsock = vq->private_data;
 946         if (sock != oldsock) {
 947                 ubufs = vhost_net_ubuf_alloc(vq,
 948                                              sock && vhost_sock_zcopy(sock));
 949                 if (IS_ERR(ubufs)) {
 950                         r = PTR_ERR(ubufs);
 951                         goto err_ubufs;
 952                 }
 953
 954                 vhost_net_disable_vq(n, vq);
 955                 vq->private_data = sock;
 956                 r = vhost_init_used(vq);
 957                 if (r)
 958                         goto err_used;
 959                 r = vhost_net_enable_vq(n, vq);
 960                 if (r)
 961                         goto err_used;
 962
 963                 oldubufs = nvq->ubufs;
 964                 nvq->ubufs = ubufs;
 965
 966                 n->tx_packets = 0;
 967                 n->tx_zcopy_err = 0;
 968                 n->tx_flush = false;
 969         }
 970
 971         mutex_unlock(&vq->mutex);
 972
 973         if (oldubufs) {
 974                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 975                 mutex_lock(&vq->mutex);
 976                 vhost_zerocopy_signal_used(n, vq);
 977                 mutex_unlock(&vq->mutex);
 978         }
 979
 980         if (oldsock) {
 981                 vhost_net_flush_vq(n, index);
 982                 sockfd_put(oldsock);
 983         }
 984
 985         mutex_unlock(&n->dev.mutex);
 986         return 0;
 987
 988 err_used:
 989         vq->private_data = oldsock;
 990         vhost_net_enable_vq(n, vq);
 991         if (ubufs)
 992                 vhost_net_ubuf_put_wait_and_free(ubufs);
 993 err_ubufs:
 994         sockfd_put(sock);
 995 err_vq:
 996         mutex_unlock(&vq->mutex);
 997 err:
 998         mutex_unlock(&n->dev.mutex);
 999         return r;
1000 }
1001
1002 static long vhost_net_reset_owner(struct vhost_net *n)
1003 {
1004         struct socket *tx_sock = NULL;
1005         struct socket *rx_sock = NULL;
1006         long err;
1007         struct vhost_memory *memory;
1008
1009         mutex_lock(&n->dev.mutex);
1010         err = vhost_dev_check_owner(&n->dev);
1011         if (err)
1012                 goto done;
1013         memory = vhost_dev_reset_owner_prepare();
1014         if (!memory) {
1015                 err = -ENOMEM;
1016                 goto done;
1017         }
1018         vhost_net_stop(n, &tx_sock, &rx_sock);
1019         vhost_net_flush(n);
1020         vhost_dev_reset_owner(&n->dev, memory);
1021         vhost_net_vq_reset(n);
1022 done:
1023         mutex_unlock(&n->dev.mutex);
1024         if (tx_sock)
1025                 sockfd_put(tx_sock);
1026         if (rx_sock)
1027                 sockfd_put(rx_sock);
1028         return err;
1029 }
1030
1031 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1032 {
1033         size_t vhost_hlen, sock_hlen, hdr_len;
1034         int i;
1035
1036         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1037                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1038                         sizeof(struct virtio_net_hdr);
1039         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1040                 /* vhost provides vnet_hdr */
1041                 vhost_hlen = hdr_len;
1042                 sock_hlen = 0;
1043         } else {
1044                 /* socket provides vnet_hdr */
1045                 vhost_hlen = 0;
1046                 sock_hlen = hdr_len;
1047         }
1048         mutex_lock(&n->dev.mutex);
1049         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1050             !vhost_log_access_ok(&n->dev)) {
1051                 mutex_unlock(&n->dev.mutex);
1052                 return -EFAULT;
1053         }
1054         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1055                 mutex_lock(&n->vqs[i].vq.mutex);
1056                 n->vqs[i].vq.acked_features = features;
1057                 n->vqs[i].vhost_hlen = vhost_hlen;
1058                 n->vqs[i].sock_hlen = sock_hlen;
1059                 mutex_unlock(&n->vqs[i].vq.mutex);
1060         }
1061         mutex_unlock(&n->dev.mutex);
1062         return 0;
1063 }
1064
1065 static long vhost_net_set_owner(struct vhost_net *n)
1066 {
1067         int r;
1068
1069         mutex_lock(&n->dev.mutex);
1070         if (vhost_dev_has_owner(&n->dev)) {
1071                 r = -EBUSY;
1072                 goto out;
1073         }
1074         r = vhost_net_set_ubuf_info(n);
1075         if (r)
1076                 goto out;
1077         r = vhost_dev_set_owner(&n->dev);
1078         if (r)
1079                 vhost_net_clear_ubuf_info(n);
1080         vhost_net_flush(n);
1081 out:
1082         mutex_unlock(&n->dev.mutex);
1083         return r;
1084 }
1085
1086 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1087                             unsigned long arg)
1088 {
1089         struct vhost_net *n = f->private_data;
1090         void __user *argp = (void __user *)arg;
1091         u64 __user *featurep = argp;
1092         struct vhost_vring_file backend;
1093         u64 features;
1094         int r;
1095
1096         switch (ioctl) {
1097         case VHOST_NET_SET_BACKEND:
1098                 if (copy_from_user(&backend, argp, sizeof backend))
1099                         return -EFAULT;
1100                 return vhost_net_set_backend(n, backend.index, backend.fd);
1101         case VHOST_GET_FEATURES:
1102                 features = VHOST_NET_FEATURES;
1103                 if (copy_to_user(featurep, &features, sizeof features))
1104                         return -EFAULT;
1105                 return 0;
1106         case VHOST_SET_FEATURES:
1107                 if (copy_from_user(&features, featurep, sizeof features))
1108                         return -EFAULT;
1109                 if (features & ~VHOST_NET_FEATURES)
1110                         return -EOPNOTSUPP;
1111                 return vhost_net_set_features(n, features);
1112         case VHOST_RESET_OWNER:
1113                 return vhost_net_reset_owner(n);
1114         case VHOST_SET_OWNER:
1115                 return vhost_net_set_owner(n);
1116         default:
1117                 mutex_lock(&n->dev.mutex);
1118                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1119                 if (r == -ENOIOCTLCMD)
1120                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1121                 else
1122                         vhost_net_flush(n);
1123                 mutex_unlock(&n->dev.mutex);
1124                 return r;
1125         }
1126 }
1127
1128 #ifdef CONFIG_COMPAT
1129 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1130                                    unsigned long arg)
1131 {
1132         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1133 }
1134 #endif
1135
1136 static const struct file_operations vhost_net_fops = {
1137         .owner          = THIS_MODULE,
1138         .release        = vhost_net_release,
1139         .unlocked_ioctl = vhost_net_ioctl,
1140 #ifdef CONFIG_COMPAT
1141         .compat_ioctl   = vhost_net_compat_ioctl,
1142 #endif
1143         .open           = vhost_net_open,
1144         .llseek         = noop_llseek,
1145 };
1146
1147 static struct miscdevice vhost_net_misc = {
1148         .minor = VHOST_NET_MINOR,
1149         .name = "vhost-net",
1150         .fops = &vhost_net_fops,
1151 };
1152
1153 static int vhost_net_init(void)
1154 {
1155         if (experimental_zcopytx)
1156                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1157         return misc_register(&vhost_net_misc);
1158 }
1159 module_init(vhost_net_init);
1160
1161 static void vhost_net_exit(void)
1162 {
1163         misc_deregister(&vhost_net_misc);
1164 }
1165 module_exit(vhost_net_exit);
1166
1167 MODULE_VERSION("0.0.1");
1168 MODULE_LICENSE("GPL v2");
1169 MODULE_AUTHOR("Michael S. Tsirkin");
1170 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1171 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1172 MODULE_ALIAS("devname:vhost-net");