Pileus Git - ~andy/linux/blob - drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/rcupdate.h>
  19 #include <linux/file.h>
  20 #include <linux/slab.h>
  21
  22 #include <linux/net.h>
  23 #include <linux/if_packet.h>
  24 #include <linux/if_arp.h>
  25 #include <linux/if_tun.h>
  26 #include <linux/if_macvlan.h>
  27 #include <linux/if_vlan.h>
  28
  29 #include <net/sock.h>
  30
  31 #include "vhost.h"
  32
  33 static int experimental_zcopytx = 1;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  36                                        " 1 -Enable; 0 - Disable");
  37
  38 /* Max number of bytes transferred before requeueing the job.
  39  * Using this limit prevents one virtqueue from starving others. */
  40 #define VHOST_NET_WEIGHT 0x80000
  41
  42 /* MAX number of TX used buffers for outstanding zerocopy */
  43 #define VHOST_MAX_PEND 128
  44 #define VHOST_GOODCOPY_LEN 256
  45
  46 /*
  47  * For transmit, used buffer len is unused; we override it to track buffer
  48  * status internally; used for zerocopy tx only.
  49  */
  50 /* Lower device DMA failed */
  51 #define VHOST_DMA_FAILED_LEN    3
  52 /* Lower device DMA done */
  53 #define VHOST_DMA_DONE_LEN      2
  54 /* Lower device DMA in progress */
  55 #define VHOST_DMA_IN_PROGRESS   1
  56 /* Buffer unused */
  57 #define VHOST_DMA_CLEAR_LEN     0
  58
  59 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
  60
  61 enum {
  62         VHOST_NET_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
 159         bool zcopy;
 160         int i;
 161
 162         for (i = 0; i < n->dev.nvqs; ++i) {
 163                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 164                 if (zcopy)
 165                         kfree(n->vqs[i].ubuf_info);
 166         }
 167 }
 168
 169 int vhost_net_set_ubuf_info(struct vhost_net *n)
 170 {
 171         bool zcopy;
 172         int i;
 173
 174         for (i = 0; i < n->dev.nvqs; ++i) {
 175                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 176                 if (!zcopy)
 177                         continue;
 178                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 179                                               UIO_MAXIOV, GFP_KERNEL);
 180                 if  (!n->vqs[i].ubuf_info)
 181                         goto err;
 182         }
 183         return 0;
 184
 185 err:
 186         while (i--) {
 187                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 188                 if (!zcopy)
 189                         continue;
 190                 kfree(n->vqs[i].ubuf_info);
 191         }
 192         return -ENOMEM;
 193 }
 194
 195 void vhost_net_vq_reset(struct vhost_net *n)
 196 {
 197         int i;
 198
 199         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 200                 n->vqs[i].done_idx = 0;
 201                 n->vqs[i].upend_idx = 0;
 202                 n->vqs[i].ubufs = NULL;
 203                 kfree(n->vqs[i].ubuf_info);
 204                 n->vqs[i].ubuf_info = NULL;
 205                 n->vqs[i].vhost_hlen = 0;
 206                 n->vqs[i].sock_hlen = 0;
 207         }
 208
 209 }
 210
 211 static void vhost_net_tx_packet(struct vhost_net *net)
 212 {
 213         ++net->tx_packets;
 214         if (net->tx_packets < 1024)
 215                 return;
 216         net->tx_packets = 0;
 217         net->tx_zcopy_err = 0;
 218 }
 219
 220 static void vhost_net_tx_err(struct vhost_net *net)
 221 {
 222         ++net->tx_zcopy_err;
 223 }
 224
 225 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 226 {
 227         /* TX flush waits for outstanding DMAs to be done.
 228          * Don't start new DMAs.
 229          */
 230         return !net->tx_flush &&
 231                 net->tx_packets / 64 >= net->tx_zcopy_err;
 232 }
 233
 234 static bool vhost_sock_zcopy(struct socket *sock)
 235 {
 236         return unlikely(experimental_zcopytx) &&
 237                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 238 }
 239
 240 /* Pop first len bytes from iovec. Return number of segments used. */
 241 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 242                           size_t len, int iov_count)
 243 {
 244         int seg = 0;
 245         size_t size;
 246
 247         while (len && seg < iov_count) {
 248                 size = min(from->iov_len, len);
 249                 to->iov_base = from->iov_base;
 250                 to->iov_len = size;
 251                 from->iov_len -= size;
 252                 from->iov_base += size;
 253                 len -= size;
 254                 ++from;
 255                 ++to;
 256                 ++seg;
 257         }
 258         return seg;
 259 }
 260 /* Copy iovec entries for len bytes from iovec. */
 261 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 262                            size_t len, int iovcount)
 263 {
 264         int seg = 0;
 265         size_t size;
 266
 267         while (len && seg < iovcount) {
 268                 size = min(from->iov_len, len);
 269                 to->iov_base = from->iov_base;
 270                 to->iov_len = size;
 271                 len -= size;
 272                 ++from;
 273                 ++to;
 274                 ++seg;
 275         }
 276 }
 277
 278 /* In case of DMA done not in order in lower device driver for some reason.
 279  * upend_idx is used to track end of used idx, done_idx is used to track head
 280  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 281  * guest used idx.
 282  */
 283 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 284                                       struct vhost_virtqueue *vq)
 285 {
 286         struct vhost_net_virtqueue *nvq =
 287                 container_of(vq, struct vhost_net_virtqueue, vq);
 288         int i;
 289         int j = 0;
 290
 291         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 292                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 293                         vhost_net_tx_err(net);
 294                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 295                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 296                         vhost_add_used_and_signal(vq->dev, vq,
 297                                                   vq->heads[i].id, 0);
 298                         ++j;
 299                 } else
 300                         break;
 301         }
 302         if (j)
 303                 nvq->done_idx = i;
 304         return j;
 305 }
 306
 307 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 308 {
 309         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 310         struct vhost_virtqueue *vq = ubufs->vq;
 311         int cnt = atomic_read(&ubufs->kref.refcount);
 312
 313         /*
 314          * Trigger polling thread if guest stopped submitting new buffers:
 315          * in this case, the refcount after decrement will eventually reach 1
 316          * so here it is 2.
 317          * We also trigger polling periodically after each 16 packets
 318          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 319          * less than 10% of times).
 320          */
 321         if (cnt <= 2 || !(cnt % 16))
 322                 vhost_poll_queue(&vq->poll);
 323         /* set len to mark this desc buffers done DMA */
 324         vq->heads[ubuf->desc].len = success ?
 325                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 326         vhost_net_ubuf_put(ubufs);
 327 }
 328
 329 /* Expects to be always run from workqueue - which acts as
 330  * read-size critical section for our kind of RCU. */
 331 static void handle_tx(struct vhost_net *net)
 332 {
 333         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 334         struct vhost_virtqueue *vq = &nvq->vq;
 335         unsigned out, in, s;
 336         int head;
 337         struct msghdr msg = {
 338                 .msg_name = NULL,
 339                 .msg_namelen = 0,
 340                 .msg_control = NULL,
 341                 .msg_controllen = 0,
 342                 .msg_iov = vq->iov,
 343                 .msg_flags = MSG_DONTWAIT,
 344         };
 345         size_t len, total_len = 0;
 346         int err;
 347         size_t hdr_size;
 348         struct socket *sock;
 349         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 350         bool zcopy, zcopy_used;
 351
 352         /* TODO: check that we are running from vhost_worker? */
 353         sock = rcu_dereference_check(vq->private_data, 1);
 354         if (!sock)
 355                 return;
 356
 357         mutex_lock(&vq->mutex);
 358         vhost_disable_notify(&net->dev, vq);
 359
 360         hdr_size = nvq->vhost_hlen;
 361         zcopy = nvq->ubufs;
 362
 363         for (;;) {
 364                 /* Release DMAs done buffers first */
 365                 if (zcopy)
 366                         vhost_zerocopy_signal_used(net, vq);
 367
 368                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 369                                          ARRAY_SIZE(vq->iov),
 370                                          &out, &in,
 371                                          NULL, NULL);
 372                 /* On error, stop handling until the next kick. */
 373                 if (unlikely(head < 0))
 374                         break;
 375                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 376                 if (head == vq->num) {
 377                         int num_pends;
 378
 379                         /* If more outstanding DMAs, queue the work.
 380                          * Handle upend_idx wrap around
 381                          */
 382                         num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
 383                                     (nvq->upend_idx - nvq->done_idx) :
 384                                     (nvq->upend_idx + UIO_MAXIOV -
 385                                      nvq->done_idx);
 386                         if (unlikely(num_pends > VHOST_MAX_PEND))
 387                                 break;
 388                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 389                                 vhost_disable_notify(&net->dev, vq);
 390                                 continue;
 391                         }
 392                         break;
 393                 }
 394                 if (in) {
 395                         vq_err(vq, "Unexpected descriptor format for TX: "
 396                                "out %d, int %d\n", out, in);
 397                         break;
 398                 }
 399                 /* Skip header. TODO: support TSO. */
 400                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 401                 msg.msg_iovlen = out;
 402                 len = iov_length(vq->iov, out);
 403                 /* Sanity check */
 404                 if (!len) {
 405                         vq_err(vq, "Unexpected header len for TX: "
 406                                "%zd expected %zd\n",
 407                                iov_length(nvq->hdr, s), hdr_size);
 408                         break;
 409                 }
 410                 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
 411                                        nvq->upend_idx != nvq->done_idx);
 412
 413                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 414                 if (zcopy_used) {
 415                         vq->heads[nvq->upend_idx].id = head;
 416                         if (!vhost_net_tx_select_zcopy(net) ||
 417                             len < VHOST_GOODCOPY_LEN) {
 418                                 /* copy don't need to wait for DMA done */
 419                                 vq->heads[nvq->upend_idx].len =
 420                                                         VHOST_DMA_DONE_LEN;
 421                                 msg.msg_control = NULL;
 422                                 msg.msg_controllen = 0;
 423                                 ubufs = NULL;
 424                         } else {
 425                                 struct ubuf_info *ubuf;
 426                                 ubuf = nvq->ubuf_info + nvq->upend_idx;
 427
 428                                 vq->heads[nvq->upend_idx].len =
 429                                         VHOST_DMA_IN_PROGRESS;
 430                                 ubuf->callback = vhost_zerocopy_callback;
 431                                 ubuf->ctx = nvq->ubufs;
 432                                 ubuf->desc = nvq->upend_idx;
 433                                 msg.msg_control = ubuf;
 434                                 msg.msg_controllen = sizeof(ubuf);
 435                                 ubufs = nvq->ubufs;
 436                                 kref_get(&ubufs->kref);
 437                         }
 438                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 439                 } else
 440                         msg.msg_control = NULL;
 441                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 442                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 443                 if (unlikely(err < 0)) {
 444                         if (zcopy_used) {
 445                                 if (ubufs)
 446                                         vhost_net_ubuf_put(ubufs);
 447                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 448                                         % UIO_MAXIOV;
 449                         }
 450                         vhost_discard_vq_desc(vq, 1);
 451                         break;
 452                 }
 453                 if (err != len)
 454                         pr_debug("Truncated TX packet: "
 455                                  " len %d != %zd\n", err, len);
 456                 if (!zcopy_used)
 457                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 458                 else
 459                         vhost_zerocopy_signal_used(net, vq);
 460                 total_len += len;
 461                 vhost_net_tx_packet(net);
 462                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 463                         vhost_poll_queue(&vq->poll);
 464                         break;
 465                 }
 466         }
 467
 468         mutex_unlock(&vq->mutex);
 469 }
 470
 471 static int peek_head_len(struct sock *sk)
 472 {
 473         struct sk_buff *head;
 474         int len = 0;
 475         unsigned long flags;
 476
 477         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 478         head = skb_peek(&sk->sk_receive_queue);
 479         if (likely(head)) {
 480                 len = head->len;
 481                 if (vlan_tx_tag_present(head))
 482                         len += VLAN_HLEN;
 483         }
 484
 485         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 486         return len;
 487 }
 488
 489 /* This is a multi-buffer version of vhost_get_desc, that works if
 490  *      vq has read descriptors only.
 491  * @vq          - the relevant virtqueue
 492  * @datalen     - data length we'll be reading
 493  * @iovcount    - returned count of io vectors we fill
 494  * @log         - vhost log
 495  * @log_num     - log offset
 496  * @quota       - headcount quota, 1 for big buffer
 497  *      returns number of buffer heads allocated, negative on error
 498  */
 499 static int get_rx_bufs(struct vhost_virtqueue *vq,
 500                        struct vring_used_elem *heads,
 501                        int datalen,
 502                        unsigned *iovcount,
 503                        struct vhost_log *log,
 504                        unsigned *log_num,
 505                        unsigned int quota)
 506 {
 507         unsigned int out, in;
 508         int seg = 0;
 509         int headcount = 0;
 510         unsigned d;
 511         int r, nlogs = 0;
 512
 513         while (datalen > 0 && headcount < quota) {
 514                 if (unlikely(seg >= UIO_MAXIOV)) {
 515                         r = -ENOBUFS;
 516                         goto err;
 517                 }
 518                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 519                                       ARRAY_SIZE(vq->iov) - seg, &out,
 520                                       &in, log, log_num);
 521                 if (d == vq->num) {
 522                         r = 0;
 523                         goto err;
 524                 }
 525                 if (unlikely(out || in <= 0)) {
 526                         vq_err(vq, "unexpected descriptor format for RX: "
 527                                 "out %d, in %d\n", out, in);
 528                         r = -EINVAL;
 529                         goto err;
 530                 }
 531                 if (unlikely(log)) {
 532                         nlogs += *log_num;
 533                         log += *log_num;
 534                 }
 535                 heads[headcount].id = d;
 536                 heads[headcount].len = iov_length(vq->iov + seg, in);
 537                 datalen -= heads[headcount].len;
 538                 ++headcount;
 539                 seg += in;
 540         }
 541         heads[headcount - 1].len += datalen;
 542         *iovcount = seg;
 543         if (unlikely(log))
 544                 *log_num = nlogs;
 545         return headcount;
 546 err:
 547         vhost_discard_vq_desc(vq, headcount);
 548         return r;
 549 }
 550
 551 /* Expects to be always run from workqueue - which acts as
 552  * read-size critical section for our kind of RCU. */
 553 static void handle_rx(struct vhost_net *net)
 554 {
 555         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 556         struct vhost_virtqueue *vq = &nvq->vq;
 557         unsigned uninitialized_var(in), log;
 558         struct vhost_log *vq_log;
 559         struct msghdr msg = {
 560                 .msg_name = NULL,
 561                 .msg_namelen = 0,
 562                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 563                 .msg_controllen = 0,
 564                 .msg_iov = vq->iov,
 565                 .msg_flags = MSG_DONTWAIT,
 566         };
 567         struct virtio_net_hdr_mrg_rxbuf hdr = {
 568                 .hdr.flags = 0,
 569                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 570         };
 571         size_t total_len = 0;
 572         int err, mergeable;
 573         s16 headcount;
 574         size_t vhost_hlen, sock_hlen;
 575         size_t vhost_len, sock_len;
 576         /* TODO: check that we are running from vhost_worker? */
 577         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 578
 579         if (!sock)
 580                 return;
 581
 582         mutex_lock(&vq->mutex);
 583         vhost_disable_notify(&net->dev, vq);
 584         vhost_hlen = nvq->vhost_hlen;
 585         sock_hlen = nvq->sock_hlen;
 586
 587         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 588                 vq->log : NULL;
 589         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 590
 591         while ((sock_len = peek_head_len(sock->sk))) {
 592                 sock_len += sock_hlen;
 593                 vhost_len = sock_len + vhost_hlen;
 594                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 595                                         &in, vq_log, &log,
 596                                         likely(mergeable) ? UIO_MAXIOV : 1);
 597                 /* On error, stop handling until the next kick. */
 598                 if (unlikely(headcount < 0))
 599                         break;
 600                 /* OK, now we need to know about added descriptors. */
 601                 if (!headcount) {
 602                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 603                                 /* They have slipped one in as we were
 604                                  * doing that: check again. */
 605                                 vhost_disable_notify(&net->dev, vq);
 606                                 continue;
 607                         }
 608                         /* Nothing new?  Wait for eventfd to tell us
 609                          * they refilled. */
 610                         break;
 611                 }
 612                 /* We don't need to be notified again. */
 613                 if (unlikely((vhost_hlen)))
 614                         /* Skip header. TODO: support TSO. */
 615                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 616                 else
 617                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 618                          * needed because recvmsg can modify msg_iov. */
 619                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 620                 msg.msg_iovlen = in;
 621                 err = sock->ops->recvmsg(NULL, sock, &msg,
 622                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 623                 /* Userspace might have consumed the packet meanwhile:
 624                  * it's not supposed to do this usually, but might be hard
 625                  * to prevent. Discard data we got (if any) and keep going. */
 626                 if (unlikely(err != sock_len)) {
 627                         pr_debug("Discarded rx packet: "
 628                                  " len %d, expected %zd\n", err, sock_len);
 629                         vhost_discard_vq_desc(vq, headcount);
 630                         continue;
 631                 }
 632                 if (unlikely(vhost_hlen) &&
 633                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 634                                       vhost_hlen)) {
 635                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 636                                vq->iov->iov_base);
 637                         break;
 638                 }
 639                 /* TODO: Should check and handle checksum. */
 640                 if (likely(mergeable) &&
 641                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 642                                       offsetof(typeof(hdr), num_buffers),
 643                                       sizeof hdr.num_buffers)) {
 644                         vq_err(vq, "Failed num_buffers write");
 645                         vhost_discard_vq_desc(vq, headcount);
 646                         break;
 647                 }
 648                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 649                                             headcount);
 650                 if (unlikely(vq_log))
 651                         vhost_log_write(vq, vq_log, log, vhost_len);
 652                 total_len += vhost_len;
 653                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 654                         vhost_poll_queue(&vq->poll);
 655                         break;
 656                 }
 657         }
 658
 659         mutex_unlock(&vq->mutex);
 660 }
 661
 662 static void handle_tx_kick(struct vhost_work *work)
 663 {
 664         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 665                                                   poll.work);
 666         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 667
 668         handle_tx(net);
 669 }
 670
 671 static void handle_rx_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_rx(net);
 678 }
 679
 680 static void handle_tx_net(struct vhost_work *work)
 681 {
 682         struct vhost_net *net = container_of(work, struct vhost_net,
 683                                              poll[VHOST_NET_VQ_TX].work);
 684         handle_tx(net);
 685 }
 686
 687 static void handle_rx_net(struct vhost_work *work)
 688 {
 689         struct vhost_net *net = container_of(work, struct vhost_net,
 690                                              poll[VHOST_NET_VQ_RX].work);
 691         handle_rx(net);
 692 }
 693
 694 static int vhost_net_open(struct inode *inode, struct file *f)
 695 {
 696         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 697         struct vhost_dev *dev;
 698         struct vhost_virtqueue **vqs;
 699         int r, i;
 700
 701         if (!n)
 702                 return -ENOMEM;
 703         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 704         if (!vqs) {
 705                 kfree(n);
 706                 return -ENOMEM;
 707         }
 708
 709         dev = &n->dev;
 710         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 711         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 712         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 713         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 714         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 715                 n->vqs[i].ubufs = NULL;
 716                 n->vqs[i].ubuf_info = NULL;
 717                 n->vqs[i].upend_idx = 0;
 718                 n->vqs[i].done_idx = 0;
 719                 n->vqs[i].vhost_hlen = 0;
 720                 n->vqs[i].sock_hlen = 0;
 721         }
 722         r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 723         if (r < 0) {
 724                 kfree(n);
 725                 kfree(vqs);
 726                 return r;
 727         }
 728
 729         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 730         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 731
 732         f->private_data = n;
 733
 734         return 0;
 735 }
 736
 737 static void vhost_net_disable_vq(struct vhost_net *n,
 738                                  struct vhost_virtqueue *vq)
 739 {
 740         struct vhost_net_virtqueue *nvq =
 741                 container_of(vq, struct vhost_net_virtqueue, vq);
 742         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 743         if (!vq->private_data)
 744                 return;
 745         vhost_poll_stop(poll);
 746 }
 747
 748 static int vhost_net_enable_vq(struct vhost_net *n,
 749                                 struct vhost_virtqueue *vq)
 750 {
 751         struct vhost_net_virtqueue *nvq =
 752                 container_of(vq, struct vhost_net_virtqueue, vq);
 753         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 754         struct socket *sock;
 755
 756         sock = rcu_dereference_protected(vq->private_data,
 757                                          lockdep_is_held(&vq->mutex));
 758         if (!sock)
 759                 return 0;
 760
 761         return vhost_poll_start(poll, sock->file);
 762 }
 763
 764 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 765                                         struct vhost_virtqueue *vq)
 766 {
 767         struct socket *sock;
 768
 769         mutex_lock(&vq->mutex);
 770         sock = rcu_dereference_protected(vq->private_data,
 771                                          lockdep_is_held(&vq->mutex));
 772         vhost_net_disable_vq(n, vq);
 773         rcu_assign_pointer(vq->private_data, NULL);
 774         mutex_unlock(&vq->mutex);
 775         return sock;
 776 }
 777
 778 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 779                            struct socket **rx_sock)
 780 {
 781         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 782         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 783 }
 784
 785 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 786 {
 787         vhost_poll_flush(n->poll + index);
 788         vhost_poll_flush(&n->vqs[index].vq.poll);
 789 }
 790
 791 static void vhost_net_flush(struct vhost_net *n)
 792 {
 793         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 794         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 795         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 796                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 797                 n->tx_flush = true;
 798                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 799                 /* Wait for all lower device DMAs done. */
 800                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 801                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 802                 n->tx_flush = false;
 803                 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
 804                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 805         }
 806 }
 807
 808 static int vhost_net_release(struct inode *inode, struct file *f)
 809 {
 810         struct vhost_net *n = f->private_data;
 811         struct socket *tx_sock;
 812         struct socket *rx_sock;
 813
 814         vhost_net_stop(n, &tx_sock, &rx_sock);
 815         vhost_net_flush(n);
 816         vhost_dev_stop(&n->dev);
 817         vhost_dev_cleanup(&n->dev, false);
 818         vhost_net_vq_reset(n);
 819         if (tx_sock)
 820                 fput(tx_sock->file);
 821         if (rx_sock)
 822                 fput(rx_sock->file);
 823         /* We do an extra flush before freeing memory,
 824          * since jobs can re-queue themselves. */
 825         vhost_net_flush(n);
 826         kfree(n->dev.vqs);
 827         kfree(n);
 828         return 0;
 829 }
 830
 831 static struct socket *get_raw_socket(int fd)
 832 {
 833         struct {
 834                 struct sockaddr_ll sa;
 835                 char  buf[MAX_ADDR_LEN];
 836         } uaddr;
 837         int uaddr_len = sizeof uaddr, r;
 838         struct socket *sock = sockfd_lookup(fd, &r);
 839
 840         if (!sock)
 841                 return ERR_PTR(-ENOTSOCK);
 842
 843         /* Parameter checking */
 844         if (sock->sk->sk_type != SOCK_RAW) {
 845                 r = -ESOCKTNOSUPPORT;
 846                 goto err;
 847         }
 848
 849         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 850                                &uaddr_len, 0);
 851         if (r)
 852                 goto err;
 853
 854         if (uaddr.sa.sll_family != AF_PACKET) {
 855                 r = -EPFNOSUPPORT;
 856                 goto err;
 857         }
 858         return sock;
 859 err:
 860         fput(sock->file);
 861         return ERR_PTR(r);
 862 }
 863
 864 static struct socket *get_tap_socket(int fd)
 865 {
 866         struct file *file = fget(fd);
 867         struct socket *sock;
 868
 869         if (!file)
 870                 return ERR_PTR(-EBADF);
 871         sock = tun_get_socket(file);
 872         if (!IS_ERR(sock))
 873                 return sock;
 874         sock = macvtap_get_socket(file);
 875         if (IS_ERR(sock))
 876                 fput(file);
 877         return sock;
 878 }
 879
 880 static struct socket *get_socket(int fd)
 881 {
 882         struct socket *sock;
 883
 884         /* special case to disable backend */
 885         if (fd == -1)
 886                 return NULL;
 887         sock = get_raw_socket(fd);
 888         if (!IS_ERR(sock))
 889                 return sock;
 890         sock = get_tap_socket(fd);
 891         if (!IS_ERR(sock))
 892                 return sock;
 893         return ERR_PTR(-ENOTSOCK);
 894 }
 895
 896 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 897 {
 898         struct socket *sock, *oldsock;
 899         struct vhost_virtqueue *vq;
 900         struct vhost_net_virtqueue *nvq;
 901         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 902         int r;
 903
 904         mutex_lock(&n->dev.mutex);
 905         r = vhost_dev_check_owner(&n->dev);
 906         if (r)
 907                 goto err;
 908
 909         if (index >= VHOST_NET_VQ_MAX) {
 910                 r = -ENOBUFS;
 911                 goto err;
 912         }
 913         vq = &n->vqs[index].vq;
 914         nvq = &n->vqs[index];
 915         mutex_lock(&vq->mutex);
 916
 917         /* Verify that ring has been setup correctly. */
 918         if (!vhost_vq_access_ok(vq)) {
 919                 r = -EFAULT;
 920                 goto err_vq;
 921         }
 922         sock = get_socket(fd);
 923         if (IS_ERR(sock)) {
 924                 r = PTR_ERR(sock);
 925                 goto err_vq;
 926         }
 927
 928         /* start polling new socket */
 929         oldsock = rcu_dereference_protected(vq->private_data,
 930                                             lockdep_is_held(&vq->mutex));
 931         if (sock != oldsock) {
 932                 ubufs = vhost_net_ubuf_alloc(vq,
 933                                              sock && vhost_sock_zcopy(sock));
 934                 if (IS_ERR(ubufs)) {
 935                         r = PTR_ERR(ubufs);
 936                         goto err_ubufs;
 937                 }
 938
 939                 vhost_net_disable_vq(n, vq);
 940                 rcu_assign_pointer(vq->private_data, sock);
 941                 r = vhost_init_used(vq);
 942                 if (r)
 943                         goto err_used;
 944                 r = vhost_net_enable_vq(n, vq);
 945                 if (r)
 946                         goto err_used;
 947
 948                 oldubufs = nvq->ubufs;
 949                 nvq->ubufs = ubufs;
 950
 951                 n->tx_packets = 0;
 952                 n->tx_zcopy_err = 0;
 953                 n->tx_flush = false;
 954         }
 955
 956         mutex_unlock(&vq->mutex);
 957
 958         if (oldubufs) {
 959                 vhost_net_ubuf_put_and_wait(oldubufs);
 960                 mutex_lock(&vq->mutex);
 961                 vhost_zerocopy_signal_used(n, vq);
 962                 mutex_unlock(&vq->mutex);
 963         }
 964
 965         if (oldsock) {
 966                 vhost_net_flush_vq(n, index);
 967                 fput(oldsock->file);
 968         }
 969
 970         mutex_unlock(&n->dev.mutex);
 971         return 0;
 972
 973 err_used:
 974         rcu_assign_pointer(vq->private_data, oldsock);
 975         vhost_net_enable_vq(n, vq);
 976         if (ubufs)
 977                 vhost_net_ubuf_put_and_wait(ubufs);
 978 err_ubufs:
 979         fput(sock->file);
 980 err_vq:
 981         mutex_unlock(&vq->mutex);
 982 err:
 983         mutex_unlock(&n->dev.mutex);
 984         return r;
 985 }
 986
 987 static long vhost_net_reset_owner(struct vhost_net *n)
 988 {
 989         struct socket *tx_sock = NULL;
 990         struct socket *rx_sock = NULL;
 991         long err;
 992         struct vhost_memory *memory;
 993
 994         mutex_lock(&n->dev.mutex);
 995         err = vhost_dev_check_owner(&n->dev);
 996         if (err)
 997                 goto done;
 998         memory = vhost_dev_reset_owner_prepare();
 999         if (!memory) {
1000                 err = -ENOMEM;
1001                 goto done;
1002         }
1003         vhost_net_stop(n, &tx_sock, &rx_sock);
1004         vhost_net_flush(n);
1005         vhost_dev_reset_owner(&n->dev, memory);
1006         vhost_net_vq_reset(n);
1007 done:
1008         mutex_unlock(&n->dev.mutex);
1009         if (tx_sock)
1010                 fput(tx_sock->file);
1011         if (rx_sock)
1012                 fput(rx_sock->file);
1013         return err;
1014 }
1015
1016 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1017 {
1018         size_t vhost_hlen, sock_hlen, hdr_len;
1019         int i;
1020
1021         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1022                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1023                         sizeof(struct virtio_net_hdr);
1024         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1025                 /* vhost provides vnet_hdr */
1026                 vhost_hlen = hdr_len;
1027                 sock_hlen = 0;
1028         } else {
1029                 /* socket provides vnet_hdr */
1030                 vhost_hlen = 0;
1031                 sock_hlen = hdr_len;
1032         }
1033         mutex_lock(&n->dev.mutex);
1034         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1035             !vhost_log_access_ok(&n->dev)) {
1036                 mutex_unlock(&n->dev.mutex);
1037                 return -EFAULT;
1038         }
1039         n->dev.acked_features = features;
1040         smp_wmb();
1041         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1042                 mutex_lock(&n->vqs[i].vq.mutex);
1043                 n->vqs[i].vhost_hlen = vhost_hlen;
1044                 n->vqs[i].sock_hlen = sock_hlen;
1045                 mutex_unlock(&n->vqs[i].vq.mutex);
1046         }
1047         vhost_net_flush(n);
1048         mutex_unlock(&n->dev.mutex);
1049         return 0;
1050 }
1051
1052 static long vhost_net_set_owner(struct vhost_net *n)
1053 {
1054         int r;
1055
1056         mutex_lock(&n->dev.mutex);
1057         r = vhost_net_set_ubuf_info(n);
1058         if (r)
1059                 goto out;
1060         r = vhost_dev_set_owner(&n->dev);
1061         if (r)
1062                 vhost_net_clear_ubuf_info(n);
1063         vhost_net_flush(n);
1064 out:
1065         mutex_unlock(&n->dev.mutex);
1066         return r;
1067 }
1068
1069 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1070                             unsigned long arg)
1071 {
1072         struct vhost_net *n = f->private_data;
1073         void __user *argp = (void __user *)arg;
1074         u64 __user *featurep = argp;
1075         struct vhost_vring_file backend;
1076         u64 features;
1077         int r;
1078
1079         switch (ioctl) {
1080         case VHOST_NET_SET_BACKEND:
1081                 if (copy_from_user(&backend, argp, sizeof backend))
1082                         return -EFAULT;
1083                 return vhost_net_set_backend(n, backend.index, backend.fd);
1084         case VHOST_GET_FEATURES:
1085                 features = VHOST_NET_FEATURES;
1086                 if (copy_to_user(featurep, &features, sizeof features))
1087                         return -EFAULT;
1088                 return 0;
1089         case VHOST_SET_FEATURES:
1090                 if (copy_from_user(&features, featurep, sizeof features))
1091                         return -EFAULT;
1092                 if (features & ~VHOST_NET_FEATURES)
1093                         return -EOPNOTSUPP;
1094                 return vhost_net_set_features(n, features);
1095         case VHOST_RESET_OWNER:
1096                 return vhost_net_reset_owner(n);
1097         case VHOST_SET_OWNER:
1098                 return vhost_net_set_owner(n);
1099         default:
1100                 mutex_lock(&n->dev.mutex);
1101                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1102                 if (r == -ENOIOCTLCMD)
1103                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1104                 else
1105                         vhost_net_flush(n);
1106                 mutex_unlock(&n->dev.mutex);
1107                 return r;
1108         }
1109 }
1110
1111 #ifdef CONFIG_COMPAT
1112 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1113                                    unsigned long arg)
1114 {
1115         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1116 }
1117 #endif
1118
1119 static const struct file_operations vhost_net_fops = {
1120         .owner          = THIS_MODULE,
1121         .release        = vhost_net_release,
1122         .unlocked_ioctl = vhost_net_ioctl,
1123 #ifdef CONFIG_COMPAT
1124         .compat_ioctl   = vhost_net_compat_ioctl,
1125 #endif
1126         .open           = vhost_net_open,
1127         .llseek         = noop_llseek,
1128 };
1129
1130 static struct miscdevice vhost_net_misc = {
1131         .minor = VHOST_NET_MINOR,
1132         .name = "vhost-net",
1133         .fops = &vhost_net_fops,
1134 };
1135
1136 static int vhost_net_init(void)
1137 {
1138         if (experimental_zcopytx)
1139                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1140         return misc_register(&vhost_net_misc);
1141 }
1142 module_init(vhost_net_init);
1143
1144 static void vhost_net_exit(void)
1145 {
1146         misc_deregister(&vhost_net_misc);
1147 }
1148 module_exit(vhost_net_exit);
1149
1150 MODULE_VERSION("0.0.1");
1151 MODULE_LICENSE("GPL v2");
1152 MODULE_AUTHOR("Michael S. Tsirkin");
1153 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1154 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1155 MODULE_ALIAS("devname:vhost-net");