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                 }
 440                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 441                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 442                 if (unlikely(err < 0)) {
 443                         if (zcopy_used) {
 444                                 if (ubufs)
 445                                         vhost_net_ubuf_put(ubufs);
 446                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 447                                         % UIO_MAXIOV;
 448                         }
 449                         vhost_discard_vq_desc(vq, 1);
 450                         break;
 451                 }
 452                 if (err != len)
 453                         pr_debug("Truncated TX packet: "
 454                                  " len %d != %zd\n", err, len);
 455                 if (!zcopy_used)
 456                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 457                 else
 458                         vhost_zerocopy_signal_used(net, vq);
 459                 total_len += len;
 460                 vhost_net_tx_packet(net);
 461                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 462                         vhost_poll_queue(&vq->poll);
 463                         break;
 464                 }
 465         }
 466
 467         mutex_unlock(&vq->mutex);
 468 }
 469
 470 static int peek_head_len(struct sock *sk)
 471 {
 472         struct sk_buff *head;
 473         int len = 0;
 474         unsigned long flags;
 475
 476         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 477         head = skb_peek(&sk->sk_receive_queue);
 478         if (likely(head)) {
 479                 len = head->len;
 480                 if (vlan_tx_tag_present(head))
 481                         len += VLAN_HLEN;
 482         }
 483
 484         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 485         return len;
 486 }
 487
 488 /* This is a multi-buffer version of vhost_get_desc, that works if
 489  *      vq has read descriptors only.
 490  * @vq          - the relevant virtqueue
 491  * @datalen     - data length we'll be reading
 492  * @iovcount    - returned count of io vectors we fill
 493  * @log         - vhost log
 494  * @log_num     - log offset
 495  * @quota       - headcount quota, 1 for big buffer
 496  *      returns number of buffer heads allocated, negative on error
 497  */
 498 static int get_rx_bufs(struct vhost_virtqueue *vq,
 499                        struct vring_used_elem *heads,
 500                        int datalen,
 501                        unsigned *iovcount,
 502                        struct vhost_log *log,
 503                        unsigned *log_num,
 504                        unsigned int quota)
 505 {
 506         unsigned int out, in;
 507         int seg = 0;
 508         int headcount = 0;
 509         unsigned d;
 510         int r, nlogs = 0;
 511
 512         while (datalen > 0 && headcount < quota) {
 513                 if (unlikely(seg >= UIO_MAXIOV)) {
 514                         r = -ENOBUFS;
 515                         goto err;
 516                 }
 517                 d = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 518                                       ARRAY_SIZE(vq->iov) - seg, &out,
 519                                       &in, log, log_num);
 520                 if (d == vq->num) {
 521                         r = 0;
 522                         goto err;
 523                 }
 524                 if (unlikely(out || in <= 0)) {
 525                         vq_err(vq, "unexpected descriptor format for RX: "
 526                                 "out %d, in %d\n", out, in);
 527                         r = -EINVAL;
 528                         goto err;
 529                 }
 530                 if (unlikely(log)) {
 531                         nlogs += *log_num;
 532                         log += *log_num;
 533                 }
 534                 heads[headcount].id = d;
 535                 heads[headcount].len = iov_length(vq->iov + seg, in);
 536                 datalen -= heads[headcount].len;
 537                 ++headcount;
 538                 seg += in;
 539         }
 540         heads[headcount - 1].len += datalen;
 541         *iovcount = seg;
 542         if (unlikely(log))
 543                 *log_num = nlogs;
 544         return headcount;
 545 err:
 546         vhost_discard_vq_desc(vq, headcount);
 547         return r;
 548 }
 549
 550 /* Expects to be always run from workqueue - which acts as
 551  * read-size critical section for our kind of RCU. */
 552 static void handle_rx(struct vhost_net *net)
 553 {
 554         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 555         struct vhost_virtqueue *vq = &nvq->vq;
 556         unsigned uninitialized_var(in), log;
 557         struct vhost_log *vq_log;
 558         struct msghdr msg = {
 559                 .msg_name = NULL,
 560                 .msg_namelen = 0,
 561                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 562                 .msg_controllen = 0,
 563                 .msg_iov = vq->iov,
 564                 .msg_flags = MSG_DONTWAIT,
 565         };
 566         struct virtio_net_hdr_mrg_rxbuf hdr = {
 567                 .hdr.flags = 0,
 568                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 569         };
 570         size_t total_len = 0;
 571         int err, mergeable;
 572         s16 headcount;
 573         size_t vhost_hlen, sock_hlen;
 574         size_t vhost_len, sock_len;
 575         /* TODO: check that we are running from vhost_worker? */
 576         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 577
 578         if (!sock)
 579                 return;
 580
 581         mutex_lock(&vq->mutex);
 582         vhost_disable_notify(&net->dev, vq);
 583         vhost_hlen = nvq->vhost_hlen;
 584         sock_hlen = nvq->sock_hlen;
 585
 586         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 587                 vq->log : NULL;
 588         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 589
 590         while ((sock_len = peek_head_len(sock->sk))) {
 591                 sock_len += sock_hlen;
 592                 vhost_len = sock_len + vhost_hlen;
 593                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 594                                         &in, vq_log, &log,
 595                                         likely(mergeable) ? UIO_MAXIOV : 1);
 596                 /* On error, stop handling until the next kick. */
 597                 if (unlikely(headcount < 0))
 598                         break;
 599                 /* OK, now we need to know about added descriptors. */
 600                 if (!headcount) {
 601                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 602                                 /* They have slipped one in as we were
 603                                  * doing that: check again. */
 604                                 vhost_disable_notify(&net->dev, vq);
 605                                 continue;
 606                         }
 607                         /* Nothing new?  Wait for eventfd to tell us
 608                          * they refilled. */
 609                         break;
 610                 }
 611                 /* We don't need to be notified again. */
 612                 if (unlikely((vhost_hlen)))
 613                         /* Skip header. TODO: support TSO. */
 614                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 615                 else
 616                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 617                          * needed because recvmsg can modify msg_iov. */
 618                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 619                 msg.msg_iovlen = in;
 620                 err = sock->ops->recvmsg(NULL, sock, &msg,
 621                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 622                 /* Userspace might have consumed the packet meanwhile:
 623                  * it's not supposed to do this usually, but might be hard
 624                  * to prevent. Discard data we got (if any) and keep going. */
 625                 if (unlikely(err != sock_len)) {
 626                         pr_debug("Discarded rx packet: "
 627                                  " len %d, expected %zd\n", err, sock_len);
 628                         vhost_discard_vq_desc(vq, headcount);
 629                         continue;
 630                 }
 631                 if (unlikely(vhost_hlen) &&
 632                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 633                                       vhost_hlen)) {
 634                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 635                                vq->iov->iov_base);
 636                         break;
 637                 }
 638                 /* TODO: Should check and handle checksum. */
 639                 if (likely(mergeable) &&
 640                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 641                                       offsetof(typeof(hdr), num_buffers),
 642                                       sizeof hdr.num_buffers)) {
 643                         vq_err(vq, "Failed num_buffers write");
 644                         vhost_discard_vq_desc(vq, headcount);
 645                         break;
 646                 }
 647                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 648                                             headcount);
 649                 if (unlikely(vq_log))
 650                         vhost_log_write(vq, vq_log, log, vhost_len);
 651                 total_len += vhost_len;
 652                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 653                         vhost_poll_queue(&vq->poll);
 654                         break;
 655                 }
 656         }
 657
 658         mutex_unlock(&vq->mutex);
 659 }
 660
 661 static void handle_tx_kick(struct vhost_work *work)
 662 {
 663         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 664                                                   poll.work);
 665         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 666
 667         handle_tx(net);
 668 }
 669
 670 static void handle_rx_kick(struct vhost_work *work)
 671 {
 672         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 673                                                   poll.work);
 674         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 675
 676         handle_rx(net);
 677 }
 678
 679 static void handle_tx_net(struct vhost_work *work)
 680 {
 681         struct vhost_net *net = container_of(work, struct vhost_net,
 682                                              poll[VHOST_NET_VQ_TX].work);
 683         handle_tx(net);
 684 }
 685
 686 static void handle_rx_net(struct vhost_work *work)
 687 {
 688         struct vhost_net *net = container_of(work, struct vhost_net,
 689                                              poll[VHOST_NET_VQ_RX].work);
 690         handle_rx(net);
 691 }
 692
 693 static int vhost_net_open(struct inode *inode, struct file *f)
 694 {
 695         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 696         struct vhost_dev *dev;
 697         struct vhost_virtqueue **vqs;
 698         int r, i;
 699
 700         if (!n)
 701                 return -ENOMEM;
 702         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 703         if (!vqs) {
 704                 kfree(n);
 705                 return -ENOMEM;
 706         }
 707
 708         dev = &n->dev;
 709         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 710         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 711         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 712         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 713         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 714                 n->vqs[i].ubufs = NULL;
 715                 n->vqs[i].ubuf_info = NULL;
 716                 n->vqs[i].upend_idx = 0;
 717                 n->vqs[i].done_idx = 0;
 718                 n->vqs[i].vhost_hlen = 0;
 719                 n->vqs[i].sock_hlen = 0;
 720         }
 721         r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 722         if (r < 0) {
 723                 kfree(n);
 724                 kfree(vqs);
 725                 return r;
 726         }
 727
 728         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 729         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 730
 731         f->private_data = n;
 732
 733         return 0;
 734 }
 735
 736 static void vhost_net_disable_vq(struct vhost_net *n,
 737                                  struct vhost_virtqueue *vq)
 738 {
 739         struct vhost_net_virtqueue *nvq =
 740                 container_of(vq, struct vhost_net_virtqueue, vq);
 741         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 742         if (!vq->private_data)
 743                 return;
 744         vhost_poll_stop(poll);
 745 }
 746
 747 static int vhost_net_enable_vq(struct vhost_net *n,
 748                                 struct vhost_virtqueue *vq)
 749 {
 750         struct vhost_net_virtqueue *nvq =
 751                 container_of(vq, struct vhost_net_virtqueue, vq);
 752         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 753         struct socket *sock;
 754
 755         sock = rcu_dereference_protected(vq->private_data,
 756                                          lockdep_is_held(&vq->mutex));
 757         if (!sock)
 758                 return 0;
 759
 760         return vhost_poll_start(poll, sock->file);
 761 }
 762
 763 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 764                                         struct vhost_virtqueue *vq)
 765 {
 766         struct socket *sock;
 767
 768         mutex_lock(&vq->mutex);
 769         sock = rcu_dereference_protected(vq->private_data,
 770                                          lockdep_is_held(&vq->mutex));
 771         vhost_net_disable_vq(n, vq);
 772         rcu_assign_pointer(vq->private_data, NULL);
 773         mutex_unlock(&vq->mutex);
 774         return sock;
 775 }
 776
 777 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 778                            struct socket **rx_sock)
 779 {
 780         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 781         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 782 }
 783
 784 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 785 {
 786         vhost_poll_flush(n->poll + index);
 787         vhost_poll_flush(&n->vqs[index].vq.poll);
 788 }
 789
 790 static void vhost_net_flush(struct vhost_net *n)
 791 {
 792         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 793         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 794         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 795                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 796                 n->tx_flush = true;
 797                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 798                 /* Wait for all lower device DMAs done. */
 799                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 800                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 801                 n->tx_flush = false;
 802                 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
 803                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 804         }
 805 }
 806
 807 static int vhost_net_release(struct inode *inode, struct file *f)
 808 {
 809         struct vhost_net *n = f->private_data;
 810         struct socket *tx_sock;
 811         struct socket *rx_sock;
 812
 813         vhost_net_stop(n, &tx_sock, &rx_sock);
 814         vhost_net_flush(n);
 815         vhost_dev_stop(&n->dev);
 816         vhost_dev_cleanup(&n->dev, false);
 817         vhost_net_vq_reset(n);
 818         if (tx_sock)
 819                 fput(tx_sock->file);
 820         if (rx_sock)
 821                 fput(rx_sock->file);
 822         /* We do an extra flush before freeing memory,
 823          * since jobs can re-queue themselves. */
 824         vhost_net_flush(n);
 825         kfree(n->dev.vqs);
 826         kfree(n);
 827         return 0;
 828 }
 829
 830 static struct socket *get_raw_socket(int fd)
 831 {
 832         struct {
 833                 struct sockaddr_ll sa;
 834                 char  buf[MAX_ADDR_LEN];
 835         } uaddr;
 836         int uaddr_len = sizeof uaddr, r;
 837         struct socket *sock = sockfd_lookup(fd, &r);
 838
 839         if (!sock)
 840                 return ERR_PTR(-ENOTSOCK);
 841
 842         /* Parameter checking */
 843         if (sock->sk->sk_type != SOCK_RAW) {
 844                 r = -ESOCKTNOSUPPORT;
 845                 goto err;
 846         }
 847
 848         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 849                                &uaddr_len, 0);
 850         if (r)
 851                 goto err;
 852
 853         if (uaddr.sa.sll_family != AF_PACKET) {
 854                 r = -EPFNOSUPPORT;
 855                 goto err;
 856         }
 857         return sock;
 858 err:
 859         fput(sock->file);
 860         return ERR_PTR(r);
 861 }
 862
 863 static struct socket *get_tap_socket(int fd)
 864 {
 865         struct file *file = fget(fd);
 866         struct socket *sock;
 867
 868         if (!file)
 869                 return ERR_PTR(-EBADF);
 870         sock = tun_get_socket(file);
 871         if (!IS_ERR(sock))
 872                 return sock;
 873         sock = macvtap_get_socket(file);
 874         if (IS_ERR(sock))
 875                 fput(file);
 876         return sock;
 877 }
 878
 879 static struct socket *get_socket(int fd)
 880 {
 881         struct socket *sock;
 882
 883         /* special case to disable backend */
 884         if (fd == -1)
 885                 return NULL;
 886         sock = get_raw_socket(fd);
 887         if (!IS_ERR(sock))
 888                 return sock;
 889         sock = get_tap_socket(fd);
 890         if (!IS_ERR(sock))
 891                 return sock;
 892         return ERR_PTR(-ENOTSOCK);
 893 }
 894
 895 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 896 {
 897         struct socket *sock, *oldsock;
 898         struct vhost_virtqueue *vq;
 899         struct vhost_net_virtqueue *nvq;
 900         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 901         int r;
 902
 903         mutex_lock(&n->dev.mutex);
 904         r = vhost_dev_check_owner(&n->dev);
 905         if (r)
 906                 goto err;
 907
 908         if (index >= VHOST_NET_VQ_MAX) {
 909                 r = -ENOBUFS;
 910                 goto err;
 911         }
 912         vq = &n->vqs[index].vq;
 913         nvq = &n->vqs[index];
 914         mutex_lock(&vq->mutex);
 915
 916         /* Verify that ring has been setup correctly. */
 917         if (!vhost_vq_access_ok(vq)) {
 918                 r = -EFAULT;
 919                 goto err_vq;
 920         }
 921         sock = get_socket(fd);
 922         if (IS_ERR(sock)) {
 923                 r = PTR_ERR(sock);
 924                 goto err_vq;
 925         }
 926
 927         /* start polling new socket */
 928         oldsock = rcu_dereference_protected(vq->private_data,
 929                                             lockdep_is_held(&vq->mutex));
 930         if (sock != oldsock) {
 931                 ubufs = vhost_net_ubuf_alloc(vq,
 932                                              sock && vhost_sock_zcopy(sock));
 933                 if (IS_ERR(ubufs)) {
 934                         r = PTR_ERR(ubufs);
 935                         goto err_ubufs;
 936                 }
 937
 938                 vhost_net_disable_vq(n, vq);
 939                 rcu_assign_pointer(vq->private_data, sock);
 940                 r = vhost_init_used(vq);
 941                 if (r)
 942                         goto err_used;
 943                 r = vhost_net_enable_vq(n, vq);
 944                 if (r)
 945                         goto err_used;
 946
 947                 oldubufs = nvq->ubufs;
 948                 nvq->ubufs = ubufs;
 949
 950                 n->tx_packets = 0;
 951                 n->tx_zcopy_err = 0;
 952                 n->tx_flush = false;
 953         }
 954
 955         mutex_unlock(&vq->mutex);
 956
 957         if (oldubufs) {
 958                 vhost_net_ubuf_put_and_wait(oldubufs);
 959                 mutex_lock(&vq->mutex);
 960                 vhost_zerocopy_signal_used(n, vq);
 961                 mutex_unlock(&vq->mutex);
 962         }
 963
 964         if (oldsock) {
 965                 vhost_net_flush_vq(n, index);
 966                 fput(oldsock->file);
 967         }
 968
 969         mutex_unlock(&n->dev.mutex);
 970         return 0;
 971
 972 err_used:
 973         rcu_assign_pointer(vq->private_data, oldsock);
 974         vhost_net_enable_vq(n, vq);
 975         if (ubufs)
 976                 vhost_net_ubuf_put_and_wait(ubufs);
 977 err_ubufs:
 978         fput(sock->file);
 979 err_vq:
 980         mutex_unlock(&vq->mutex);
 981 err:
 982         mutex_unlock(&n->dev.mutex);
 983         return r;
 984 }
 985
 986 static long vhost_net_reset_owner(struct vhost_net *n)
 987 {
 988         struct socket *tx_sock = NULL;
 989         struct socket *rx_sock = NULL;
 990         long err;
 991         struct vhost_memory *memory;
 992
 993         mutex_lock(&n->dev.mutex);
 994         err = vhost_dev_check_owner(&n->dev);
 995         if (err)
 996                 goto done;
 997         memory = vhost_dev_reset_owner_prepare();
 998         if (!memory) {
 999                 err = -ENOMEM;
1000                 goto done;
1001         }
1002         vhost_net_stop(n, &tx_sock, &rx_sock);
1003         vhost_net_flush(n);
1004         vhost_dev_reset_owner(&n->dev, memory);
1005         vhost_net_vq_reset(n);
1006 done:
1007         mutex_unlock(&n->dev.mutex);
1008         if (tx_sock)
1009                 fput(tx_sock->file);
1010         if (rx_sock)
1011                 fput(rx_sock->file);
1012         return err;
1013 }
1014
1015 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1016 {
1017         size_t vhost_hlen, sock_hlen, hdr_len;
1018         int i;
1019
1020         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1021                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1022                         sizeof(struct virtio_net_hdr);
1023         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1024                 /* vhost provides vnet_hdr */
1025                 vhost_hlen = hdr_len;
1026                 sock_hlen = 0;
1027         } else {
1028                 /* socket provides vnet_hdr */
1029                 vhost_hlen = 0;
1030                 sock_hlen = hdr_len;
1031         }
1032         mutex_lock(&n->dev.mutex);
1033         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1034             !vhost_log_access_ok(&n->dev)) {
1035                 mutex_unlock(&n->dev.mutex);
1036                 return -EFAULT;
1037         }
1038         n->dev.acked_features = features;
1039         smp_wmb();
1040         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1041                 mutex_lock(&n->vqs[i].vq.mutex);
1042                 n->vqs[i].vhost_hlen = vhost_hlen;
1043                 n->vqs[i].sock_hlen = sock_hlen;
1044                 mutex_unlock(&n->vqs[i].vq.mutex);
1045         }
1046         vhost_net_flush(n);
1047         mutex_unlock(&n->dev.mutex);
1048         return 0;
1049 }
1050
1051 static long vhost_net_set_owner(struct vhost_net *n)
1052 {
1053         int r;
1054
1055         mutex_lock(&n->dev.mutex);
1056         r = vhost_net_set_ubuf_info(n);
1057         if (r)
1058                 goto out;
1059         r = vhost_dev_set_owner(&n->dev);
1060         if (r)
1061                 vhost_net_clear_ubuf_info(n);
1062         vhost_net_flush(n);
1063 out:
1064         mutex_unlock(&n->dev.mutex);
1065         return r;
1066 }
1067
1068 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1069                             unsigned long arg)
1070 {
1071         struct vhost_net *n = f->private_data;
1072         void __user *argp = (void __user *)arg;
1073         u64 __user *featurep = argp;
1074         struct vhost_vring_file backend;
1075         u64 features;
1076         int r;
1077
1078         switch (ioctl) {
1079         case VHOST_NET_SET_BACKEND:
1080                 if (copy_from_user(&backend, argp, sizeof backend))
1081                         return -EFAULT;
1082                 return vhost_net_set_backend(n, backend.index, backend.fd);
1083         case VHOST_GET_FEATURES:
1084                 features = VHOST_NET_FEATURES;
1085                 if (copy_to_user(featurep, &features, sizeof features))
1086                         return -EFAULT;
1087                 return 0;
1088         case VHOST_SET_FEATURES:
1089                 if (copy_from_user(&features, featurep, sizeof features))
1090                         return -EFAULT;
1091                 if (features & ~VHOST_NET_FEATURES)
1092                         return -EOPNOTSUPP;
1093                 return vhost_net_set_features(n, features);
1094         case VHOST_RESET_OWNER:
1095                 return vhost_net_reset_owner(n);
1096         case VHOST_SET_OWNER:
1097                 return vhost_net_set_owner(n);
1098         default:
1099                 mutex_lock(&n->dev.mutex);
1100                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1101                 if (r == -ENOIOCTLCMD)
1102                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1103                 else
1104                         vhost_net_flush(n);
1105                 mutex_unlock(&n->dev.mutex);
1106                 return r;
1107         }
1108 }
1109
1110 #ifdef CONFIG_COMPAT
1111 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1112                                    unsigned long arg)
1113 {
1114         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1115 }
1116 #endif
1117
1118 static const struct file_operations vhost_net_fops = {
1119         .owner          = THIS_MODULE,
1120         .release        = vhost_net_release,
1121         .unlocked_ioctl = vhost_net_ioctl,
1122 #ifdef CONFIG_COMPAT
1123         .compat_ioctl   = vhost_net_compat_ioctl,
1124 #endif
1125         .open           = vhost_net_open,
1126         .llseek         = noop_llseek,
1127 };
1128
1129 static struct miscdevice vhost_net_misc = {
1130         .minor = VHOST_NET_MINOR,
1131         .name = "vhost-net",
1132         .fops = &vhost_net_fops,
1133 };
1134
1135 static int vhost_net_init(void)
1136 {
1137         if (experimental_zcopytx)
1138                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1139         return misc_register(&vhost_net_misc);
1140 }
1141 module_init(vhost_net_init);
1142
1143 static void vhost_net_exit(void)
1144 {
1145         misc_deregister(&vhost_net_misc);
1146 }
1147 module_exit(vhost_net_exit);
1148
1149 MODULE_VERSION("0.0.1");
1150 MODULE_LICENSE("GPL v2");
1151 MODULE_AUTHOR("Michael S. Tsirkin");
1152 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1153 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1154 MODULE_ALIAS("devname:vhost-net");