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
  21 #include <linux/net.h>
  22 #include <linux/if_packet.h>
  23 #include <linux/if_arp.h>
  24 #include <linux/if_tun.h>
  25 #include <linux/if_macvlan.h>
  26 #include <linux/if_vlan.h>
  27
  28 #include <net/sock.h>
  29
  30 #include "vhost.h"
  31
  32 static int experimental_zcopytx = 1;
  33 module_param(experimental_zcopytx, int, 0444);
  34 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  35                                        " 1 -Enable; 0 - Disable");
  36
  37 /* Max number of bytes transferred before requeueing the job.
  38  * Using this limit prevents one virtqueue from starving others. */
  39 #define VHOST_NET_WEIGHT 0x80000
  40
  41 /* MAX number of TX used buffers for outstanding zerocopy */
  42 #define VHOST_MAX_PEND 128
  43 #define VHOST_GOODCOPY_LEN 256
  44
  45 /*
  46  * For transmit, used buffer len is unused; we override it to track buffer
  47  * status internally; used for zerocopy tx only.
  48  */
  49 /* Lower device DMA failed */
  50 #define VHOST_DMA_FAILED_LEN    3
  51 /* Lower device DMA done */
  52 #define VHOST_DMA_DONE_LEN      2
  53 /* Lower device DMA in progress */
  54 #define VHOST_DMA_IN_PROGRESS   1
  55 /* Buffer unused */
  56 #define VHOST_DMA_CLEAR_LEN     0
  57
  58 #define VHOST_DMA_IS_DONE(len) ((len) >= VHOST_DMA_DONE_LEN)
  59
  60 enum {
  61         VHOST_NET_FEATURES = VHOST_FEATURES |
  62                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  63                          (1ULL << VIRTIO_NET_F_MRG_RXBUF),
  64 };
  65
  66 enum {
  67         VHOST_NET_VQ_RX = 0,
  68         VHOST_NET_VQ_TX = 1,
  69         VHOST_NET_VQ_MAX = 2,
  70 };
  71
  72 struct vhost_net_ubuf_ref {
  73         struct kref kref;
  74         wait_queue_head_t wait;
  75         struct vhost_virtqueue *vq;
  76 };
  77
  78 struct vhost_net_virtqueue {
  79         struct vhost_virtqueue vq;
  80         /* hdr is used to store the virtio header.
  81          * Since each iovec has >= 1 byte length, we never need more than
  82          * header length entries to store the header. */
  83         struct iovec hdr[sizeof(struct virtio_net_hdr_mrg_rxbuf)];
  84         size_t vhost_hlen;
  85         size_t sock_hlen;
  86         /* vhost zerocopy support fields below: */
  87         /* last used idx for outstanding DMA zerocopy buffers */
  88         int upend_idx;
  89         /* first used idx for DMA done zerocopy buffers */
  90         int done_idx;
  91         /* an array of userspace buffers info */
  92         struct ubuf_info *ubuf_info;
  93         /* Reference counting for outstanding ubufs.
  94          * Protected by vq mutex. Writers must also take device mutex. */
  95         struct vhost_net_ubuf_ref *ubufs;
  96 };
  97
  98 struct vhost_net {
  99         struct vhost_dev dev;
 100         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 101         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 102         /* Number of TX recently submitted.
 103          * Protected by tx vq lock. */
 104         unsigned tx_packets;
 105         /* Number of times zerocopy TX recently failed.
 106          * Protected by tx vq lock. */
 107         unsigned tx_zcopy_err;
 108         /* Flush in progress. Protected by tx vq lock. */
 109         bool tx_flush;
 110 };
 111
 112 static unsigned vhost_net_zcopy_mask __read_mostly;
 113
 114 static void vhost_net_enable_zcopy(int vq)
 115 {
 116         vhost_net_zcopy_mask |= 0x1 << vq;
 117 }
 118
 119 static void vhost_net_zerocopy_done_signal(struct kref *kref)
 120 {
 121         struct vhost_net_ubuf_ref *ubufs;
 122
 123         ubufs = container_of(kref, struct vhost_net_ubuf_ref, kref);
 124         wake_up(&ubufs->wait);
 125 }
 126
 127 static struct vhost_net_ubuf_ref *
 128 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 129 {
 130         struct vhost_net_ubuf_ref *ubufs;
 131         /* No zero copy backend? Nothing to count. */
 132         if (!zcopy)
 133                 return NULL;
 134         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 135         if (!ubufs)
 136                 return ERR_PTR(-ENOMEM);
 137         kref_init(&ubufs->kref);
 138         init_waitqueue_head(&ubufs->wait);
 139         ubufs->vq = vq;
 140         return ubufs;
 141 }
 142
 143 static void vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 144 {
 145         kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
 146 }
 147
 148 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 149 {
 150         kref_put(&ubufs->kref, vhost_net_zerocopy_done_signal);
 151         wait_event(ubufs->wait, !atomic_read(&ubufs->kref.refcount));
 152 }
 153
 154 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 155 {
 156         vhost_net_ubuf_put_and_wait(ubufs);
 157         kfree(ubufs);
 158 }
 159
 160 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 161 {
 162         int i;
 163
 164         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 165                 kfree(n->vqs[i].ubuf_info);
 166                 n->vqs[i].ubuf_info = NULL;
 167         }
 168 }
 169
 170 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 171 {
 172         bool zcopy;
 173         int i;
 174
 175         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 176                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 177                 if (!zcopy)
 178                         continue;
 179                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 180                                               UIO_MAXIOV, GFP_KERNEL);
 181                 if  (!n->vqs[i].ubuf_info)
 182                         goto err;
 183         }
 184         return 0;
 185
 186 err:
 187         vhost_net_clear_ubuf_info(n);
 188         return -ENOMEM;
 189 }
 190
 191 static void vhost_net_vq_reset(struct vhost_net *n)
 192 {
 193         int i;
 194
 195         vhost_net_clear_ubuf_info(n);
 196
 197         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 198                 n->vqs[i].done_idx = 0;
 199                 n->vqs[i].upend_idx = 0;
 200                 n->vqs[i].ubufs = NULL;
 201                 n->vqs[i].vhost_hlen = 0;
 202                 n->vqs[i].sock_hlen = 0;
 203         }
 204
 205 }
 206
 207 static void vhost_net_tx_packet(struct vhost_net *net)
 208 {
 209         ++net->tx_packets;
 210         if (net->tx_packets < 1024)
 211                 return;
 212         net->tx_packets = 0;
 213         net->tx_zcopy_err = 0;
 214 }
 215
 216 static void vhost_net_tx_err(struct vhost_net *net)
 217 {
 218         ++net->tx_zcopy_err;
 219 }
 220
 221 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 222 {
 223         /* TX flush waits for outstanding DMAs to be done.
 224          * Don't start new DMAs.
 225          */
 226         return !net->tx_flush &&
 227                 net->tx_packets / 64 >= net->tx_zcopy_err;
 228 }
 229
 230 static bool vhost_sock_zcopy(struct socket *sock)
 231 {
 232         return unlikely(experimental_zcopytx) &&
 233                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 234 }
 235
 236 /* Pop first len bytes from iovec. Return number of segments used. */
 237 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 238                           size_t len, int iov_count)
 239 {
 240         int seg = 0;
 241         size_t size;
 242
 243         while (len && seg < iov_count) {
 244                 size = min(from->iov_len, len);
 245                 to->iov_base = from->iov_base;
 246                 to->iov_len = size;
 247                 from->iov_len -= size;
 248                 from->iov_base += size;
 249                 len -= size;
 250                 ++from;
 251                 ++to;
 252                 ++seg;
 253         }
 254         return seg;
 255 }
 256 /* Copy iovec entries for len bytes from iovec. */
 257 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 258                            size_t len, int iovcount)
 259 {
 260         int seg = 0;
 261         size_t size;
 262
 263         while (len && seg < iovcount) {
 264                 size = min(from->iov_len, len);
 265                 to->iov_base = from->iov_base;
 266                 to->iov_len = size;
 267                 len -= size;
 268                 ++from;
 269                 ++to;
 270                 ++seg;
 271         }
 272 }
 273
 274 /* In case of DMA done not in order in lower device driver for some reason.
 275  * upend_idx is used to track end of used idx, done_idx is used to track head
 276  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 277  * guest used idx.
 278  */
 279 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 280                                        struct vhost_virtqueue *vq)
 281 {
 282         struct vhost_net_virtqueue *nvq =
 283                 container_of(vq, struct vhost_net_virtqueue, vq);
 284         int i, add;
 285         int j = 0;
 286
 287         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 288                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 289                         vhost_net_tx_err(net);
 290                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 291                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 292                         ++j;
 293                 } else
 294                         break;
 295         }
 296         while (j) {
 297                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 298                 vhost_add_used_and_signal_n(vq->dev, vq,
 299                                             &vq->heads[nvq->done_idx], add);
 300                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 301                 j -= add;
 302         }
 303 }
 304
 305 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 306 {
 307         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 308         struct vhost_virtqueue *vq = ubufs->vq;
 309         int cnt = atomic_read(&ubufs->kref.refcount);
 310
 311         /* set len to mark this desc buffers done DMA */
 312         vq->heads[ubuf->desc].len = success ?
 313                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 314         vhost_net_ubuf_put(ubufs);
 315
 316         /*
 317          * Trigger polling thread if guest stopped submitting new buffers:
 318          * in this case, the refcount after decrement will eventually reach 1
 319          * so here it is 2.
 320          * We also trigger polling periodically after each 16 packets
 321          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 322          * less than 10% of times).
 323          */
 324         if (cnt <= 2 || !(cnt % 16))
 325                 vhost_poll_queue(&vq->poll);
 326 }
 327
 328 /* Expects to be always run from workqueue - which acts as
 329  * read-size critical section for our kind of RCU. */
 330 static void handle_tx(struct vhost_net *net)
 331 {
 332         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 333         struct vhost_virtqueue *vq = &nvq->vq;
 334         unsigned out, in, s;
 335         int head;
 336         struct msghdr msg = {
 337                 .msg_name = NULL,
 338                 .msg_namelen = 0,
 339                 .msg_control = NULL,
 340                 .msg_controllen = 0,
 341                 .msg_iov = vq->iov,
 342                 .msg_flags = MSG_DONTWAIT,
 343         };
 344         size_t len, total_len = 0;
 345         int err;
 346         size_t hdr_size;
 347         struct socket *sock;
 348         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 349         bool zcopy, zcopy_used;
 350
 351         mutex_lock(&vq->mutex);
 352         sock = vq->private_data;
 353         if (!sock)
 354                 goto out;
 355
 356         vhost_disable_notify(&net->dev, vq);
 357
 358         hdr_size = nvq->vhost_hlen;
 359         zcopy = nvq->ubufs;
 360
 361         for (;;) {
 362                 /* Release DMAs done buffers first */
 363                 if (zcopy)
 364                         vhost_zerocopy_signal_used(net, vq);
 365
 366                 /* If more outstanding DMAs, queue the work.
 367                  * Handle upend_idx wrap around
 368                  */
 369                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
 370                               % UIO_MAXIOV == nvq->done_idx))
 371                         break;
 372
 373                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 374                                          ARRAY_SIZE(vq->iov),
 375                                          &out, &in,
 376                                          NULL, NULL);
 377                 /* On error, stop handling until the next kick. */
 378                 if (unlikely(head < 0))
 379                         break;
 380                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 381                 if (head == vq->num) {
 382                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 383                                 vhost_disable_notify(&net->dev, vq);
 384                                 continue;
 385                         }
 386                         break;
 387                 }
 388                 if (in) {
 389                         vq_err(vq, "Unexpected descriptor format for TX: "
 390                                "out %d, int %d\n", out, in);
 391                         break;
 392                 }
 393                 /* Skip header. TODO: support TSO. */
 394                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 395                 msg.msg_iovlen = out;
 396                 len = iov_length(vq->iov, out);
 397                 /* Sanity check */
 398                 if (!len) {
 399                         vq_err(vq, "Unexpected header len for TX: "
 400                                "%zd expected %zd\n",
 401                                iov_length(nvq->hdr, s), hdr_size);
 402                         break;
 403                 }
 404
 405                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 406                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 407                                       nvq->done_idx
 408                                    && vhost_net_tx_select_zcopy(net);
 409
 410                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 411                 if (zcopy_used) {
 412                         struct ubuf_info *ubuf;
 413                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 414
 415                         vq->heads[nvq->upend_idx].id = head;
 416                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 417                         ubuf->callback = vhost_zerocopy_callback;
 418                         ubuf->ctx = nvq->ubufs;
 419                         ubuf->desc = nvq->upend_idx;
 420                         msg.msg_control = ubuf;
 421                         msg.msg_controllen = sizeof(ubuf);
 422                         ubufs = nvq->ubufs;
 423                         kref_get(&ubufs->kref);
 424                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 425                 } else {
 426                         msg.msg_control = NULL;
 427                         ubufs = NULL;
 428                 }
 429                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 430                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 431                 if (unlikely(err < 0)) {
 432                         if (zcopy_used) {
 433                                 vhost_net_ubuf_put(ubufs);
 434                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 435                                         % UIO_MAXIOV;
 436                         }
 437                         vhost_discard_vq_desc(vq, 1);
 438                         break;
 439                 }
 440                 if (err != len)
 441                         pr_debug("Truncated TX packet: "
 442                                  " len %d != %zd\n", err, len);
 443                 if (!zcopy_used)
 444                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 445                 else
 446                         vhost_zerocopy_signal_used(net, vq);
 447                 total_len += len;
 448                 vhost_net_tx_packet(net);
 449                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 450                         vhost_poll_queue(&vq->poll);
 451                         break;
 452                 }
 453         }
 454 out:
 455         mutex_unlock(&vq->mutex);
 456 }
 457
 458 static int peek_head_len(struct sock *sk)
 459 {
 460         struct sk_buff *head;
 461         int len = 0;
 462         unsigned long flags;
 463
 464         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 465         head = skb_peek(&sk->sk_receive_queue);
 466         if (likely(head)) {
 467                 len = head->len;
 468                 if (vlan_tx_tag_present(head))
 469                         len += VLAN_HLEN;
 470         }
 471
 472         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 473         return len;
 474 }
 475
 476 /* This is a multi-buffer version of vhost_get_desc, that works if
 477  *      vq has read descriptors only.
 478  * @vq          - the relevant virtqueue
 479  * @datalen     - data length we'll be reading
 480  * @iovcount    - returned count of io vectors we fill
 481  * @log         - vhost log
 482  * @log_num     - log offset
 483  * @quota       - headcount quota, 1 for big buffer
 484  *      returns number of buffer heads allocated, negative on error
 485  */
 486 static int get_rx_bufs(struct vhost_virtqueue *vq,
 487                        struct vring_used_elem *heads,
 488                        int datalen,
 489                        unsigned *iovcount,
 490                        struct vhost_log *log,
 491                        unsigned *log_num,
 492                        unsigned int quota)
 493 {
 494         unsigned int out, in;
 495         int seg = 0;
 496         int headcount = 0;
 497         unsigned d;
 498         int r, nlogs = 0;
 499
 500         while (datalen > 0 && headcount < quota) {
 501                 if (unlikely(seg >= UIO_MAXIOV)) {
 502                         r = -ENOBUFS;
 503                         goto err;
 504                 }
 505                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 506                                       ARRAY_SIZE(vq->iov) - seg, &out,
 507                                       &in, log, log_num);
 508                 if (d == vq->num) {
 509                         r = 0;
 510                         goto err;
 511                 }
 512                 if (unlikely(out || in <= 0)) {
 513                         vq_err(vq, "unexpected descriptor format for RX: "
 514                                 "out %d, in %d\n", out, in);
 515                         r = -EINVAL;
 516                         goto err;
 517                 }
 518                 if (unlikely(log)) {
 519                         nlogs += *log_num;
 520                         log += *log_num;
 521                 }
 522                 heads[headcount].id = d;
 523                 heads[headcount].len = iov_length(vq->iov + seg, in);
 524                 datalen -= heads[headcount].len;
 525                 ++headcount;
 526                 seg += in;
 527         }
 528         heads[headcount - 1].len += datalen;
 529         *iovcount = seg;
 530         if (unlikely(log))
 531                 *log_num = nlogs;
 532         return headcount;
 533 err:
 534         vhost_discard_vq_desc(vq, headcount);
 535         return r;
 536 }
 537
 538 /* Expects to be always run from workqueue - which acts as
 539  * read-size critical section for our kind of RCU. */
 540 static void handle_rx(struct vhost_net *net)
 541 {
 542         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 543         struct vhost_virtqueue *vq = &nvq->vq;
 544         unsigned uninitialized_var(in), log;
 545         struct vhost_log *vq_log;
 546         struct msghdr msg = {
 547                 .msg_name = NULL,
 548                 .msg_namelen = 0,
 549                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 550                 .msg_controllen = 0,
 551                 .msg_iov = vq->iov,
 552                 .msg_flags = MSG_DONTWAIT,
 553         };
 554         struct virtio_net_hdr_mrg_rxbuf hdr = {
 555                 .hdr.flags = 0,
 556                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 557         };
 558         size_t total_len = 0;
 559         int err, mergeable;
 560         s16 headcount;
 561         size_t vhost_hlen, sock_hlen;
 562         size_t vhost_len, sock_len;
 563         struct socket *sock;
 564
 565         mutex_lock(&vq->mutex);
 566         sock = vq->private_data;
 567         if (!sock)
 568                 goto out;
 569         vhost_disable_notify(&net->dev, vq);
 570
 571         vhost_hlen = nvq->vhost_hlen;
 572         sock_hlen = nvq->sock_hlen;
 573
 574         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 575                 vq->log : NULL;
 576         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 577
 578         while ((sock_len = peek_head_len(sock->sk))) {
 579                 sock_len += sock_hlen;
 580                 vhost_len = sock_len + vhost_hlen;
 581                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 582                                         &in, vq_log, &log,
 583                                         likely(mergeable) ? UIO_MAXIOV : 1);
 584                 /* On error, stop handling until the next kick. */
 585                 if (unlikely(headcount < 0))
 586                         break;
 587                 /* OK, now we need to know about added descriptors. */
 588                 if (!headcount) {
 589                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 590                                 /* They have slipped one in as we were
 591                                  * doing that: check again. */
 592                                 vhost_disable_notify(&net->dev, vq);
 593                                 continue;
 594                         }
 595                         /* Nothing new?  Wait for eventfd to tell us
 596                          * they refilled. */
 597                         break;
 598                 }
 599                 /* We don't need to be notified again. */
 600                 if (unlikely((vhost_hlen)))
 601                         /* Skip header. TODO: support TSO. */
 602                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 603                 else
 604                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 605                          * needed because recvmsg can modify msg_iov. */
 606                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 607                 msg.msg_iovlen = in;
 608                 err = sock->ops->recvmsg(NULL, sock, &msg,
 609                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 610                 /* Userspace might have consumed the packet meanwhile:
 611                  * it's not supposed to do this usually, but might be hard
 612                  * to prevent. Discard data we got (if any) and keep going. */
 613                 if (unlikely(err != sock_len)) {
 614                         pr_debug("Discarded rx packet: "
 615                                  " len %d, expected %zd\n", err, sock_len);
 616                         vhost_discard_vq_desc(vq, headcount);
 617                         continue;
 618                 }
 619                 if (unlikely(vhost_hlen) &&
 620                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 621                                       vhost_hlen)) {
 622                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 623                                vq->iov->iov_base);
 624                         break;
 625                 }
 626                 /* TODO: Should check and handle checksum. */
 627                 if (likely(mergeable) &&
 628                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 629                                       offsetof(typeof(hdr), num_buffers),
 630                                       sizeof hdr.num_buffers)) {
 631                         vq_err(vq, "Failed num_buffers write");
 632                         vhost_discard_vq_desc(vq, headcount);
 633                         break;
 634                 }
 635                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 636                                             headcount);
 637                 if (unlikely(vq_log))
 638                         vhost_log_write(vq, vq_log, log, vhost_len);
 639                 total_len += vhost_len;
 640                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 641                         vhost_poll_queue(&vq->poll);
 642                         break;
 643                 }
 644         }
 645 out:
 646         mutex_unlock(&vq->mutex);
 647 }
 648
 649 static void handle_tx_kick(struct vhost_work *work)
 650 {
 651         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 652                                                   poll.work);
 653         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 654
 655         handle_tx(net);
 656 }
 657
 658 static void handle_rx_kick(struct vhost_work *work)
 659 {
 660         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 661                                                   poll.work);
 662         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 663
 664         handle_rx(net);
 665 }
 666
 667 static void handle_tx_net(struct vhost_work *work)
 668 {
 669         struct vhost_net *net = container_of(work, struct vhost_net,
 670                                              poll[VHOST_NET_VQ_TX].work);
 671         handle_tx(net);
 672 }
 673
 674 static void handle_rx_net(struct vhost_work *work)
 675 {
 676         struct vhost_net *net = container_of(work, struct vhost_net,
 677                                              poll[VHOST_NET_VQ_RX].work);
 678         handle_rx(net);
 679 }
 680
 681 static int vhost_net_open(struct inode *inode, struct file *f)
 682 {
 683         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 684         struct vhost_dev *dev;
 685         struct vhost_virtqueue **vqs;
 686         int i;
 687
 688         if (!n)
 689                 return -ENOMEM;
 690         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 691         if (!vqs) {
 692                 kfree(n);
 693                 return -ENOMEM;
 694         }
 695
 696         dev = &n->dev;
 697         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 698         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 699         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 700         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 701         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 702                 n->vqs[i].ubufs = NULL;
 703                 n->vqs[i].ubuf_info = NULL;
 704                 n->vqs[i].upend_idx = 0;
 705                 n->vqs[i].done_idx = 0;
 706                 n->vqs[i].vhost_hlen = 0;
 707                 n->vqs[i].sock_hlen = 0;
 708         }
 709         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 710
 711         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 712         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 713
 714         f->private_data = n;
 715
 716         return 0;
 717 }
 718
 719 static void vhost_net_disable_vq(struct vhost_net *n,
 720                                  struct vhost_virtqueue *vq)
 721 {
 722         struct vhost_net_virtqueue *nvq =
 723                 container_of(vq, struct vhost_net_virtqueue, vq);
 724         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 725         if (!vq->private_data)
 726                 return;
 727         vhost_poll_stop(poll);
 728 }
 729
 730 static int vhost_net_enable_vq(struct vhost_net *n,
 731                                 struct vhost_virtqueue *vq)
 732 {
 733         struct vhost_net_virtqueue *nvq =
 734                 container_of(vq, struct vhost_net_virtqueue, vq);
 735         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 736         struct socket *sock;
 737
 738         sock = vq->private_data;
 739         if (!sock)
 740                 return 0;
 741
 742         return vhost_poll_start(poll, sock->file);
 743 }
 744
 745 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 746                                         struct vhost_virtqueue *vq)
 747 {
 748         struct socket *sock;
 749
 750         mutex_lock(&vq->mutex);
 751         sock = vq->private_data;
 752         vhost_net_disable_vq(n, vq);
 753         vq->private_data = NULL;
 754         mutex_unlock(&vq->mutex);
 755         return sock;
 756 }
 757
 758 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 759                            struct socket **rx_sock)
 760 {
 761         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 762         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 763 }
 764
 765 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 766 {
 767         vhost_poll_flush(n->poll + index);
 768         vhost_poll_flush(&n->vqs[index].vq.poll);
 769 }
 770
 771 static void vhost_net_flush(struct vhost_net *n)
 772 {
 773         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 774         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 775         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 776                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 777                 n->tx_flush = true;
 778                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 779                 /* Wait for all lower device DMAs done. */
 780                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 781                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 782                 n->tx_flush = false;
 783                 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
 784                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 785         }
 786 }
 787
 788 static int vhost_net_release(struct inode *inode, struct file *f)
 789 {
 790         struct vhost_net *n = f->private_data;
 791         struct socket *tx_sock;
 792         struct socket *rx_sock;
 793
 794         vhost_net_stop(n, &tx_sock, &rx_sock);
 795         vhost_net_flush(n);
 796         vhost_dev_stop(&n->dev);
 797         vhost_dev_cleanup(&n->dev, false);
 798         vhost_net_vq_reset(n);
 799         if (tx_sock)
 800                 fput(tx_sock->file);
 801         if (rx_sock)
 802                 fput(rx_sock->file);
 803         /* We do an extra flush before freeing memory,
 804          * since jobs can re-queue themselves. */
 805         vhost_net_flush(n);
 806         kfree(n->dev.vqs);
 807         kfree(n);
 808         return 0;
 809 }
 810
 811 static struct socket *get_raw_socket(int fd)
 812 {
 813         struct {
 814                 struct sockaddr_ll sa;
 815                 char  buf[MAX_ADDR_LEN];
 816         } uaddr;
 817         int uaddr_len = sizeof uaddr, r;
 818         struct socket *sock = sockfd_lookup(fd, &r);
 819
 820         if (!sock)
 821                 return ERR_PTR(-ENOTSOCK);
 822
 823         /* Parameter checking */
 824         if (sock->sk->sk_type != SOCK_RAW) {
 825                 r = -ESOCKTNOSUPPORT;
 826                 goto err;
 827         }
 828
 829         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 830                                &uaddr_len, 0);
 831         if (r)
 832                 goto err;
 833
 834         if (uaddr.sa.sll_family != AF_PACKET) {
 835                 r = -EPFNOSUPPORT;
 836                 goto err;
 837         }
 838         return sock;
 839 err:
 840         fput(sock->file);
 841         return ERR_PTR(r);
 842 }
 843
 844 static struct socket *get_tap_socket(int fd)
 845 {
 846         struct file *file = fget(fd);
 847         struct socket *sock;
 848
 849         if (!file)
 850                 return ERR_PTR(-EBADF);
 851         sock = tun_get_socket(file);
 852         if (!IS_ERR(sock))
 853                 return sock;
 854         sock = macvtap_get_socket(file);
 855         if (IS_ERR(sock))
 856                 fput(file);
 857         return sock;
 858 }
 859
 860 static struct socket *get_socket(int fd)
 861 {
 862         struct socket *sock;
 863
 864         /* special case to disable backend */
 865         if (fd == -1)
 866                 return NULL;
 867         sock = get_raw_socket(fd);
 868         if (!IS_ERR(sock))
 869                 return sock;
 870         sock = get_tap_socket(fd);
 871         if (!IS_ERR(sock))
 872                 return sock;
 873         return ERR_PTR(-ENOTSOCK);
 874 }
 875
 876 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 877 {
 878         struct socket *sock, *oldsock;
 879         struct vhost_virtqueue *vq;
 880         struct vhost_net_virtqueue *nvq;
 881         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 882         int r;
 883
 884         mutex_lock(&n->dev.mutex);
 885         r = vhost_dev_check_owner(&n->dev);
 886         if (r)
 887                 goto err;
 888
 889         if (index >= VHOST_NET_VQ_MAX) {
 890                 r = -ENOBUFS;
 891                 goto err;
 892         }
 893         vq = &n->vqs[index].vq;
 894         nvq = &n->vqs[index];
 895         mutex_lock(&vq->mutex);
 896
 897         /* Verify that ring has been setup correctly. */
 898         if (!vhost_vq_access_ok(vq)) {
 899                 r = -EFAULT;
 900                 goto err_vq;
 901         }
 902         sock = get_socket(fd);
 903         if (IS_ERR(sock)) {
 904                 r = PTR_ERR(sock);
 905                 goto err_vq;
 906         }
 907
 908         /* start polling new socket */
 909         oldsock = vq->private_data;
 910         if (sock != oldsock) {
 911                 ubufs = vhost_net_ubuf_alloc(vq,
 912                                              sock && vhost_sock_zcopy(sock));
 913                 if (IS_ERR(ubufs)) {
 914                         r = PTR_ERR(ubufs);
 915                         goto err_ubufs;
 916                 }
 917
 918                 vhost_net_disable_vq(n, vq);
 919                 vq->private_data = sock;
 920                 r = vhost_init_used(vq);
 921                 if (r)
 922                         goto err_used;
 923                 r = vhost_net_enable_vq(n, vq);
 924                 if (r)
 925                         goto err_used;
 926
 927                 oldubufs = nvq->ubufs;
 928                 nvq->ubufs = ubufs;
 929
 930                 n->tx_packets = 0;
 931                 n->tx_zcopy_err = 0;
 932                 n->tx_flush = false;
 933         }
 934
 935         mutex_unlock(&vq->mutex);
 936
 937         if (oldubufs) {
 938                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 939                 mutex_lock(&vq->mutex);
 940                 vhost_zerocopy_signal_used(n, vq);
 941                 mutex_unlock(&vq->mutex);
 942         }
 943
 944         if (oldsock) {
 945                 vhost_net_flush_vq(n, index);
 946                 fput(oldsock->file);
 947         }
 948
 949         mutex_unlock(&n->dev.mutex);
 950         return 0;
 951
 952 err_used:
 953         vq->private_data = oldsock;
 954         vhost_net_enable_vq(n, vq);
 955         if (ubufs)
 956                 vhost_net_ubuf_put_wait_and_free(ubufs);
 957 err_ubufs:
 958         fput(sock->file);
 959 err_vq:
 960         mutex_unlock(&vq->mutex);
 961 err:
 962         mutex_unlock(&n->dev.mutex);
 963         return r;
 964 }
 965
 966 static long vhost_net_reset_owner(struct vhost_net *n)
 967 {
 968         struct socket *tx_sock = NULL;
 969         struct socket *rx_sock = NULL;
 970         long err;
 971         struct vhost_memory *memory;
 972
 973         mutex_lock(&n->dev.mutex);
 974         err = vhost_dev_check_owner(&n->dev);
 975         if (err)
 976                 goto done;
 977         memory = vhost_dev_reset_owner_prepare();
 978         if (!memory) {
 979                 err = -ENOMEM;
 980                 goto done;
 981         }
 982         vhost_net_stop(n, &tx_sock, &rx_sock);
 983         vhost_net_flush(n);
 984         vhost_dev_reset_owner(&n->dev, memory);
 985         vhost_net_vq_reset(n);
 986 done:
 987         mutex_unlock(&n->dev.mutex);
 988         if (tx_sock)
 989                 fput(tx_sock->file);
 990         if (rx_sock)
 991                 fput(rx_sock->file);
 992         return err;
 993 }
 994
 995 static int vhost_net_set_features(struct vhost_net *n, u64 features)
 996 {
 997         size_t vhost_hlen, sock_hlen, hdr_len;
 998         int i;
 999
1000         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1001                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1002                         sizeof(struct virtio_net_hdr);
1003         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1004                 /* vhost provides vnet_hdr */
1005                 vhost_hlen = hdr_len;
1006                 sock_hlen = 0;
1007         } else {
1008                 /* socket provides vnet_hdr */
1009                 vhost_hlen = 0;
1010                 sock_hlen = hdr_len;
1011         }
1012         mutex_lock(&n->dev.mutex);
1013         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1014             !vhost_log_access_ok(&n->dev)) {
1015                 mutex_unlock(&n->dev.mutex);
1016                 return -EFAULT;
1017         }
1018         n->dev.acked_features = features;
1019         smp_wmb();
1020         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1021                 mutex_lock(&n->vqs[i].vq.mutex);
1022                 n->vqs[i].vhost_hlen = vhost_hlen;
1023                 n->vqs[i].sock_hlen = sock_hlen;
1024                 mutex_unlock(&n->vqs[i].vq.mutex);
1025         }
1026         vhost_net_flush(n);
1027         mutex_unlock(&n->dev.mutex);
1028         return 0;
1029 }
1030
1031 static long vhost_net_set_owner(struct vhost_net *n)
1032 {
1033         int r;
1034
1035         mutex_lock(&n->dev.mutex);
1036         if (vhost_dev_has_owner(&n->dev)) {
1037                 r = -EBUSY;
1038                 goto out;
1039         }
1040         r = vhost_net_set_ubuf_info(n);
1041         if (r)
1042                 goto out;
1043         r = vhost_dev_set_owner(&n->dev);
1044         if (r)
1045                 vhost_net_clear_ubuf_info(n);
1046         vhost_net_flush(n);
1047 out:
1048         mutex_unlock(&n->dev.mutex);
1049         return r;
1050 }
1051
1052 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1053                             unsigned long arg)
1054 {
1055         struct vhost_net *n = f->private_data;
1056         void __user *argp = (void __user *)arg;
1057         u64 __user *featurep = argp;
1058         struct vhost_vring_file backend;
1059         u64 features;
1060         int r;
1061
1062         switch (ioctl) {
1063         case VHOST_NET_SET_BACKEND:
1064                 if (copy_from_user(&backend, argp, sizeof backend))
1065                         return -EFAULT;
1066                 return vhost_net_set_backend(n, backend.index, backend.fd);
1067         case VHOST_GET_FEATURES:
1068                 features = VHOST_NET_FEATURES;
1069                 if (copy_to_user(featurep, &features, sizeof features))
1070                         return -EFAULT;
1071                 return 0;
1072         case VHOST_SET_FEATURES:
1073                 if (copy_from_user(&features, featurep, sizeof features))
1074                         return -EFAULT;
1075                 if (features & ~VHOST_NET_FEATURES)
1076                         return -EOPNOTSUPP;
1077                 return vhost_net_set_features(n, features);
1078         case VHOST_RESET_OWNER:
1079                 return vhost_net_reset_owner(n);
1080         case VHOST_SET_OWNER:
1081                 return vhost_net_set_owner(n);
1082         default:
1083                 mutex_lock(&n->dev.mutex);
1084                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1085                 if (r == -ENOIOCTLCMD)
1086                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1087                 else
1088                         vhost_net_flush(n);
1089                 mutex_unlock(&n->dev.mutex);
1090                 return r;
1091         }
1092 }
1093
1094 #ifdef CONFIG_COMPAT
1095 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1096                                    unsigned long arg)
1097 {
1098         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1099 }
1100 #endif
1101
1102 static const struct file_operations vhost_net_fops = {
1103         .owner          = THIS_MODULE,
1104         .release        = vhost_net_release,
1105         .unlocked_ioctl = vhost_net_ioctl,
1106 #ifdef CONFIG_COMPAT
1107         .compat_ioctl   = vhost_net_compat_ioctl,
1108 #endif
1109         .open           = vhost_net_open,
1110         .llseek         = noop_llseek,
1111 };
1112
1113 static struct miscdevice vhost_net_misc = {
1114         .minor = VHOST_NET_MINOR,
1115         .name = "vhost-net",
1116         .fops = &vhost_net_fops,
1117 };
1118
1119 static int vhost_net_init(void)
1120 {
1121         if (experimental_zcopytx)
1122                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1123         return misc_register(&vhost_net_misc);
1124 }
1125 module_init(vhost_net_init);
1126
1127 static void vhost_net_exit(void)
1128 {
1129         misc_deregister(&vhost_net_misc);
1130 }
1131 module_exit(vhost_net_exit);
1132
1133 MODULE_VERSION("0.0.1");
1134 MODULE_LICENSE("GPL v2");
1135 MODULE_AUTHOR("Michael S. Tsirkin");
1136 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1137 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1138 MODULE_ALIAS("devname:vhost-net");