drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <mst@redhat.com>
   3  *
   4  * This work is licensed under the terms of the GNU GPL, version 2.
   5  *
   6  * virtio-net server in host kernel.
   7  */
   8
   9 #include <linux/compat.h>
  10 #include <linux/eventfd.h>
  11 #include <linux/vhost.h>
  12 #include <linux/virtio_net.h>
  13 #include <linux/miscdevice.h>
  14 #include <linux/module.h>
  15 #include <linux/moduleparam.h>
  16 #include <linux/mutex.h>
  17 #include <linux/workqueue.h>
  18 #include <linux/file.h>
  19 #include <linux/slab.h>
  20 #include <linux/vmalloc.h>
  21
  22 #include <linux/net.h>
  23 #include <linux/if_packet.h>
  24 #include <linux/if_arp.h>
  25 #include <linux/if_tun.h>
  26 #include <linux/if_macvlan.h>
  27 #include <linux/if_vlan.h>
  28
  29 #include <net/sock.h>
  30
  31 #include "vhost.h"
  32
  33 static int experimental_zcopytx = 1;
  34 module_param(experimental_zcopytx, int, 0444);
  35 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  36                                        " 1 -Enable; 0 - Disable");
  37
  38 /* Max number of bytes transferred before requeueing the job.
  39  * Using this limit prevents one virtqueue from starving others. */
  40 #define VHOST_NET_WEIGHT 0x80000
  41
  42 /* MAX number of TX used buffers for outstanding zerocopy */
  43 #define VHOST_MAX_PEND 128
  44 #define VHOST_GOODCOPY_LEN 256
  45
  46 /*
  47  * For transmit, used buffer len is unused; we override it to track buffer
  48  * status internally; used for zerocopy tx only.
  49  */
  50 /* Lower device DMA failed */
  51 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  52 /* Lower device DMA done */
  53 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  54 /* Lower device DMA in progress */
  55 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  56 /* Buffer unused */
  57 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  58
  59 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  60
  61 enum {
  62         VHOST_NET_FEATURES = VHOST_FEATURES |
  63                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  64                          (1ULL << VIRTIO_NET_F_MRG_RXBUF)
  65 };
  66
  67 enum {
  68         VHOST_NET_VQ_RX = 0,
  69         VHOST_NET_VQ_TX = 1,
  70         VHOST_NET_VQ_MAX = 2,
  71 };
  72
  73 struct vhost_net_ubuf_ref {
  74         /* refcount follows semantics similar to kref:
  75          *  0: object is released
  76          *  1: no outstanding ubufs
  77          * >1: outstanding ubufs
  78          */
  79         atomic_t refcount;
  80         wait_queue_head_t wait;
  81         struct vhost_virtqueue *vq;
  82 };
  83
  84 struct vhost_net_virtqueue {
  85         struct vhost_virtqueue vq;
  86         size_t vhost_hlen;
  87         size_t sock_hlen;
  88         /* vhost zerocopy support fields below: */
  89         /* last used idx for outstanding DMA zerocopy buffers */
  90         int upend_idx;
  91         /* first used idx for DMA done zerocopy buffers */
  92         int done_idx;
  93         /* an array of userspace buffers info */
  94         struct ubuf_info *ubuf_info;
  95         /* Reference counting for outstanding ubufs.
  96          * Protected by vq mutex. Writers must also take device mutex. */
  97         struct vhost_net_ubuf_ref *ubufs;
  98 };
  99
 100 struct vhost_net {
 101         struct vhost_dev dev;
 102         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 103         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 104         /* Number of TX recently submitted.
 105          * Protected by tx vq lock. */
 106         unsigned tx_packets;
 107         /* Number of times zerocopy TX recently failed.
 108          * Protected by tx vq lock. */
 109         unsigned tx_zcopy_err;
 110         /* Flush in progress. Protected by tx vq lock. */
 111         bool tx_flush;
 112 };
 113
 114 static unsigned vhost_net_zcopy_mask __read_mostly;
 115
 116 static void vhost_net_enable_zcopy(int vq)
 117 {
 118         vhost_net_zcopy_mask |= 0x1 << vq;
 119 }
 120
 121 static struct vhost_net_ubuf_ref *
 122 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 123 {
 124         struct vhost_net_ubuf_ref *ubufs;
 125         /* No zero copy backend? Nothing to count. */
 126         if (!zcopy)
 127                 return NULL;
 128         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 129         if (!ubufs)
 130                 return ERR_PTR(-ENOMEM);
 131         atomic_set(&ubufs->refcount, 1);
 132         init_waitqueue_head(&ubufs->wait);
 133         ubufs->vq = vq;
 134         return ubufs;
 135 }
 136
 137 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 138 {
 139         int r = atomic_sub_return(1, &ubufs->refcount);
 140         if (unlikely(!r))
 141                 wake_up(&ubufs->wait);
 142         return r;
 143 }
 144
 145 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 146 {
 147         vhost_net_ubuf_put(ubufs);
 148         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 149 }
 150
 151 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 152 {
 153         vhost_net_ubuf_put_and_wait(ubufs);
 154         kfree(ubufs);
 155 }
 156
 157 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 158 {
 159         int i;
 160
 161         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 162                 kfree(n->vqs[i].ubuf_info);
 163                 n->vqs[i].ubuf_info = NULL;
 164         }
 165 }
 166
 167 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 168 {
 169         bool zcopy;
 170         int i;
 171
 172         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 173                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 174                 if (!zcopy)
 175                         continue;
 176                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 177                                               UIO_MAXIOV, GFP_KERNEL);
 178                 if  (!n->vqs[i].ubuf_info)
 179                         goto err;
 180         }
 181         return 0;
 182
 183 err:
 184         vhost_net_clear_ubuf_info(n);
 185         return -ENOMEM;
 186 }
 187
 188 static void vhost_net_vq_reset(struct vhost_net *n)
 189 {
 190         int i;
 191
 192         vhost_net_clear_ubuf_info(n);
 193
 194         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 195                 n->vqs[i].done_idx = 0;
 196                 n->vqs[i].upend_idx = 0;
 197                 n->vqs[i].ubufs = NULL;
 198                 n->vqs[i].vhost_hlen = 0;
 199                 n->vqs[i].sock_hlen = 0;
 200         }
 201
 202 }
 203
 204 static void vhost_net_tx_packet(struct vhost_net *net)
 205 {
 206         ++net->tx_packets;
 207         if (net->tx_packets < 1024)
 208                 return;
 209         net->tx_packets = 0;
 210         net->tx_zcopy_err = 0;
 211 }
 212
 213 static void vhost_net_tx_err(struct vhost_net *net)
 214 {
 215         ++net->tx_zcopy_err;
 216 }
 217
 218 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 219 {
 220         /* TX flush waits for outstanding DMAs to be done.
 221          * Don't start new DMAs.
 222          */
 223         return !net->tx_flush &&
 224                 net->tx_packets / 64 >= net->tx_zcopy_err;
 225 }
 226
 227 static bool vhost_sock_zcopy(struct socket *sock)
 228 {
 229         return unlikely(experimental_zcopytx) &&
 230                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 231 }
 232
 233 /* In case of DMA done not in order in lower device driver for some reason.
 234  * upend_idx is used to track end of used idx, done_idx is used to track head
 235  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 236  * guest used idx.
 237  */
 238 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 239                                        struct vhost_virtqueue *vq)
 240 {
 241         struct vhost_net_virtqueue *nvq =
 242                 container_of(vq, struct vhost_net_virtqueue, vq);
 243         int i, add;
 244         int j = 0;
 245
 246         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 247                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 248                         vhost_net_tx_err(net);
 249                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 250                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 251                         ++j;
 252                 } else
 253                         break;
 254         }
 255         while (j) {
 256                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 257                 vhost_add_used_and_signal_n(vq->dev, vq,
 258                                             &vq->heads[nvq->done_idx], add);
 259                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 260                 j -= add;
 261         }
 262 }
 263
 264 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 265 {
 266         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 267         struct vhost_virtqueue *vq = ubufs->vq;
 268         int cnt;
 269
 270         rcu_read_lock_bh();
 271
 272         /* set len to mark this desc buffers done DMA */
 273         vq->heads[ubuf->desc].len = success ?
 274                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 275         cnt = vhost_net_ubuf_put(ubufs);
 276
 277         /*
 278          * Trigger polling thread if guest stopped submitting new buffers:
 279          * in this case, the refcount after decrement will eventually reach 1.
 280          * We also trigger polling periodically after each 16 packets
 281          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 282          * less than 10% of times).
 283          */
 284         if (cnt <= 1 || !(cnt % 16))
 285                 vhost_poll_queue(&vq->poll);
 286
 287         rcu_read_unlock_bh();
 288 }
 289
 290 static inline unsigned long busy_clock(void)
 291 {
 292         return local_clock() >> 10;
 293 }
 294
 295 static bool vhost_can_busy_poll(struct vhost_dev *dev,
 296                                 unsigned long endtime)
 297 {
 298         return likely(!need_resched()) &&
 299                likely(!time_after(busy_clock(), endtime)) &&
 300                likely(!signal_pending(current)) &&
 301                !vhost_has_work(dev);
 302 }
 303
 304 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 305                                     struct vhost_virtqueue *vq,
 306                                     struct iovec iov[], unsigned int iov_size,
 307                                     unsigned int *out_num, unsigned int *in_num)
 308 {
 309         unsigned long uninitialized_var(endtime);
 310         int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 311                                     out_num, in_num, NULL, NULL);
 312
 313         if (r == vq->num && vq->busyloop_timeout) {
 314                 preempt_disable();
 315                 endtime = busy_clock() + vq->busyloop_timeout;
 316                 while (vhost_can_busy_poll(vq->dev, endtime) &&
 317                        vhost_vq_avail_empty(vq->dev, vq))
 318                         cpu_relax_lowlatency();
 319                 preempt_enable();
 320                 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 321                                         out_num, in_num, NULL, NULL);
 322         }
 323
 324         return r;
 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;
 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_flags = MSG_DONTWAIT,
 341         };
 342         size_t len, total_len = 0;
 343         int err;
 344         size_t hdr_size;
 345         struct socket *sock;
 346         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 347         bool zcopy, zcopy_used;
 348
 349         mutex_lock(&vq->mutex);
 350         sock = vq->private_data;
 351         if (!sock)
 352                 goto out;
 353
 354         vhost_disable_notify(&net->dev, vq);
 355
 356         hdr_size = nvq->vhost_hlen;
 357         zcopy = nvq->ubufs;
 358
 359         for (;;) {
 360                 /* Release DMAs done buffers first */
 361                 if (zcopy)
 362                         vhost_zerocopy_signal_used(net, vq);
 363
 364                 /* If more outstanding DMAs, queue the work.
 365                  * Handle upend_idx wrap around
 366                  */
 367                 if (unlikely((nvq->upend_idx + vq->num - VHOST_MAX_PEND)
 368                               % UIO_MAXIOV == nvq->done_idx))
 369                         break;
 370
 371                 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
 372                                                 ARRAY_SIZE(vq->iov),
 373                                                 &out, &in);
 374                 /* On error, stop handling until the next kick. */
 375                 if (unlikely(head < 0))
 376                         break;
 377                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 378                 if (head == vq->num) {
 379                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 380                                 vhost_disable_notify(&net->dev, vq);
 381                                 continue;
 382                         }
 383                         break;
 384                 }
 385                 if (in) {
 386                         vq_err(vq, "Unexpected descriptor format for TX: "
 387                                "out %d, int %d\n", out, in);
 388                         break;
 389                 }
 390                 /* Skip header. TODO: support TSO. */
 391                 len = iov_length(vq->iov, out);
 392                 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
 393                 iov_iter_advance(&msg.msg_iter, hdr_size);
 394                 /* Sanity check */
 395                 if (!msg_data_left(&msg)) {
 396                         vq_err(vq, "Unexpected header len for TX: "
 397                                "%zd expected %zd\n",
 398                                len, hdr_size);
 399                         break;
 400                 }
 401                 len = msg_data_left(&msg);
 402
 403                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 404                                    && (nvq->upend_idx + 1) % UIO_MAXIOV !=
 405                                       nvq->done_idx
 406                                    && vhost_net_tx_select_zcopy(net);
 407
 408                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 409                 if (zcopy_used) {
 410                         struct ubuf_info *ubuf;
 411                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 412
 413                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 414                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 415                         ubuf->callback = vhost_zerocopy_callback;
 416                         ubuf->ctx = nvq->ubufs;
 417                         ubuf->desc = nvq->upend_idx;
 418                         msg.msg_control = ubuf;
 419                         msg.msg_controllen = sizeof(ubuf);
 420                         ubufs = nvq->ubufs;
 421                         atomic_inc(&ubufs->refcount);
 422                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 423                 } else {
 424                         msg.msg_control = NULL;
 425                         ubufs = NULL;
 426                 }
 427                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 428                 err = sock->ops->sendmsg(sock, &msg, len);
 429                 if (unlikely(err < 0)) {
 430                         if (zcopy_used) {
 431                                 vhost_net_ubuf_put(ubufs);
 432                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 433                                         % UIO_MAXIOV;
 434                         }
 435                         vhost_discard_vq_desc(vq, 1);
 436                         break;
 437                 }
 438                 if (err != len)
 439                         pr_debug("Truncated TX packet: "
 440                                  " len %d != %zd\n", err, len);
 441                 if (!zcopy_used)
 442                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 443                 else
 444                         vhost_zerocopy_signal_used(net, vq);
 445                 total_len += len;
 446                 vhost_net_tx_packet(net);
 447                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 448                         vhost_poll_queue(&vq->poll);
 449                         break;
 450                 }
 451         }
 452 out:
 453         mutex_unlock(&vq->mutex);
 454 }
 455
 456 static int peek_head_len(struct sock *sk)
 457 {
 458         struct sk_buff *head;
 459         int len = 0;
 460         unsigned long flags;
 461
 462         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 463         head = skb_peek(&sk->sk_receive_queue);
 464         if (likely(head)) {
 465                 len = head->len;
 466                 if (skb_vlan_tag_present(head))
 467                         len += VLAN_HLEN;
 468         }
 469
 470         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 471         return len;
 472 }
 473
 474 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
 475 {
 476         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 477         struct vhost_virtqueue *vq = &nvq->vq;
 478         unsigned long uninitialized_var(endtime);
 479         int len = peek_head_len(sk);
 480
 481         if (!len && vq->busyloop_timeout) {
 482                 /* Both tx vq and rx socket were polled here */
 483                 mutex_lock(&vq->mutex);
 484                 vhost_disable_notify(&net->dev, vq);
 485
 486                 preempt_disable();
 487                 endtime = busy_clock() + vq->busyloop_timeout;
 488
 489                 while (vhost_can_busy_poll(&net->dev, endtime) &&
 490                        skb_queue_empty(&sk->sk_receive_queue) &&
 491                        vhost_vq_avail_empty(&net->dev, vq))
 492                         cpu_relax_lowlatency();
 493
 494                 preempt_enable();
 495
 496                 if (vhost_enable_notify(&net->dev, vq))
 497                         vhost_poll_queue(&vq->poll);
 498                 mutex_unlock(&vq->mutex);
 499
 500                 len = peek_head_len(sk);
 501         }
 502
 503         return len;
 504 }
 505
 506 /* This is a multi-buffer version of vhost_get_desc, that works if
 507  *      vq has read descriptors only.
 508  * @vq          - the relevant virtqueue
 509  * @datalen     - data length we'll be reading
 510  * @iovcount    - returned count of io vectors we fill
 511  * @log         - vhost log
 512  * @log_num     - log offset
 513  * @quota       - headcount quota, 1 for big buffer
 514  *      returns number of buffer heads allocated, negative on error
 515  */
 516 static int get_rx_bufs(struct vhost_virtqueue *vq,
 517                        struct vring_used_elem *heads,
 518                        int datalen,
 519                        unsigned *iovcount,
 520                        struct vhost_log *log,
 521                        unsigned *log_num,
 522                        unsigned int quota)
 523 {
 524         unsigned int out, in;
 525         int seg = 0;
 526         int headcount = 0;
 527         unsigned d;
 528         int r, nlogs = 0;
 529         /* len is always initialized before use since we are always called with
 530          * datalen > 0.
 531          */
 532         u32 uninitialized_var(len);
 533
 534         while (datalen > 0 && headcount < quota) {
 535                 if (unlikely(seg >= UIO_MAXIOV)) {
 536                         r = -ENOBUFS;
 537                         goto err;
 538                 }
 539                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 540                                       ARRAY_SIZE(vq->iov) - seg, &out,
 541                                       &in, log, log_num);
 542                 if (unlikely(r < 0))
 543                         goto err;
 544
 545                 d = r;
 546                 if (d == vq->num) {
 547                         r = 0;
 548                         goto err;
 549                 }
 550                 if (unlikely(out || in <= 0)) {
 551                         vq_err(vq, "unexpected descriptor format for RX: "
 552                                 "out %d, in %d\n", out, in);
 553                         r = -EINVAL;
 554                         goto err;
 555                 }
 556                 if (unlikely(log)) {
 557                         nlogs += *log_num;
 558                         log += *log_num;
 559                 }
 560                 heads[headcount].id = cpu_to_vhost32(vq, d);
 561                 len = iov_length(vq->iov + seg, in);
 562                 heads[headcount].len = cpu_to_vhost32(vq, len);
 563                 datalen -= len;
 564                 ++headcount;
 565                 seg += in;
 566         }
 567         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 568         *iovcount = seg;
 569         if (unlikely(log))
 570                 *log_num = nlogs;
 571
 572         /* Detect overrun */
 573         if (unlikely(datalen > 0)) {
 574                 r = UIO_MAXIOV + 1;
 575                 goto err;
 576         }
 577         return headcount;
 578 err:
 579         vhost_discard_vq_desc(vq, headcount);
 580         return r;
 581 }
 582
 583 /* Expects to be always run from workqueue - which acts as
 584  * read-size critical section for our kind of RCU. */
 585 static void handle_rx(struct vhost_net *net)
 586 {
 587         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 588         struct vhost_virtqueue *vq = &nvq->vq;
 589         unsigned uninitialized_var(in), log;
 590         struct vhost_log *vq_log;
 591         struct msghdr msg = {
 592                 .msg_name = NULL,
 593                 .msg_namelen = 0,
 594                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 595                 .msg_controllen = 0,
 596                 .msg_flags = MSG_DONTWAIT,
 597         };
 598         struct virtio_net_hdr hdr = {
 599                 .flags = 0,
 600                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 601         };
 602         size_t total_len = 0;
 603         int err, mergeable;
 604         s16 headcount;
 605         size_t vhost_hlen, sock_hlen;
 606         size_t vhost_len, sock_len;
 607         struct socket *sock;
 608         struct iov_iter fixup;
 609         __virtio16 num_buffers;
 610
 611         mutex_lock(&vq->mutex);
 612         sock = vq->private_data;
 613         if (!sock)
 614                 goto out;
 615         vhost_disable_notify(&net->dev, vq);
 616
 617         vhost_hlen = nvq->vhost_hlen;
 618         sock_hlen = nvq->sock_hlen;
 619
 620         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 621                 vq->log : NULL;
 622         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 623
 624         while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
 625                 sock_len += sock_hlen;
 626                 vhost_len = sock_len + vhost_hlen;
 627                 headcount = get_rx_bufs(vq, vq->heads, vhost_len,
 628                                         &in, vq_log, &log,
 629                                         likely(mergeable) ? UIO_MAXIOV : 1);
 630                 /* On error, stop handling until the next kick. */
 631                 if (unlikely(headcount < 0))
 632                         break;
 633                 /* On overrun, truncate and discard */
 634                 if (unlikely(headcount > UIO_MAXIOV)) {
 635                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 636                         err = sock->ops->recvmsg(sock, &msg,
 637                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 638                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 639                         continue;
 640                 }
 641                 /* OK, now we need to know about added descriptors. */
 642                 if (!headcount) {
 643                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 644                                 /* They have slipped one in as we were
 645                                  * doing that: check again. */
 646                                 vhost_disable_notify(&net->dev, vq);
 647                                 continue;
 648                         }
 649                         /* Nothing new?  Wait for eventfd to tell us
 650                          * they refilled. */
 651                         break;
 652                 }
 653                 /* We don't need to be notified again. */
 654                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
 655                 fixup = msg.msg_iter;
 656                 if (unlikely((vhost_hlen))) {
 657                         /* We will supply the header ourselves
 658                          * TODO: support TSO.
 659                          */
 660                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
 661                 }
 662                 err = sock->ops->recvmsg(sock, &msg,
 663                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 664                 /* Userspace might have consumed the packet meanwhile:
 665                  * it's not supposed to do this usually, but might be hard
 666                  * to prevent. Discard data we got (if any) and keep going. */
 667                 if (unlikely(err != sock_len)) {
 668                         pr_debug("Discarded rx packet: "
 669                                  " len %d, expected %zd\n", err, sock_len);
 670                         vhost_discard_vq_desc(vq, headcount);
 671                         continue;
 672                 }
 673                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
 674                 if (unlikely(vhost_hlen)) {
 675                         if (copy_to_iter(&hdr, sizeof(hdr),
 676                                          &fixup) != sizeof(hdr)) {
 677                                 vq_err(vq, "Unable to write vnet_hdr "
 678                                        "at addr %p\n", vq->iov->iov_base);
 679                                 break;
 680                         }
 681                 } else {
 682                         /* Header came from socket; we'll need to patch
 683                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
 684                          */
 685                         iov_iter_advance(&fixup, sizeof(hdr));
 686                 }
 687                 /* TODO: Should check and handle checksum. */
 688
 689                 num_buffers = cpu_to_vhost16(vq, headcount);
 690                 if (likely(mergeable) &&
 691                     copy_to_iter(&num_buffers, sizeof num_buffers,
 692                                  &fixup) != sizeof num_buffers) {
 693                         vq_err(vq, "Failed num_buffers write");
 694                         vhost_discard_vq_desc(vq, headcount);
 695                         break;
 696                 }
 697                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 698                                             headcount);
 699                 if (unlikely(vq_log))
 700                         vhost_log_write(vq, vq_log, log, vhost_len);
 701                 total_len += vhost_len;
 702                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 703                         vhost_poll_queue(&vq->poll);
 704                         break;
 705                 }
 706         }
 707 out:
 708         mutex_unlock(&vq->mutex);
 709 }
 710
 711 static void handle_tx_kick(struct vhost_work *work)
 712 {
 713         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 714                                                   poll.work);
 715         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 716
 717         handle_tx(net);
 718 }
 719
 720 static void handle_rx_kick(struct vhost_work *work)
 721 {
 722         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 723                                                   poll.work);
 724         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 725
 726         handle_rx(net);
 727 }
 728
 729 static void handle_tx_net(struct vhost_work *work)
 730 {
 731         struct vhost_net *net = container_of(work, struct vhost_net,
 732                                              poll[VHOST_NET_VQ_TX].work);
 733         handle_tx(net);
 734 }
 735
 736 static void handle_rx_net(struct vhost_work *work)
 737 {
 738         struct vhost_net *net = container_of(work, struct vhost_net,
 739                                              poll[VHOST_NET_VQ_RX].work);
 740         handle_rx(net);
 741 }
 742
 743 static int vhost_net_open(struct inode *inode, struct file *f)
 744 {
 745         struct vhost_net *n;
 746         struct vhost_dev *dev;
 747         struct vhost_virtqueue **vqs;
 748         int i;
 749
 750         n = kmalloc(sizeof *n, GFP_KERNEL | __GFP_NOWARN | __GFP_REPEAT);
 751         if (!n) {
 752                 n = vmalloc(sizeof *n);
 753                 if (!n)
 754                         return -ENOMEM;
 755         }
 756         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 757         if (!vqs) {
 758                 kvfree(n);
 759                 return -ENOMEM;
 760         }
 761
 762         dev = &n->dev;
 763         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 764         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 765         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 766         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 767         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 768                 n->vqs[i].ubufs = NULL;
 769                 n->vqs[i].ubuf_info = NULL;
 770                 n->vqs[i].upend_idx = 0;
 771                 n->vqs[i].done_idx = 0;
 772                 n->vqs[i].vhost_hlen = 0;
 773                 n->vqs[i].sock_hlen = 0;
 774         }
 775         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 776
 777         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, POLLOUT, dev);
 778         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, POLLIN, dev);
 779
 780         f->private_data = n;
 781
 782         return 0;
 783 }
 784
 785 static void vhost_net_disable_vq(struct vhost_net *n,
 786                                  struct vhost_virtqueue *vq)
 787 {
 788         struct vhost_net_virtqueue *nvq =
 789                 container_of(vq, struct vhost_net_virtqueue, vq);
 790         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 791         if (!vq->private_data)
 792                 return;
 793         vhost_poll_stop(poll);
 794 }
 795
 796 static int vhost_net_enable_vq(struct vhost_net *n,
 797                                 struct vhost_virtqueue *vq)
 798 {
 799         struct vhost_net_virtqueue *nvq =
 800                 container_of(vq, struct vhost_net_virtqueue, vq);
 801         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 802         struct socket *sock;
 803
 804         sock = vq->private_data;
 805         if (!sock)
 806                 return 0;
 807
 808         return vhost_poll_start(poll, sock->file);
 809 }
 810
 811 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 812                                         struct vhost_virtqueue *vq)
 813 {
 814         struct socket *sock;
 815
 816         mutex_lock(&vq->mutex);
 817         sock = vq->private_data;
 818         vhost_net_disable_vq(n, vq);
 819         vq->private_data = NULL;
 820         mutex_unlock(&vq->mutex);
 821         return sock;
 822 }
 823
 824 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 825                            struct socket **rx_sock)
 826 {
 827         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 828         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 829 }
 830
 831 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 832 {
 833         vhost_poll_flush(n->poll + index);
 834         vhost_poll_flush(&n->vqs[index].vq.poll);
 835 }
 836
 837 static void vhost_net_flush(struct vhost_net *n)
 838 {
 839         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
 840         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
 841         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
 842                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 843                 n->tx_flush = true;
 844                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 845                 /* Wait for all lower device DMAs done. */
 846                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
 847                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 848                 n->tx_flush = false;
 849                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
 850                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
 851         }
 852 }
 853
 854 static int vhost_net_release(struct inode *inode, struct file *f)
 855 {
 856         struct vhost_net *n = f->private_data;
 857         struct socket *tx_sock;
 858         struct socket *rx_sock;
 859
 860         vhost_net_stop(n, &tx_sock, &rx_sock);
 861         vhost_net_flush(n);
 862         vhost_dev_stop(&n->dev);
 863         vhost_dev_cleanup(&n->dev, false);
 864         vhost_net_vq_reset(n);
 865         if (tx_sock)
 866                 sockfd_put(tx_sock);
 867         if (rx_sock)
 868                 sockfd_put(rx_sock);
 869         /* Make sure no callbacks are outstanding */
 870         synchronize_rcu_bh();
 871         /* We do an extra flush before freeing memory,
 872          * since jobs can re-queue themselves. */
 873         vhost_net_flush(n);
 874         kfree(n->dev.vqs);
 875         kvfree(n);
 876         return 0;
 877 }
 878
 879 static struct socket *get_raw_socket(int fd)
 880 {
 881         struct {
 882                 struct sockaddr_ll sa;
 883                 char  buf[MAX_ADDR_LEN];
 884         } uaddr;
 885         int uaddr_len = sizeof uaddr, r;
 886         struct socket *sock = sockfd_lookup(fd, &r);
 887
 888         if (!sock)
 889                 return ERR_PTR(-ENOTSOCK);
 890
 891         /* Parameter checking */
 892         if (sock->sk->sk_type != SOCK_RAW) {
 893                 r = -ESOCKTNOSUPPORT;
 894                 goto err;
 895         }
 896
 897         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa,
 898                                &uaddr_len, 0);
 899         if (r)
 900                 goto err;
 901
 902         if (uaddr.sa.sll_family != AF_PACKET) {
 903                 r = -EPFNOSUPPORT;
 904                 goto err;
 905         }
 906         return sock;
 907 err:
 908         sockfd_put(sock);
 909         return ERR_PTR(r);
 910 }
 911
 912 static struct socket *get_tap_socket(int fd)
 913 {
 914         struct file *file = fget(fd);
 915         struct socket *sock;
 916
 917         if (!file)
 918                 return ERR_PTR(-EBADF);
 919         sock = tun_get_socket(file);
 920         if (!IS_ERR(sock))
 921                 return sock;
 922         sock = macvtap_get_socket(file);
 923         if (IS_ERR(sock))
 924                 fput(file);
 925         return sock;
 926 }
 927
 928 static struct socket *get_socket(int fd)
 929 {
 930         struct socket *sock;
 931
 932         /* special case to disable backend */
 933         if (fd == -1)
 934                 return NULL;
 935         sock = get_raw_socket(fd);
 936         if (!IS_ERR(sock))
 937                 return sock;
 938         sock = get_tap_socket(fd);
 939         if (!IS_ERR(sock))
 940                 return sock;
 941         return ERR_PTR(-ENOTSOCK);
 942 }
 943
 944 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
 945 {
 946         struct socket *sock, *oldsock;
 947         struct vhost_virtqueue *vq;
 948         struct vhost_net_virtqueue *nvq;
 949         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
 950         int r;
 951
 952         mutex_lock(&n->dev.mutex);
 953         r = vhost_dev_check_owner(&n->dev);
 954         if (r)
 955                 goto err;
 956
 957         if (index >= VHOST_NET_VQ_MAX) {
 958                 r = -ENOBUFS;
 959                 goto err;
 960         }
 961         vq = &n->vqs[index].vq;
 962         nvq = &n->vqs[index];
 963         mutex_lock(&vq->mutex);
 964
 965         /* Verify that ring has been setup correctly. */
 966         if (!vhost_vq_access_ok(vq)) {
 967                 r = -EFAULT;
 968                 goto err_vq;
 969         }
 970         sock = get_socket(fd);
 971         if (IS_ERR(sock)) {
 972                 r = PTR_ERR(sock);
 973                 goto err_vq;
 974         }
 975
 976         /* start polling new socket */
 977         oldsock = vq->private_data;
 978         if (sock != oldsock) {
 979                 ubufs = vhost_net_ubuf_alloc(vq,
 980                                              sock && vhost_sock_zcopy(sock));
 981                 if (IS_ERR(ubufs)) {
 982                         r = PTR_ERR(ubufs);
 983                         goto err_ubufs;
 984                 }
 985
 986                 vhost_net_disable_vq(n, vq);
 987                 vq->private_data = sock;
 988                 r = vhost_vq_init_access(vq);
 989                 if (r)
 990                         goto err_used;
 991                 r = vhost_net_enable_vq(n, vq);
 992                 if (r)
 993                         goto err_used;
 994
 995                 oldubufs = nvq->ubufs;
 996                 nvq->ubufs = ubufs;
 997
 998                 n->tx_packets = 0;
 999                 n->tx_zcopy_err = 0;
1000                 n->tx_flush = false;
1001         }
1002
1003         mutex_unlock(&vq->mutex);
1004
1005         if (oldubufs) {
1006                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1007                 mutex_lock(&vq->mutex);
1008                 vhost_zerocopy_signal_used(n, vq);
1009                 mutex_unlock(&vq->mutex);
1010         }
1011
1012         if (oldsock) {
1013                 vhost_net_flush_vq(n, index);
1014                 sockfd_put(oldsock);
1015         }
1016
1017         mutex_unlock(&n->dev.mutex);
1018         return 0;
1019
1020 err_used:
1021         vq->private_data = oldsock;
1022         vhost_net_enable_vq(n, vq);
1023         if (ubufs)
1024                 vhost_net_ubuf_put_wait_and_free(ubufs);
1025 err_ubufs:
1026         sockfd_put(sock);
1027 err_vq:
1028         mutex_unlock(&vq->mutex);
1029 err:
1030         mutex_unlock(&n->dev.mutex);
1031         return r;
1032 }
1033
1034 static long vhost_net_reset_owner(struct vhost_net *n)
1035 {
1036         struct socket *tx_sock = NULL;
1037         struct socket *rx_sock = NULL;
1038         long err;
1039         struct vhost_memory *memory;
1040
1041         mutex_lock(&n->dev.mutex);
1042         err = vhost_dev_check_owner(&n->dev);
1043         if (err)
1044                 goto done;
1045         memory = vhost_dev_reset_owner_prepare();
1046         if (!memory) {
1047                 err = -ENOMEM;
1048                 goto done;
1049         }
1050         vhost_net_stop(n, &tx_sock, &rx_sock);
1051         vhost_net_flush(n);
1052         vhost_dev_reset_owner(&n->dev, memory);
1053         vhost_net_vq_reset(n);
1054 done:
1055         mutex_unlock(&n->dev.mutex);
1056         if (tx_sock)
1057                 sockfd_put(tx_sock);
1058         if (rx_sock)
1059                 sockfd_put(rx_sock);
1060         return err;
1061 }
1062
1063 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1064 {
1065         size_t vhost_hlen, sock_hlen, hdr_len;
1066         int i;
1067
1068         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1069                                (1ULL << VIRTIO_F_VERSION_1))) ?
1070                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1071                         sizeof(struct virtio_net_hdr);
1072         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1073                 /* vhost provides vnet_hdr */
1074                 vhost_hlen = hdr_len;
1075                 sock_hlen = 0;
1076         } else {
1077                 /* socket provides vnet_hdr */
1078                 vhost_hlen = 0;
1079                 sock_hlen = hdr_len;
1080         }
1081         mutex_lock(&n->dev.mutex);
1082         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1083             !vhost_log_access_ok(&n->dev)) {
1084                 mutex_unlock(&n->dev.mutex);
1085                 return -EFAULT;
1086         }
1087         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1088                 mutex_lock(&n->vqs[i].vq.mutex);
1089                 n->vqs[i].vq.acked_features = features;
1090                 n->vqs[i].vhost_hlen = vhost_hlen;
1091                 n->vqs[i].sock_hlen = sock_hlen;
1092                 mutex_unlock(&n->vqs[i].vq.mutex);
1093         }
1094         mutex_unlock(&n->dev.mutex);
1095         return 0;
1096 }
1097
1098 static long vhost_net_set_owner(struct vhost_net *n)
1099 {
1100         int r;
1101
1102         mutex_lock(&n->dev.mutex);
1103         if (vhost_dev_has_owner(&n->dev)) {
1104                 r = -EBUSY;
1105                 goto out;
1106         }
1107         r = vhost_net_set_ubuf_info(n);
1108         if (r)
1109                 goto out;
1110         r = vhost_dev_set_owner(&n->dev);
1111         if (r)
1112                 vhost_net_clear_ubuf_info(n);
1113         vhost_net_flush(n);
1114 out:
1115         mutex_unlock(&n->dev.mutex);
1116         return r;
1117 }
1118
1119 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1120                             unsigned long arg)
1121 {
1122         struct vhost_net *n = f->private_data;
1123         void __user *argp = (void __user *)arg;
1124         u64 __user *featurep = argp;
1125         struct vhost_vring_file backend;
1126         u64 features;
1127         int r;
1128
1129         switch (ioctl) {
1130         case VHOST_NET_SET_BACKEND:
1131                 if (copy_from_user(&backend, argp, sizeof backend))
1132                         return -EFAULT;
1133                 return vhost_net_set_backend(n, backend.index, backend.fd);
1134         case VHOST_GET_FEATURES:
1135                 features = VHOST_NET_FEATURES;
1136                 if (copy_to_user(featurep, &features, sizeof features))
1137                         return -EFAULT;
1138                 return 0;
1139         case VHOST_SET_FEATURES:
1140                 if (copy_from_user(&features, featurep, sizeof features))
1141                         return -EFAULT;
1142                 if (features & ~VHOST_NET_FEATURES)
1143                         return -EOPNOTSUPP;
1144                 return vhost_net_set_features(n, features);
1145         case VHOST_RESET_OWNER:
1146                 return vhost_net_reset_owner(n);
1147         case VHOST_SET_OWNER:
1148                 return vhost_net_set_owner(n);
1149         default:
1150                 mutex_lock(&n->dev.mutex);
1151                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1152                 if (r == -ENOIOCTLCMD)
1153                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1154                 else
1155                         vhost_net_flush(n);
1156                 mutex_unlock(&n->dev.mutex);
1157                 return r;
1158         }
1159 }
1160
1161 #ifdef CONFIG_COMPAT
1162 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1163                                    unsigned long arg)
1164 {
1165         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1166 }
1167 #endif
1168
1169 static const struct file_operations vhost_net_fops = {
1170         .owner          = THIS_MODULE,
1171         .release        = vhost_net_release,
1172         .unlocked_ioctl = vhost_net_ioctl,
1173 #ifdef CONFIG_COMPAT
1174         .compat_ioctl   = vhost_net_compat_ioctl,
1175 #endif
1176         .open           = vhost_net_open,
1177         .llseek         = noop_llseek,
1178 };
1179
1180 static struct miscdevice vhost_net_misc = {
1181         .minor = VHOST_NET_MINOR,
1182         .name = "vhost-net",
1183         .fops = &vhost_net_fops,
1184 };
1185
1186 static int vhost_net_init(void)
1187 {
1188         if (experimental_zcopytx)
1189                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1190         return misc_register(&vhost_net_misc);
1191 }
1192 module_init(vhost_net_init);
1193
1194 static void vhost_net_exit(void)
1195 {
1196         misc_deregister(&vhost_net_misc);
1197 }
1198 module_exit(vhost_net_exit);
1199
1200 MODULE_VERSION("0.0.1");
1201 MODULE_LICENSE("GPL v2");
1202 MODULE_AUTHOR("Michael S. Tsirkin");
1203 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1204 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1205 MODULE_ALIAS("devname:vhost-net");