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 }
 154
 155 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 156 {
 157         vhost_net_ubuf_put_and_wait(ubufs);
 158         kfree(ubufs);
 159 }
 160
 161 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 162 {
 163         int i;
 164
 165         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 166                 kfree(n->vqs[i].ubuf_info);
 167                 n->vqs[i].ubuf_info = NULL;
 168         }
 169 }
 170
 171 int vhost_net_set_ubuf_info(struct vhost_net *n)
 172 {
 173         bool zcopy;
 174         int i;
 175
 176         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 177                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 178                 if (!zcopy)
 179                         continue;
 180                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 181                                               UIO_MAXIOV, GFP_KERNEL);
 182                 if  (!n->vqs[i].ubuf_info)
 183                         goto err;
 184         }
 185         return 0;
 186
 187 err:
 188         vhost_net_clear_ubuf_info(n);
 189         return -ENOMEM;
 190 }
 191
 192 void vhost_net_vq_reset(struct vhost_net *n)
 193 {
 194         int i;
 195
 196         vhost_net_clear_ubuf_info(n);
 197
 198         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 199                 n->vqs[i].done_idx = 0;
 200                 n->vqs[i].upend_idx = 0;
 201                 n->vqs[i].ubufs = NULL;
 202                 n->vqs[i].vhost_hlen = 0;
 203                 n->vqs[i].sock_hlen = 0;
 204         }
 205
 206 }
 207
 208 static void vhost_net_tx_packet(struct vhost_net *net)
 209 {
 210         ++net->tx_packets;
 211         if (net->tx_packets < 1024)
 212                 return;
 213         net->tx_packets = 0;
 214         net->tx_zcopy_err = 0;
 215 }
 216
 217 static void vhost_net_tx_err(struct vhost_net *net)
 218 {
 219         ++net->tx_zcopy_err;
 220 }
 221
 222 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 223 {
 224         /* TX flush waits for outstanding DMAs to be done.
 225          * Don't start new DMAs.
 226          */
 227         return !net->tx_flush &&
 228                 net->tx_packets / 64 >= net->tx_zcopy_err;
 229 }
 230
 231 static bool vhost_sock_zcopy(struct socket *sock)
 232 {
 233         return unlikely(experimental_zcopytx) &&
 234                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 235 }
 236
 237 /* Pop first len bytes from iovec. Return number of segments used. */
 238 static int move_iovec_hdr(struct iovec *from, struct iovec *to,
 239                           size_t len, int iov_count)
 240 {
 241         int seg = 0;
 242         size_t size;
 243
 244         while (len && seg < iov_count) {
 245                 size = min(from->iov_len, len);
 246                 to->iov_base = from->iov_base;
 247                 to->iov_len = size;
 248                 from->iov_len -= size;
 249                 from->iov_base += size;
 250                 len -= size;
 251                 ++from;
 252                 ++to;
 253                 ++seg;
 254         }
 255         return seg;
 256 }
 257 /* Copy iovec entries for len bytes from iovec. */
 258 static void copy_iovec_hdr(const struct iovec *from, struct iovec *to,
 259                            size_t len, int iovcount)
 260 {
 261         int seg = 0;
 262         size_t size;
 263
 264         while (len && seg < iovcount) {
 265                 size = min(from->iov_len, len);
 266                 to->iov_base = from->iov_base;
 267                 to->iov_len = size;
 268                 len -= size;
 269                 ++from;
 270                 ++to;
 271                 ++seg;
 272         }
 273 }
 274
 275 /* In case of DMA done not in order in lower device driver for some reason.
 276  * upend_idx is used to track end of used idx, done_idx is used to track head
 277  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 278  * guest used idx.
 279  */
 280 static int vhost_zerocopy_signal_used(struct vhost_net *net,
 281                                       struct vhost_virtqueue *vq)
 282 {
 283         struct vhost_net_virtqueue *nvq =
 284                 container_of(vq, struct vhost_net_virtqueue, vq);
 285         int i;
 286         int j = 0;
 287
 288         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 289                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 290                         vhost_net_tx_err(net);
 291                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 292                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 293                         vhost_add_used_and_signal(vq->dev, vq,
 294                                                   vq->heads[i].id, 0);
 295                         ++j;
 296                 } else
 297                         break;
 298         }
 299         if (j)
 300                 nvq->done_idx = i;
 301         return j;
 302 }
 303
 304 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 305 {
 306         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 307         struct vhost_virtqueue *vq = ubufs->vq;
 308         int cnt = atomic_read(&ubufs->kref.refcount);
 309
 310         /* set len to mark this desc buffers done DMA */
 311         vq->heads[ubuf->desc].len = success ?
 312                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 313         vhost_net_ubuf_put(ubufs);
 314
 315         /*
 316          * Trigger polling thread if guest stopped submitting new buffers:
 317          * in this case, the refcount after decrement will eventually reach 1
 318          * so here it is 2.
 319          * We also trigger polling periodically after each 16 packets
 320          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 321          * less than 10% of times).
 322          */
 323         if (cnt <= 2 || !(cnt % 16))
 324                 vhost_poll_queue(&vq->poll);
 325 }
 326
 327 /* Expects to be always run from workqueue - which acts as
 328  * read-size critical section for our kind of RCU. */
 329 static void handle_tx(struct vhost_net *net)
 330 {
 331         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 332         struct vhost_virtqueue *vq = &nvq->vq;
 333         unsigned out, in, s;
 334         int head;
 335         struct msghdr msg = {
 336                 .msg_name = NULL,
 337                 .msg_namelen = 0,
 338                 .msg_control = NULL,
 339                 .msg_controllen = 0,
 340                 .msg_iov = vq->iov,
 341                 .msg_flags = MSG_DONTWAIT,
 342         };
 343         size_t len, total_len = 0;
 344         int err;
 345         size_t hdr_size;
 346         struct socket *sock;
 347         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 348         bool zcopy, zcopy_used;
 349
 350         /* TODO: check that we are running from vhost_worker? */
 351         sock = rcu_dereference_check(vq->private_data, 1);
 352         if (!sock)
 353                 return;
 354
 355         mutex_lock(&vq->mutex);
 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                 head = vhost_get_vq_desc(&net->dev, vq, vq->iov,
 367                                          ARRAY_SIZE(vq->iov),
 368                                          &out, &in,
 369                                          NULL, NULL);
 370                 /* On error, stop handling until the next kick. */
 371                 if (unlikely(head < 0))
 372                         break;
 373                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 374                 if (head == vq->num) {
 375                         int num_pends;
 376
 377                         /* If more outstanding DMAs, queue the work.
 378                          * Handle upend_idx wrap around
 379                          */
 380                         num_pends = likely(nvq->upend_idx >= nvq->done_idx) ?
 381                                     (nvq->upend_idx - nvq->done_idx) :
 382                                     (nvq->upend_idx + UIO_MAXIOV -
 383                                      nvq->done_idx);
 384                         if (unlikely(num_pends > VHOST_MAX_PEND))
 385                                 break;
 386                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 387                                 vhost_disable_notify(&net->dev, vq);
 388                                 continue;
 389                         }
 390                         break;
 391                 }
 392                 if (in) {
 393                         vq_err(vq, "Unexpected descriptor format for TX: "
 394                                "out %d, int %d\n", out, in);
 395                         break;
 396                 }
 397                 /* Skip header. TODO: support TSO. */
 398                 s = move_iovec_hdr(vq->iov, nvq->hdr, hdr_size, out);
 399                 msg.msg_iovlen = out;
 400                 len = iov_length(vq->iov, out);
 401                 /* Sanity check */
 402                 if (!len) {
 403                         vq_err(vq, "Unexpected header len for TX: "
 404                                "%zd expected %zd\n",
 405                                iov_length(nvq->hdr, s), hdr_size);
 406                         break;
 407                 }
 408                 zcopy_used = zcopy && (len >= VHOST_GOODCOPY_LEN ||
 409                                        nvq->upend_idx != nvq->done_idx);
 410
 411                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 412                 if (zcopy_used) {
 413                         vq->heads[nvq->upend_idx].id = head;
 414                         if (!vhost_net_tx_select_zcopy(net) ||
 415                             len < VHOST_GOODCOPY_LEN) {
 416                                 /* copy don't need to wait for DMA done */
 417                                 vq->heads[nvq->upend_idx].len =
 418                                                         VHOST_DMA_DONE_LEN;
 419                                 msg.msg_control = NULL;
 420                                 msg.msg_controllen = 0;
 421                                 ubufs = NULL;
 422                         } else {
 423                                 struct ubuf_info *ubuf;
 424                                 ubuf = nvq->ubuf_info + nvq->upend_idx;
 425
 426                                 vq->heads[nvq->upend_idx].len =
 427                                         VHOST_DMA_IN_PROGRESS;
 428                                 ubuf->callback = vhost_zerocopy_callback;
 429                                 ubuf->ctx = nvq->ubufs;
 430                                 ubuf->desc = nvq->upend_idx;
 431                                 msg.msg_control = ubuf;
 432                                 msg.msg_controllen = sizeof(ubuf);
 433                                 ubufs = nvq->ubufs;
 434                                 kref_get(&ubufs->kref);
 435                         }
 436                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 437                 } else
 438                         msg.msg_control = NULL;
 439                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 440                 err = sock->ops->sendmsg(NULL, sock, &msg, len);
 441                 if (unlikely(err < 0)) {
 442                         if (zcopy_used) {
 443                                 if (ubufs)
 444                                         vhost_net_ubuf_put(ubufs);
 445                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 446                                         % UIO_MAXIOV;
 447                         }
 448                         vhost_discard_vq_desc(vq, 1);
 449                         break;
 450                 }
 451                 if (err != len)
 452                         pr_debug("Truncated TX packet: "
 453                                  " len %d != %zd\n", err, len);
 454                 if (!zcopy_used)
 455                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 456                 else
 457                         vhost_zerocopy_signal_used(net, vq);
 458                 total_len += len;
 459                 vhost_net_tx_packet(net);
 460                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 461                         vhost_poll_queue(&vq->poll);
 462                         break;
 463                 }
 464         }
 465
 466         mutex_unlock(&vq->mutex);
 467 }
 468
 469 static int peek_head_len(struct sock *sk)
 470 {
 471         struct sk_buff *head;
 472         int len = 0;
 473         unsigned long flags;
 474
 475         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 476         head = skb_peek(&sk->sk_receive_queue);
 477         if (likely(head)) {
 478                 len = head->len;
 479                 if (vlan_tx_tag_present(head))
 480                         len += VLAN_HLEN;
 481         }
 482
 483         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 484         return len;
 485 }
 486
 487 /* This is a multi-buffer version of vhost_get_desc, that works if
 488  *      vq has read descriptors only.
 489  * @vq          - the relevant virtqueue
 490  * @datalen     - data length we'll be reading
 491  * @iovcount    - returned count of io vectors we fill
 492  * @log         - vhost log
 493  * @log_num     - log offset
 494  * @quota       - headcount quota, 1 for big buffer
 495  *      returns number of buffer heads allocated, negative on error
 496  */
 497 static int get_rx_bufs(struct vhost_virtqueue *vq,
 498                        struct vring_used_elem *heads,
 499                        int datalen,
 500                        unsigned *iovcount,
 501                        struct vhost_log *log,
 502                        unsigned *log_num,
 503                        unsigned int quota)
 504 {
 505         unsigned int out, in;
 506         int seg = 0;
 507         int headcount = 0;
 508         unsigned d;
 509         int r, nlogs = 0;
 510
 511         while (datalen > 0 && headcount < quota) {
 512                 if (unlikely(seg >= UIO_MAXIOV)) {
 513                         r = -ENOBUFS;
 514                         goto err;
 515                 }
 516                 r = vhost_get_vq_desc(vq->dev, vq, vq->iov + seg,
 517                                       ARRAY_SIZE(vq->iov) - seg, &out,
 518                                       &in, log, log_num);
 519                 if (unlikely(r < 0))
 520                         goto err;
 521
 522                 d = r;
 523                 if (d == vq->num) {
 524                         r = 0;
 525                         goto err;
 526                 }
 527                 if (unlikely(out || in <= 0)) {
 528                         vq_err(vq, "unexpected descriptor format for RX: "
 529                                 "out %d, in %d\n", out, in);
 530                         r = -EINVAL;
 531                         goto err;
 532                 }
 533                 if (unlikely(log)) {
 534                         nlogs += *log_num;
 535                         log += *log_num;
 536                 }
 537                 heads[headcount].id = d;
 538                 heads[headcount].len = iov_length(vq->iov + seg, in);
 539                 datalen -= heads[headcount].len;
 540                 ++headcount;
 541                 seg += in;
 542         }
 543         heads[headcount - 1].len += datalen;
 544         *iovcount = seg;
 545         if (unlikely(log))
 546                 *log_num = nlogs;
 547
 548         /* Detect overrun */
 549         if (unlikely(datalen > 0)) {
 550                 r = UIO_MAXIOV + 1;
 551                 goto err;
 552         }
 553         return headcount;
 554 err:
 555         vhost_discard_vq_desc(vq, headcount);
 556         return r;
 557 }
 558
 559 /* Expects to be always run from workqueue - which acts as
 560  * read-size critical section for our kind of RCU. */
 561 static void handle_rx(struct vhost_net *net)
 562 {
 563         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 564         struct vhost_virtqueue *vq = &nvq->vq;
 565         unsigned uninitialized_var(in), log;
 566         struct vhost_log *vq_log;
 567         struct msghdr msg = {
 568                 .msg_name = NULL,
 569                 .msg_namelen = 0,
 570                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 571                 .msg_controllen = 0,
 572                 .msg_iov = vq->iov,
 573                 .msg_flags = MSG_DONTWAIT,
 574         };
 575         struct virtio_net_hdr_mrg_rxbuf hdr = {
 576                 .hdr.flags = 0,
 577                 .hdr.gso_type = VIRTIO_NET_HDR_GSO_NONE
 578         };
 579         size_t total_len = 0;
 580         int err, mergeable;
 581         s16 headcount;
 582         size_t vhost_hlen, sock_hlen;
 583         size_t vhost_len, sock_len;
 584         /* TODO: check that we are running from vhost_worker? */
 585         struct socket *sock = rcu_dereference_check(vq->private_data, 1);
 586
 587         if (!sock)
 588                 return;
 589
 590         mutex_lock(&vq->mutex);
 591         vhost_disable_notify(&net->dev, vq);
 592         vhost_hlen = nvq->vhost_hlen;
 593         sock_hlen = nvq->sock_hlen;
 594
 595         vq_log = unlikely(vhost_has_feature(&net->dev, VHOST_F_LOG_ALL)) ?
 596                 vq->log : NULL;
 597         mergeable = vhost_has_feature(&net->dev, VIRTIO_NET_F_MRG_RXBUF);
 598
 599         while ((sock_len = peek_head_len(sock->sk))) {
 600                 sock_len += sock_hlen;
 601                 vhost_len = sock_len + vhost_hlen;
 602                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 603                                         &in, vq_log, &log,
 604                                         likely(mergeable) ? UIO_MAXIOV : 1);
 605                 /* On error, stop handling until the next kick. */
 606                 if (unlikely(headcount < 0))
 607                         break;
 608                 /* On overrun, truncate and discard */
 609                 if (unlikely(headcount > UIO_MAXIOV)) {
 610                         msg.msg_iovlen = 1;
 611                         err = sock->ops->recvmsg(NULL, sock, &msg,
 612                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 613                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 614                         continue;
 615                 }
 616                 /* OK, now we need to know about added descriptors. */
 617                 if (!headcount) {
 618                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 619                                 /* They have slipped one in as we were
 620                                  * doing that: check again. */
 621                                 vhost_disable_notify(&net->dev, vq);
 622                                 continue;
 623                         }
 624                         /* Nothing new?  Wait for eventfd to tell us
 625                          * they refilled. */
 626                         break;
 627                 }
 628                 /* We don't need to be notified again. */
 629                 if (unlikely((vhost_hlen)))
 630                         /* Skip header. TODO: support TSO. */
 631                         move_iovec_hdr(vq->iov, nvq->hdr, vhost_hlen, in);
 632                 else
 633                         /* Copy the header for use in VIRTIO_NET_F_MRG_RXBUF:
 634                          * needed because recvmsg can modify msg_iov. */
 635                         copy_iovec_hdr(vq->iov, nvq->hdr, sock_hlen, in);
 636                 msg.msg_iovlen = in;
 637                 err = sock->ops->recvmsg(NULL, sock, &msg,
 638                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 639                 /* Userspace might have consumed the packet meanwhile:
 640                  * it's not supposed to do this usually, but might be hard
 641                  * to prevent. Discard data we got (if any) and keep going. */
 642                 if (unlikely(err != sock_len)) {
 643                         pr_debug("Discarded rx packet: "
 644                                  " len %d, expected %zd\n", err, sock_len);
 645                         vhost_discard_vq_desc(vq, headcount);
 646                         continue;
 647                 }
 648                 if (unlikely(vhost_hlen) &&
 649                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&hdr, 0,
 650                                       vhost_hlen)) {
 651                         vq_err(vq, "Unable to write vnet_hdr at addr %p\n",
 652                                vq->iov->iov_base);
 653                         break;
 654                 }
 655                 /* TODO: Should check and handle checksum. */
 656                 if (likely(mergeable) &&
 657                     memcpy_toiovecend(nvq->hdr, (unsigned char *)&headcount,
 658                                       offsetof(typeof(hdr), num_buffers),
 659                                       sizeof hdr.num_buffers)) {
 660                         vq_err(vq, "Failed num_buffers write");
 661                         vhost_discard_vq_desc(vq, headcount);
 662                         break;
 663                 }
 664                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 665                                             headcount);
 666                 if (unlikely(vq_log))
 667                         vhost_log_write(vq, vq_log, log, vhost_len);
 668                 total_len += vhost_len;
 669                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 670                         vhost_poll_queue(&vq->poll);
 671                         break;
 672                 }
 673         }
 674
 675         mutex_unlock(&vq->mutex);
 676 }
 677
 678 static void handle_tx_kick(struct vhost_work *work)
 679 {
 680         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 681                                                   poll.work);
 682         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 683
 684         handle_tx(net);
 685 }
 686
 687 static void handle_rx_kick(struct vhost_work *work)
 688 {
 689         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 690                                                   poll.work);
 691         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 692
 693         handle_rx(net);
 694 }
 695
 696 static void handle_tx_net(struct vhost_work *work)
 697 {
 698         struct vhost_net *net = container_of(work, struct vhost_net,
 699                                              poll[VHOST_NET_VQ_TX].work);
 700         handle_tx(net);
 701 }
 702
 703 static void handle_rx_net(struct vhost_work *work)
 704 {
 705         struct vhost_net *net = container_of(work, struct vhost_net,
 706                                              poll[VHOST_NET_VQ_RX].work);
 707         handle_rx(net);
 708 }
 709
 710 static int vhost_net_open(struct inode *inode, struct file *f)
 711 {
 712         struct vhost_net *n = kmalloc(sizeof *n, GFP_KERNEL);
 713         struct vhost_dev *dev;
 714         struct vhost_virtqueue **vqs;
 715         int r, i;
 716
 717         if (!n)
 718                 return -ENOMEM;
 719         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 720         if (!vqs) {
 721                 kfree(n);
 722                 return -ENOMEM;
 723         }
 724
 725         dev = &n->dev;
 726         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 727         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 728         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 729         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 730         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 731                 n->vqs[i].ubufs = NULL;
 732                 n->vqs[i].ubuf_info = NULL;
 733                 n->vqs[i].upend_idx = 0;
 734                 n->vqs[i].done_idx = 0;
 735                 n->vqs[i].vhost_hlen = 0;
 736                 n->vqs[i].sock_hlen = 0;
 737         }
 738         r = vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 739         if (r < 0) {
 740                 kfree(n);
 741                 kfree(vqs);
 742                 return r;
 743         }
 744
 745         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 746         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 747
 748         f->private_data = n;
 749
 750         return 0;
 751 }
 752
 753 static void vhost_net_disable_vq(struct vhost_net *n,
 754                                  struct vhost_virtqueue *vq)
 755 {
 756         struct vhost_net_virtqueue *nvq =
 757                 container_of(vq, struct vhost_net_virtqueue, vq);
 758         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 759         if (!vq->private_data)
 760                 return;
 761         vhost_poll_stop(poll);
 762 }
 763
 764 static int vhost_net_enable_vq(struct vhost_net *n,
 765                                 struct vhost_virtqueue *vq)
 766 {
 767         struct vhost_net_virtqueue *nvq =
 768                 container_of(vq, struct vhost_net_virtqueue, vq);
 769         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 770         struct socket *sock;
 771
 772         sock = rcu_dereference_protected(vq->private_data,
 773                                          lockdep_is_held(&vq->mutex));
 774         if (!sock)
 775                 return 0;
 776
 777         return vhost_poll_start(poll, sock->file);
 778 }
 779
 780 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 781                                         struct vhost_virtqueue *vq)
 782 {
 783         struct socket *sock;
 784
 785         mutex_lock(&vq->mutex);
 786         sock = rcu_dereference_protected(vq->private_data,
 787                                          lockdep_is_held(&vq->mutex));
 788         vhost_net_disable_vq(n, vq);
 789         rcu_assign_pointer(vq->private_data, NULL);
 790         mutex_unlock(&vq->mutex);
 791         return sock;
 792 }
 793
 794 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 795                            struct socket **rx_sock)
 796 {
 797         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 798         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 799 }
 800
 801 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 802 {
 803         vhost_poll_flush(n->poll + index);
 804         vhost_poll_flush(&n->vqs[index].vq.poll);
 805 }
 806
 807 static void vhost_net_flush(struct vhost_net *n)
 808 {
 809         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 810         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 811         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 812                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 813                 n->tx_flush = true;
 814                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 815                 /* Wait for all lower device DMAs done. */
 816                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 817                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 818                 n->tx_flush = false;
 819                 kref_init(&n->vqs[VHOST_NET_VQ_TX].ubufs->kref);
 820                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 821         }
 822 }
 823
 824 static int vhost_net_release(struct inode *inode, struct file *f)
 825 {
 826         struct vhost_net *n = f->private_data;
 827         struct socket *tx_sock;
 828         struct socket *rx_sock;
 829
 830         vhost_net_stop(n, &tx_sock, &rx_sock);
 831         vhost_net_flush(n);
 832         vhost_dev_stop(&n->dev);
 833         vhost_dev_cleanup(&n->dev, false);
 834         vhost_net_vq_reset(n);
 835         if (tx_sock)
 836                 fput(tx_sock->file);
 837         if (rx_sock)
 838                 fput(rx_sock->file);
 839         /* We do an extra flush before freeing memory,
 840          * since jobs can re-queue themselves. */
 841         vhost_net_flush(n);
 842         kfree(n->dev.vqs);
 843         kfree(n);
 844         return 0;
 845 }
 846
 847 static struct socket *get_raw_socket(int fd)
 848 {
 849         struct {
 850                 struct sockaddr_ll sa;
 851                 char  buf[MAX_ADDR_LEN];
 852         } uaddr;
 853         int uaddr_len = sizeof uaddr, r;
 854         struct socket *sock = sockfd_lookup(fd, &r);
 855
 856         if (!sock)
 857                 return ERR_PTR(-ENOTSOCK);
 858
 859         /* Parameter checking */
 860         if (sock->sk->sk_type != SOCK_RAW) {
 861                 r = -ESOCKTNOSUPPORT;
 862                 goto err;
 863         }
 864
 865         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 866                                &uaddr_len, 0);
 867         if (r)
 868                 goto err;
 869
 870         if (uaddr.sa.sll_family != AF_PACKET) {
 871                 r = -EPFNOSUPPORT;
 872                 goto err;
 873         }
 874         return sock;
 875 err:
 876         fput(sock->file);
 877         return ERR_PTR(r);
 878 }
 879
 880 static struct socket *get_tap_socket(int fd)
 881 {
 882         struct file *file = fget(fd);
 883         struct socket *sock;
 884
 885         if (!file)
 886                 return ERR_PTR(-EBADF);
 887         sock = tun_get_socket(file);
 888         if (!IS_ERR(sock))
 889                 return sock;
 890         sock = macvtap_get_socket(file);
 891         if (IS_ERR(sock))
 892                 fput(file);
 893         return sock;
 894 }
 895
 896 static struct socket *get_socket(int fd)
 897 {
 898         struct socket *sock;
 899
 900         /* special case to disable backend */
 901         if (fd == -1)
 902                 return NULL;
 903         sock = get_raw_socket(fd);
 904         if (!IS_ERR(sock))
 905                 return sock;
 906         sock = get_tap_socket(fd);
 907         if (!IS_ERR(sock))
 908                 return sock;
 909         return ERR_PTR(-ENOTSOCK);
 910 }
 911
 912 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 913 {
 914         struct socket *sock, *oldsock;
 915         struct vhost_virtqueue *vq;
 916         struct vhost_net_virtqueue *nvq;
 917         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 918         int r;
 919
 920         mutex_lock(&n->dev.mutex);
 921         r = vhost_dev_check_owner(&n->dev);
 922         if (r)
 923                 goto err;
 924
 925         if (index >= VHOST_NET_VQ_MAX) {
 926                 r = -ENOBUFS;
 927                 goto err;
 928         }
 929         vq = &n->vqs[index].vq;
 930         nvq = &n->vqs[index];
 931         mutex_lock(&vq->mutex);
 932
 933         /* Verify that ring has been setup correctly. */
 934         if (!vhost_vq_access_ok(vq)) {
 935                 r = -EFAULT;
 936                 goto err_vq;
 937         }
 938         sock = get_socket(fd);
 939         if (IS_ERR(sock)) {
 940                 r = PTR_ERR(sock);
 941                 goto err_vq;
 942         }
 943
 944         /* start polling new socket */
 945         oldsock = rcu_dereference_protected(vq->private_data,
 946                                             lockdep_is_held(&vq->mutex));
 947         if (sock != oldsock) {
 948                 ubufs = vhost_net_ubuf_alloc(vq,
 949                                              sock && vhost_sock_zcopy(sock));
 950                 if (IS_ERR(ubufs)) {
 951                         r = PTR_ERR(ubufs);
 952                         goto err_ubufs;
 953                 }
 954
 955                 vhost_net_disable_vq(n, vq);
 956                 rcu_assign_pointer(vq->private_data, sock);
 957                 r = vhost_init_used(vq);
 958                 if (r)
 959                         goto err_used;
 960                 r = vhost_net_enable_vq(n, vq);
 961                 if (r)
 962                         goto err_used;
 963
 964                 oldubufs = nvq->ubufs;
 965                 nvq->ubufs = ubufs;
 966
 967                 n->tx_packets = 0;
 968                 n->tx_zcopy_err = 0;
 969                 n->tx_flush = false;
 970         }
 971
 972         mutex_unlock(&vq->mutex);
 973
 974         if (oldubufs) {
 975                 vhost_net_ubuf_put_wait_and_free(oldubufs);
 976                 mutex_lock(&vq->mutex);
 977                 vhost_zerocopy_signal_used(n, vq);
 978                 mutex_unlock(&vq->mutex);
 979         }
 980
 981         if (oldsock) {
 982                 vhost_net_flush_vq(n, index);
 983                 fput(oldsock->file);
 984         }
 985
 986         mutex_unlock(&n->dev.mutex);
 987         return 0;
 988
 989 err_used:
 990         rcu_assign_pointer(vq->private_data, oldsock);
 991         vhost_net_enable_vq(n, vq);
 992         if (ubufs)
 993                 vhost_net_ubuf_put_wait_and_free(ubufs);
 994 err_ubufs:
 995         fput(sock->file);
 996 err_vq:
 997         mutex_unlock(&vq->mutex);
 998 err:
 999         mutex_unlock(&n->dev.mutex);
1000         return r;
1001 }
1002
1003 static long vhost_net_reset_owner(struct vhost_net *n)
1004 {
1005         struct socket *tx_sock = NULL;
1006         struct socket *rx_sock = NULL;
1007         long err;
1008         struct vhost_memory *memory;
1009
1010         mutex_lock(&n->dev.mutex);
1011         err = vhost_dev_check_owner(&n->dev);
1012         if (err)
1013                 goto done;
1014         memory = vhost_dev_reset_owner_prepare();
1015         if (!memory) {
1016                 err = -ENOMEM;
1017                 goto done;
1018         }
1019         vhost_net_stop(n, &tx_sock, &rx_sock);
1020         vhost_net_flush(n);
1021         vhost_dev_reset_owner(&n->dev, memory);
1022         vhost_net_vq_reset(n);
1023 done:
1024         mutex_unlock(&n->dev.mutex);
1025         if (tx_sock)
1026                 fput(tx_sock->file);
1027         if (rx_sock)
1028                 fput(rx_sock->file);
1029         return err;
1030 }
1031
1032 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1033 {
1034         size_t vhost_hlen, sock_hlen, hdr_len;
1035         int i;
1036
1037         hdr_len = (features & (1 << VIRTIO_NET_F_MRG_RXBUF)) ?
1038                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1039                         sizeof(struct virtio_net_hdr);
1040         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1041                 /* vhost provides vnet_hdr */
1042                 vhost_hlen = hdr_len;
1043                 sock_hlen = 0;
1044         } else {
1045                 /* socket provides vnet_hdr */
1046                 vhost_hlen = 0;
1047                 sock_hlen = hdr_len;
1048         }
1049         mutex_lock(&n->dev.mutex);
1050         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1051             !vhost_log_access_ok(&n->dev)) {
1052                 mutex_unlock(&n->dev.mutex);
1053                 return -EFAULT;
1054         }
1055         n->dev.acked_features = features;
1056         smp_wmb();
1057         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1058                 mutex_lock(&n->vqs[i].vq.mutex);
1059                 n->vqs[i].vhost_hlen = vhost_hlen;
1060                 n->vqs[i].sock_hlen = sock_hlen;
1061                 mutex_unlock(&n->vqs[i].vq.mutex);
1062         }
1063         vhost_net_flush(n);
1064         mutex_unlock(&n->dev.mutex);
1065         return 0;
1066 }
1067
1068 static long vhost_net_set_owner(struct vhost_net *n)
1069 {
1070         int r;
1071
1072         mutex_lock(&n->dev.mutex);
1073         if (vhost_dev_has_owner(&n->dev)) {
1074                 r = -EBUSY;
1075                 goto out;
1076         }
1077         r = vhost_net_set_ubuf_info(n);
1078         if (r)
1079                 goto out;
1080         r = vhost_dev_set_owner(&n->dev);
1081         if (r)
1082                 vhost_net_clear_ubuf_info(n);
1083         vhost_net_flush(n);
1084 out:
1085         mutex_unlock(&n->dev.mutex);
1086         return r;
1087 }
1088
1089 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1090                             unsigned long arg)
1091 {
1092         struct vhost_net *n = f->private_data;
1093         void __user *argp = (void __user *)arg;
1094         u64 __user *featurep = argp;
1095         struct vhost_vring_file backend;
1096         u64 features;
1097         int r;
1098
1099         switch (ioctl) {
1100         case VHOST_NET_SET_BACKEND:
1101                 if (copy_from_user(&backend, argp, sizeof backend))
1102                         return -EFAULT;
1103                 return vhost_net_set_backend(n, backend.index, backend.fd);
1104         case VHOST_GET_FEATURES:
1105                 features = VHOST_NET_FEATURES;
1106                 if (copy_to_user(featurep, &features, sizeof features))
1107                         return -EFAULT;
1108                 return 0;
1109         case VHOST_SET_FEATURES:
1110                 if (copy_from_user(&features, featurep, sizeof features))
1111                         return -EFAULT;
1112                 if (features & ~VHOST_NET_FEATURES)
1113                         return -EOPNOTSUPP;
1114                 return vhost_net_set_features(n, features);
1115         case VHOST_RESET_OWNER:
1116                 return vhost_net_reset_owner(n);
1117         case VHOST_SET_OWNER:
1118                 return vhost_net_set_owner(n);
1119         default:
1120                 mutex_lock(&n->dev.mutex);
1121                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1122                 if (r == -ENOIOCTLCMD)
1123                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1124                 else
1125                         vhost_net_flush(n);
1126                 mutex_unlock(&n->dev.mutex);
1127                 return r;
1128         }
1129 }
1130
1131 #ifdef CONFIG_COMPAT
1132 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1133                                    unsigned long arg)
1134 {
1135         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1136 }
1137 #endif
1138
1139 static const struct file_operations vhost_net_fops = {
1140         .owner          = THIS_MODULE,
1141         .release        = vhost_net_release,
1142         .unlocked_ioctl = vhost_net_ioctl,
1143 #ifdef CONFIG_COMPAT
1144         .compat_ioctl   = vhost_net_compat_ioctl,
1145 #endif
1146         .open           = vhost_net_open,
1147         .llseek         = noop_llseek,
1148 };
1149
1150 static struct miscdevice vhost_net_misc = {
1151         .minor = VHOST_NET_MINOR,
1152         .name = "vhost-net",
1153         .fops = &vhost_net_fops,
1154 };
1155
1156 static int vhost_net_init(void)
1157 {
1158         if (experimental_zcopytx)
1159                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1160         return misc_register(&vhost_net_misc);
1161 }
1162 module_init(vhost_net_init);
1163
1164 static void vhost_net_exit(void)
1165 {
1166         misc_deregister(&vhost_net_misc);
1167 }
1168 module_exit(vhost_net_exit);
1169
1170 MODULE_VERSION("0.0.1");
1171 MODULE_LICENSE("GPL v2");
1172 MODULE_AUTHOR("Michael S. Tsirkin");
1173 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1174 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1175 MODULE_ALIAS("devname:vhost-net");