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/sched/clock.h>
  21 #include <linux/sched/signal.h>
  22 #include <linux/vmalloc.h>
  23
  24 #include <linux/net.h>
  25 #include <linux/if_packet.h>
  26 #include <linux/if_arp.h>
  27 #include <linux/if_tun.h>
  28 #include <linux/if_macvlan.h>
  29 #include <linux/if_tap.h>
  30 #include <linux/if_vlan.h>
  31 #include <linux/skb_array.h>
  32 #include <linux/skbuff.h>
  33
  34 #include <net/sock.h>
  35 #include <net/xdp.h>
  36
  37 #include "vhost.h"
  38
  39 static int experimental_zcopytx = 0;
  40 module_param(experimental_zcopytx, int, 0444);
  41 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  42                                        " 1 -Enable; 0 - Disable");
  43
  44 /* Max number of bytes transferred before requeueing the job.
  45  * Using this limit prevents one virtqueue from starving others. */
  46 #define VHOST_NET_WEIGHT 0x80000
  47
  48 /* Max number of packets transferred before requeueing the job.
  49  * Using this limit prevents one virtqueue from starving others with small
  50  * pkts.
  51  */
  52 #define VHOST_NET_PKT_WEIGHT 256
  53
  54 /* MAX number of TX used buffers for outstanding zerocopy */
  55 #define VHOST_MAX_PEND 128
  56 #define VHOST_GOODCOPY_LEN 256
  57
  58 /*
  59  * For transmit, used buffer len is unused; we override it to track buffer
  60  * status internally; used for zerocopy tx only.
  61  */
  62 /* Lower device DMA failed */
  63 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  64 /* Lower device DMA done */
  65 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  66 /* Lower device DMA in progress */
  67 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  68 /* Buffer unused */
  69 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  70
  71 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  72
  73 enum {
  74         VHOST_NET_FEATURES = VHOST_FEATURES |
  75                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  76                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  77                          (1ULL << VIRTIO_F_IOMMU_PLATFORM)
  78 };
  79
  80 enum {
  81         VHOST_NET_BACKEND_FEATURES = (1ULL << VHOST_BACKEND_F_IOTLB_MSG_V2)
  82 };
  83
  84 enum {
  85         VHOST_NET_VQ_RX = 0,
  86         VHOST_NET_VQ_TX = 1,
  87         VHOST_NET_VQ_MAX = 2,
  88 };
  89
  90 struct vhost_net_ubuf_ref {
  91         /* refcount follows semantics similar to kref:
  92          *  0: object is released
  93          *  1: no outstanding ubufs
  94          * >1: outstanding ubufs
  95          */
  96         atomic_t refcount;
  97         wait_queue_head_t wait;
  98         struct vhost_virtqueue *vq;
  99 };
 100
 101 #define VHOST_NET_BATCH 64
 102 struct vhost_net_buf {
 103         void **queue;
 104         int tail;
 105         int head;
 106 };
 107
 108 struct vhost_net_virtqueue {
 109         struct vhost_virtqueue vq;
 110         size_t vhost_hlen;
 111         size_t sock_hlen;
 112         /* vhost zerocopy support fields below: */
 113         /* last used idx for outstanding DMA zerocopy buffers */
 114         int upend_idx;
 115         /* For TX, first used idx for DMA done zerocopy buffers
 116          * For RX, number of batched heads
 117          */
 118         int done_idx;
 119         /* an array of userspace buffers info */
 120         struct ubuf_info *ubuf_info;
 121         /* Reference counting for outstanding ubufs.
 122          * Protected by vq mutex. Writers must also take device mutex. */
 123         struct vhost_net_ubuf_ref *ubufs;
 124         struct ptr_ring *rx_ring;
 125         struct vhost_net_buf rxq;
 126 };
 127
 128 struct vhost_net {
 129         struct vhost_dev dev;
 130         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 131         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 132         /* Number of TX recently submitted.
 133          * Protected by tx vq lock. */
 134         unsigned tx_packets;
 135         /* Number of times zerocopy TX recently failed.
 136          * Protected by tx vq lock. */
 137         unsigned tx_zcopy_err;
 138         /* Flush in progress. Protected by tx vq lock. */
 139         bool tx_flush;
 140 };
 141
 142 static unsigned vhost_net_zcopy_mask __read_mostly;
 143
 144 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 145 {
 146         if (rxq->tail != rxq->head)
 147                 return rxq->queue[rxq->head];
 148         else
 149                 return NULL;
 150 }
 151
 152 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 153 {
 154         return rxq->tail - rxq->head;
 155 }
 156
 157 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 158 {
 159         return rxq->tail == rxq->head;
 160 }
 161
 162 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 163 {
 164         void *ret = vhost_net_buf_get_ptr(rxq);
 165         ++rxq->head;
 166         return ret;
 167 }
 168
 169 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 170 {
 171         struct vhost_net_buf *rxq = &nvq->rxq;
 172
 173         rxq->head = 0;
 174         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 175                                               VHOST_NET_BATCH);
 176         return rxq->tail;
 177 }
 178
 179 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 180 {
 181         struct vhost_net_buf *rxq = &nvq->rxq;
 182
 183         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 184                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 185                                    vhost_net_buf_get_size(rxq),
 186                                    tun_ptr_free);
 187                 rxq->head = rxq->tail = 0;
 188         }
 189 }
 190
 191 static int vhost_net_buf_peek_len(void *ptr)
 192 {
 193         if (tun_is_xdp_frame(ptr)) {
 194                 struct xdp_frame *xdpf = tun_ptr_to_xdp(ptr);
 195
 196                 return xdpf->len;
 197         }
 198
 199         return __skb_array_len_with_tag(ptr);
 200 }
 201
 202 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 203 {
 204         struct vhost_net_buf *rxq = &nvq->rxq;
 205
 206         if (!vhost_net_buf_is_empty(rxq))
 207                 goto out;
 208
 209         if (!vhost_net_buf_produce(nvq))
 210                 return 0;
 211
 212 out:
 213         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 214 }
 215
 216 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 217 {
 218         rxq->head = rxq->tail = 0;
 219 }
 220
 221 static void vhost_net_enable_zcopy(int vq)
 222 {
 223         vhost_net_zcopy_mask |= 0x1 << vq;
 224 }
 225
 226 static struct vhost_net_ubuf_ref *
 227 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 228 {
 229         struct vhost_net_ubuf_ref *ubufs;
 230         /* No zero copy backend? Nothing to count. */
 231         if (!zcopy)
 232                 return NULL;
 233         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 234         if (!ubufs)
 235                 return ERR_PTR(-ENOMEM);
 236         atomic_set(&ubufs->refcount, 1);
 237         init_waitqueue_head(&ubufs->wait);
 238         ubufs->vq = vq;
 239         return ubufs;
 240 }
 241
 242 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 243 {
 244         int r = atomic_sub_return(1, &ubufs->refcount);
 245         if (unlikely(!r))
 246                 wake_up(&ubufs->wait);
 247         return r;
 248 }
 249
 250 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 251 {
 252         vhost_net_ubuf_put(ubufs);
 253         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 254 }
 255
 256 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 257 {
 258         vhost_net_ubuf_put_and_wait(ubufs);
 259         kfree(ubufs);
 260 }
 261
 262 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 263 {
 264         int i;
 265
 266         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 267                 kfree(n->vqs[i].ubuf_info);
 268                 n->vqs[i].ubuf_info = NULL;
 269         }
 270 }
 271
 272 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 273 {
 274         bool zcopy;
 275         int i;
 276
 277         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 278                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 279                 if (!zcopy)
 280                         continue;
 281                 n->vqs[i].ubuf_info =
 282                         kmalloc_array(UIO_MAXIOV,
 283                                       sizeof(*n->vqs[i].ubuf_info),
 284                                       GFP_KERNEL);
 285                 if  (!n->vqs[i].ubuf_info)
 286                         goto err;
 287         }
 288         return 0;
 289
 290 err:
 291         vhost_net_clear_ubuf_info(n);
 292         return -ENOMEM;
 293 }
 294
 295 static void vhost_net_vq_reset(struct vhost_net *n)
 296 {
 297         int i;
 298
 299         vhost_net_clear_ubuf_info(n);
 300
 301         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 302                 n->vqs[i].done_idx = 0;
 303                 n->vqs[i].upend_idx = 0;
 304                 n->vqs[i].ubufs = NULL;
 305                 n->vqs[i].vhost_hlen = 0;
 306                 n->vqs[i].sock_hlen = 0;
 307                 vhost_net_buf_init(&n->vqs[i].rxq);
 308         }
 309
 310 }
 311
 312 static void vhost_net_tx_packet(struct vhost_net *net)
 313 {
 314         ++net->tx_packets;
 315         if (net->tx_packets < 1024)
 316                 return;
 317         net->tx_packets = 0;
 318         net->tx_zcopy_err = 0;
 319 }
 320
 321 static void vhost_net_tx_err(struct vhost_net *net)
 322 {
 323         ++net->tx_zcopy_err;
 324 }
 325
 326 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 327 {
 328         /* TX flush waits for outstanding DMAs to be done.
 329          * Don't start new DMAs.
 330          */
 331         return !net->tx_flush &&
 332                 net->tx_packets / 64 >= net->tx_zcopy_err;
 333 }
 334
 335 static bool vhost_sock_zcopy(struct socket *sock)
 336 {
 337         return unlikely(experimental_zcopytx) &&
 338                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 339 }
 340
 341 /* In case of DMA done not in order in lower device driver for some reason.
 342  * upend_idx is used to track end of used idx, done_idx is used to track head
 343  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 344  * guest used idx.
 345  */
 346 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 347                                        struct vhost_virtqueue *vq)
 348 {
 349         struct vhost_net_virtqueue *nvq =
 350                 container_of(vq, struct vhost_net_virtqueue, vq);
 351         int i, add;
 352         int j = 0;
 353
 354         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 355                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 356                         vhost_net_tx_err(net);
 357                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 358                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 359                         ++j;
 360                 } else
 361                         break;
 362         }
 363         while (j) {
 364                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 365                 vhost_add_used_and_signal_n(vq->dev, vq,
 366                                             &vq->heads[nvq->done_idx], add);
 367                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 368                 j -= add;
 369         }
 370 }
 371
 372 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 373 {
 374         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 375         struct vhost_virtqueue *vq = ubufs->vq;
 376         int cnt;
 377
 378         rcu_read_lock_bh();
 379
 380         /* set len to mark this desc buffers done DMA */
 381         vq->heads[ubuf->desc].len = success ?
 382                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 383         cnt = vhost_net_ubuf_put(ubufs);
 384
 385         /*
 386          * Trigger polling thread if guest stopped submitting new buffers:
 387          * in this case, the refcount after decrement will eventually reach 1.
 388          * We also trigger polling periodically after each 16 packets
 389          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 390          * less than 10% of times).
 391          */
 392         if (cnt <= 1 || !(cnt % 16))
 393                 vhost_poll_queue(&vq->poll);
 394
 395         rcu_read_unlock_bh();
 396 }
 397
 398 static inline unsigned long busy_clock(void)
 399 {
 400         return local_clock() >> 10;
 401 }
 402
 403 static bool vhost_can_busy_poll(unsigned long endtime)
 404 {
 405         return likely(!need_resched() && !time_after(busy_clock(), endtime) &&
 406                       !signal_pending(current));
 407 }
 408
 409 static void vhost_net_disable_vq(struct vhost_net *n,
 410                                  struct vhost_virtqueue *vq)
 411 {
 412         struct vhost_net_virtqueue *nvq =
 413                 container_of(vq, struct vhost_net_virtqueue, vq);
 414         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 415         if (!vq->private_data)
 416                 return;
 417         vhost_poll_stop(poll);
 418 }
 419
 420 static int vhost_net_enable_vq(struct vhost_net *n,
 421                                 struct vhost_virtqueue *vq)
 422 {
 423         struct vhost_net_virtqueue *nvq =
 424                 container_of(vq, struct vhost_net_virtqueue, vq);
 425         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 426         struct socket *sock;
 427
 428         sock = vq->private_data;
 429         if (!sock)
 430                 return 0;
 431
 432         return vhost_poll_start(poll, sock->file);
 433 }
 434
 435 static void vhost_net_signal_used(struct vhost_net_virtqueue *nvq)
 436 {
 437         struct vhost_virtqueue *vq = &nvq->vq;
 438         struct vhost_dev *dev = vq->dev;
 439
 440         if (!nvq->done_idx)
 441                 return;
 442
 443         vhost_add_used_and_signal_n(dev, vq, vq->heads, nvq->done_idx);
 444         nvq->done_idx = 0;
 445 }
 446
 447 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 448                                     struct vhost_net_virtqueue *nvq,
 449                                     unsigned int *out_num, unsigned int *in_num,
 450                                     bool *busyloop_intr)
 451 {
 452         struct vhost_virtqueue *vq = &nvq->vq;
 453         unsigned long uninitialized_var(endtime);
 454         int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 455                                   out_num, in_num, NULL, NULL);
 456
 457         if (r == vq->num && vq->busyloop_timeout) {
 458                 if (!vhost_sock_zcopy(vq->private_data))
 459                         vhost_net_signal_used(nvq);
 460                 preempt_disable();
 461                 endtime = busy_clock() + vq->busyloop_timeout;
 462                 while (vhost_can_busy_poll(endtime)) {
 463                         if (vhost_has_work(vq->dev)) {
 464                                 *busyloop_intr = true;
 465                                 break;
 466                         }
 467                         if (!vhost_vq_avail_empty(vq->dev, vq))
 468                                 break;
 469                         cpu_relax();
 470                 }
 471                 preempt_enable();
 472                 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 473                                       out_num, in_num, NULL, NULL);
 474         }
 475
 476         return r;
 477 }
 478
 479 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 480 {
 481         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 482         struct vhost_virtqueue *vq = &nvq->vq;
 483
 484         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 485                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 486 }
 487
 488 static size_t init_iov_iter(struct vhost_virtqueue *vq, struct iov_iter *iter,
 489                             size_t hdr_size, int out)
 490 {
 491         /* Skip header. TODO: support TSO. */
 492         size_t len = iov_length(vq->iov, out);
 493
 494         iov_iter_init(iter, WRITE, vq->iov, out, len);
 495         iov_iter_advance(iter, hdr_size);
 496
 497         return iov_iter_count(iter);
 498 }
 499
 500 static int get_tx_bufs(struct vhost_net *net,
 501                        struct vhost_net_virtqueue *nvq,
 502                        struct msghdr *msg,
 503                        unsigned int *out, unsigned int *in,
 504                        size_t *len, bool *busyloop_intr)
 505 {
 506         struct vhost_virtqueue *vq = &nvq->vq;
 507         int ret;
 508
 509         ret = vhost_net_tx_get_vq_desc(net, nvq, out, in, busyloop_intr);
 510
 511         if (ret < 0 || ret == vq->num)
 512                 return ret;
 513
 514         if (*in) {
 515                 vq_err(vq, "Unexpected descriptor format for TX: out %d, int %d\n",
 516                         *out, *in);
 517                 return -EFAULT;
 518         }
 519
 520         /* Sanity check */
 521         *len = init_iov_iter(vq, &msg->msg_iter, nvq->vhost_hlen, *out);
 522         if (*len == 0) {
 523                 vq_err(vq, "Unexpected header len for TX: %zd expected %zd\n",
 524                         *len, nvq->vhost_hlen);
 525                 return -EFAULT;
 526         }
 527
 528         return ret;
 529 }
 530
 531 static bool tx_can_batch(struct vhost_virtqueue *vq, size_t total_len)
 532 {
 533         return total_len < VHOST_NET_WEIGHT &&
 534                !vhost_vq_avail_empty(vq->dev, vq);
 535 }
 536
 537 static void handle_tx_copy(struct vhost_net *net, struct socket *sock)
 538 {
 539         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 540         struct vhost_virtqueue *vq = &nvq->vq;
 541         unsigned out, in;
 542         int head;
 543         struct msghdr msg = {
 544                 .msg_name = NULL,
 545                 .msg_namelen = 0,
 546                 .msg_control = NULL,
 547                 .msg_controllen = 0,
 548                 .msg_flags = MSG_DONTWAIT,
 549         };
 550         size_t len, total_len = 0;
 551         int err;
 552         int sent_pkts = 0;
 553
 554         do {
 555                 bool busyloop_intr = false;
 556
 557                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 558                                    &busyloop_intr);
 559                 /* On error, stop handling until the next kick. */
 560                 if (unlikely(head < 0))
 561                         break;
 562                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 563                 if (head == vq->num) {
 564                         if (unlikely(busyloop_intr)) {
 565                                 vhost_poll_queue(&vq->poll);
 566                         } else if (unlikely(vhost_enable_notify(&net->dev,
 567                                                                 vq))) {
 568                                 vhost_disable_notify(&net->dev, vq);
 569                                 continue;
 570                         }
 571                         break;
 572                 }
 573
 574                 vq->heads[nvq->done_idx].id = cpu_to_vhost32(vq, head);
 575                 vq->heads[nvq->done_idx].len = 0;
 576
 577                 total_len += len;
 578                 if (tx_can_batch(vq, total_len))
 579                         msg.msg_flags |= MSG_MORE;
 580                 else
 581                         msg.msg_flags &= ~MSG_MORE;
 582
 583                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 584                 err = sock->ops->sendmsg(sock, &msg, len);
 585                 if (unlikely(err < 0)) {
 586                         vhost_discard_vq_desc(vq, 1);
 587                         vhost_net_enable_vq(net, vq);
 588                         break;
 589                 }
 590                 if (err != len)
 591                         pr_debug("Truncated TX packet: len %d != %zd\n",
 592                                  err, len);
 593                 if (++nvq->done_idx >= VHOST_NET_BATCH)
 594                         vhost_net_signal_used(nvq);
 595         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 596
 597         vhost_net_signal_used(nvq);
 598 }
 599
 600 static void handle_tx_zerocopy(struct vhost_net *net, struct socket *sock)
 601 {
 602         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 603         struct vhost_virtqueue *vq = &nvq->vq;
 604         unsigned out, in;
 605         int head;
 606         struct msghdr msg = {
 607                 .msg_name = NULL,
 608                 .msg_namelen = 0,
 609                 .msg_control = NULL,
 610                 .msg_controllen = 0,
 611                 .msg_flags = MSG_DONTWAIT,
 612         };
 613         size_t len, total_len = 0;
 614         int err;
 615         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 616         struct ubuf_info *ubuf;
 617         bool zcopy_used;
 618         int sent_pkts = 0;
 619
 620         do {
 621                 bool busyloop_intr;
 622
 623                 /* Release DMAs done buffers first */
 624                 vhost_zerocopy_signal_used(net, vq);
 625
 626                 busyloop_intr = false;
 627                 head = get_tx_bufs(net, nvq, &msg, &out, &in, &len,
 628                                    &busyloop_intr);
 629                 /* On error, stop handling until the next kick. */
 630                 if (unlikely(head < 0))
 631                         break;
 632                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 633                 if (head == vq->num) {
 634                         if (unlikely(busyloop_intr)) {
 635                                 vhost_poll_queue(&vq->poll);
 636                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 637                                 vhost_disable_notify(&net->dev, vq);
 638                                 continue;
 639                         }
 640                         break;
 641                 }
 642
 643                 zcopy_used = len >= VHOST_GOODCOPY_LEN
 644                              && !vhost_exceeds_maxpend(net)
 645                              && vhost_net_tx_select_zcopy(net);
 646
 647                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 648                 if (zcopy_used) {
 649                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 650                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 651                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 652                         ubuf->callback = vhost_zerocopy_callback;
 653                         ubuf->ctx = nvq->ubufs;
 654                         ubuf->desc = nvq->upend_idx;
 655                         refcount_set(&ubuf->refcnt, 1);
 656                         msg.msg_control = ubuf;
 657                         msg.msg_controllen = sizeof(ubuf);
 658                         ubufs = nvq->ubufs;
 659                         atomic_inc(&ubufs->refcount);
 660                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 661                 } else {
 662                         msg.msg_control = NULL;
 663                         ubufs = NULL;
 664                 }
 665                 total_len += len;
 666                 if (tx_can_batch(vq, total_len) &&
 667                     likely(!vhost_exceeds_maxpend(net))) {
 668                         msg.msg_flags |= MSG_MORE;
 669                 } else {
 670                         msg.msg_flags &= ~MSG_MORE;
 671                 }
 672
 673                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 674                 err = sock->ops->sendmsg(sock, &msg, len);
 675                 if (unlikely(err < 0)) {
 676                         if (zcopy_used) {
 677                                 if (vq->heads[ubuf->desc].len == VHOST_DMA_IN_PROGRESS)
 678                                         vhost_net_ubuf_put(ubufs);
 679                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 680                                         % UIO_MAXIOV;
 681                         }
 682                         vhost_discard_vq_desc(vq, 1);
 683                         vhost_net_enable_vq(net, vq);
 684                         break;
 685                 }
 686                 if (err != len)
 687                         pr_debug("Truncated TX packet: "
 688                                  " len %d != %zd\n", err, len);
 689                 if (!zcopy_used)
 690                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 691                 else
 692                         vhost_zerocopy_signal_used(net, vq);
 693                 vhost_net_tx_packet(net);
 694         } while (likely(!vhost_exceeds_weight(vq, ++sent_pkts, total_len)));
 695 }
 696
 697 /* Expects to be always run from workqueue - which acts as
 698  * read-size critical section for our kind of RCU. */
 699 static void handle_tx(struct vhost_net *net)
 700 {
 701         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 702         struct vhost_virtqueue *vq = &nvq->vq;
 703         struct socket *sock;
 704
 705         mutex_lock(&vq->mutex);
 706         sock = vq->private_data;
 707         if (!sock)
 708                 goto out;
 709
 710         if (!vq_iotlb_prefetch(vq))
 711                 goto out;
 712
 713         vhost_disable_notify(&net->dev, vq);
 714         vhost_net_disable_vq(net, vq);
 715
 716         if (vhost_sock_zcopy(sock))
 717                 handle_tx_zerocopy(net, sock);
 718         else
 719                 handle_tx_copy(net, sock);
 720
 721 out:
 722         mutex_unlock(&vq->mutex);
 723 }
 724
 725 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 726 {
 727         struct sk_buff *head;
 728         int len = 0;
 729         unsigned long flags;
 730
 731         if (rvq->rx_ring)
 732                 return vhost_net_buf_peek(rvq);
 733
 734         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 735         head = skb_peek(&sk->sk_receive_queue);
 736         if (likely(head)) {
 737                 len = head->len;
 738                 if (skb_vlan_tag_present(head))
 739                         len += VLAN_HLEN;
 740         }
 741
 742         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 743         return len;
 744 }
 745
 746 static int sk_has_rx_data(struct sock *sk)
 747 {
 748         struct socket *sock = sk->sk_socket;
 749
 750         if (sock->ops->peek_len)
 751                 return sock->ops->peek_len(sock);
 752
 753         return skb_queue_empty(&sk->sk_receive_queue);
 754 }
 755
 756 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk,
 757                                       bool *busyloop_intr)
 758 {
 759         struct vhost_net_virtqueue *rnvq = &net->vqs[VHOST_NET_VQ_RX];
 760         struct vhost_net_virtqueue *tnvq = &net->vqs[VHOST_NET_VQ_TX];
 761         struct vhost_virtqueue *rvq = &rnvq->vq;
 762         struct vhost_virtqueue *tvq = &tnvq->vq;
 763         unsigned long uninitialized_var(endtime);
 764         int len = peek_head_len(rnvq, sk);
 765
 766         if (!len && tvq->busyloop_timeout) {
 767                 /* Flush batched heads first */
 768                 vhost_net_signal_used(rnvq);
 769                 /* Both tx vq and rx socket were polled here */
 770                 mutex_lock_nested(&tvq->mutex, 1);
 771                 vhost_disable_notify(&net->dev, tvq);
 772
 773                 preempt_disable();
 774                 endtime = busy_clock() + tvq->busyloop_timeout;
 775
 776                 while (vhost_can_busy_poll(endtime)) {
 777                         if (vhost_has_work(&net->dev)) {
 778                                 *busyloop_intr = true;
 779                                 break;
 780                         }
 781                         if ((sk_has_rx_data(sk) &&
 782                              !vhost_vq_avail_empty(&net->dev, rvq)) ||
 783                             !vhost_vq_avail_empty(&net->dev, tvq))
 784                                 break;
 785                         cpu_relax();
 786                 }
 787
 788                 preempt_enable();
 789
 790                 if (!vhost_vq_avail_empty(&net->dev, tvq)) {
 791                         vhost_poll_queue(&tvq->poll);
 792                 } else if (unlikely(vhost_enable_notify(&net->dev, tvq))) {
 793                         vhost_disable_notify(&net->dev, tvq);
 794                         vhost_poll_queue(&tvq->poll);
 795                 }
 796
 797                 mutex_unlock(&tvq->mutex);
 798
 799                 len = peek_head_len(rnvq, sk);
 800         }
 801
 802         return len;
 803 }
 804
 805 /* This is a multi-buffer version of vhost_get_desc, that works if
 806  *      vq has read descriptors only.
 807  * @vq          - the relevant virtqueue
 808  * @datalen     - data length we'll be reading
 809  * @iovcount    - returned count of io vectors we fill
 810  * @log         - vhost log
 811  * @log_num     - log offset
 812  * @quota       - headcount quota, 1 for big buffer
 813  *      returns number of buffer heads allocated, negative on error
 814  */
 815 static int get_rx_bufs(struct vhost_virtqueue *vq,
 816                        struct vring_used_elem *heads,
 817                        int datalen,
 818                        unsigned *iovcount,
 819                        struct vhost_log *log,
 820                        unsigned *log_num,
 821                        unsigned int quota)
 822 {
 823         unsigned int out, in;
 824         int seg = 0;
 825         int headcount = 0;
 826         unsigned d;
 827         int r, nlogs = 0;
 828         /* len is always initialized before use since we are always called with
 829          * datalen > 0.
 830          */
 831         u32 uninitialized_var(len);
 832
 833         while (datalen > 0 && headcount < quota) {
 834                 if (unlikely(seg >= UIO_MAXIOV)) {
 835                         r = -ENOBUFS;
 836                         goto err;
 837                 }
 838                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 839                                       ARRAY_SIZE(vq->iov) - seg, &out,
 840                                       &in, log, log_num);
 841                 if (unlikely(r < 0))
 842                         goto err;
 843
 844                 d = r;
 845                 if (d == vq->num) {
 846                         r = 0;
 847                         goto err;
 848                 }
 849                 if (unlikely(out || in <= 0)) {
 850                         vq_err(vq, "unexpected descriptor format for RX: "
 851                                 "out %d, in %d\n", out, in);
 852                         r = -EINVAL;
 853                         goto err;
 854                 }
 855                 if (unlikely(log)) {
 856                         nlogs += *log_num;
 857                         log += *log_num;
 858                 }
 859                 heads[headcount].id = cpu_to_vhost32(vq, d);
 860                 len = iov_length(vq->iov + seg, in);
 861                 heads[headcount].len = cpu_to_vhost32(vq, len);
 862                 datalen -= len;
 863                 ++headcount;
 864                 seg += in;
 865         }
 866         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 867         *iovcount = seg;
 868         if (unlikely(log))
 869                 *log_num = nlogs;
 870
 871         /* Detect overrun */
 872         if (unlikely(datalen > 0)) {
 873                 r = UIO_MAXIOV + 1;
 874                 goto err;
 875         }
 876         return headcount;
 877 err:
 878         vhost_discard_vq_desc(vq, headcount);
 879         return r;
 880 }
 881
 882 /* Expects to be always run from workqueue - which acts as
 883  * read-size critical section for our kind of RCU. */
 884 static void handle_rx(struct vhost_net *net)
 885 {
 886         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 887         struct vhost_virtqueue *vq = &nvq->vq;
 888         unsigned uninitialized_var(in), log;
 889         struct vhost_log *vq_log;
 890         struct msghdr msg = {
 891                 .msg_name = NULL,
 892                 .msg_namelen = 0,
 893                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 894                 .msg_controllen = 0,
 895                 .msg_flags = MSG_DONTWAIT,
 896         };
 897         struct virtio_net_hdr hdr = {
 898                 .flags = 0,
 899                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 900         };
 901         size_t total_len = 0;
 902         int err, mergeable;
 903         s16 headcount;
 904         size_t vhost_hlen, sock_hlen;
 905         size_t vhost_len, sock_len;
 906         bool busyloop_intr = false;
 907         struct socket *sock;
 908         struct iov_iter fixup;
 909         __virtio16 num_buffers;
 910         int recv_pkts = 0;
 911
 912         mutex_lock_nested(&vq->mutex, 0);
 913         sock = vq->private_data;
 914         if (!sock)
 915                 goto out;
 916
 917         if (!vq_iotlb_prefetch(vq))
 918                 goto out;
 919
 920         vhost_disable_notify(&net->dev, vq);
 921         vhost_net_disable_vq(net, vq);
 922
 923         vhost_hlen = nvq->vhost_hlen;
 924         sock_hlen = nvq->sock_hlen;
 925
 926         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 927                 vq->log : NULL;
 928         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 929
 930         do {
 931                 sock_len = vhost_net_rx_peek_head_len(net, sock->sk,
 932                                                       &busyloop_intr);
 933                 if (!sock_len)
 934                         break;
 935                 sock_len += sock_hlen;
 936                 vhost_len = sock_len + vhost_hlen;
 937                 headcount = get_rx_bufs(vq, vq->heads + nvq->done_idx,
 938                                         vhost_len, &in, vq_log, &log,
 939                                         likely(mergeable) ? UIO_MAXIOV : 1);
 940                 /* On error, stop handling until the next kick. */
 941                 if (unlikely(headcount < 0))
 942                         goto out;
 943                 /* OK, now we need to know about added descriptors. */
 944                 if (!headcount) {
 945                         if (unlikely(busyloop_intr)) {
 946                                 vhost_poll_queue(&vq->poll);
 947                         } else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 948                                 /* They have slipped one in as we were
 949                                  * doing that: check again. */
 950                                 vhost_disable_notify(&net->dev, vq);
 951                                 continue;
 952                         }
 953                         /* Nothing new?  Wait for eventfd to tell us
 954                          * they refilled. */
 955                         goto out;
 956                 }
 957                 busyloop_intr = false;
 958                 if (nvq->rx_ring)
 959                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
 960                 /* On overrun, truncate and discard */
 961                 if (unlikely(headcount > UIO_MAXIOV)) {
 962                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 963                         err = sock->ops->recvmsg(sock, &msg,
 964                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 965                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 966                         continue;
 967                 }
 968                 /* We don't need to be notified again. */
 969                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
 970                 fixup = msg.msg_iter;
 971                 if (unlikely((vhost_hlen))) {
 972                         /* We will supply the header ourselves
 973                          * TODO: support TSO.
 974                          */
 975                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
 976                 }
 977                 err = sock->ops->recvmsg(sock, &msg,
 978                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 979                 /* Userspace might have consumed the packet meanwhile:
 980                  * it's not supposed to do this usually, but might be hard
 981                  * to prevent. Discard data we got (if any) and keep going. */
 982                 if (unlikely(err != sock_len)) {
 983                         pr_debug("Discarded rx packet: "
 984                                  " len %d, expected %zd\n", err, sock_len);
 985                         vhost_discard_vq_desc(vq, headcount);
 986                         continue;
 987                 }
 988                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
 989                 if (unlikely(vhost_hlen)) {
 990                         if (copy_to_iter(&hdr, sizeof(hdr),
 991                                          &fixup) != sizeof(hdr)) {
 992                                 vq_err(vq, "Unable to write vnet_hdr "
 993                                        "at addr %p\n", vq->iov->iov_base);
 994                                 goto out;
 995                         }
 996                 } else {
 997                         /* Header came from socket; we'll need to patch
 998                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
 999                          */
1000                         iov_iter_advance(&fixup, sizeof(hdr));
1001                 }
1002                 /* TODO: Should check and handle checksum. */
1003
1004                 num_buffers = cpu_to_vhost16(vq, headcount);
1005                 if (likely(mergeable) &&
1006                     copy_to_iter(&num_buffers, sizeof num_buffers,
1007                                  &fixup) != sizeof num_buffers) {
1008                         vq_err(vq, "Failed num_buffers write");
1009                         vhost_discard_vq_desc(vq, headcount);
1010                         goto out;
1011                 }
1012                 nvq->done_idx += headcount;
1013                 if (nvq->done_idx > VHOST_NET_BATCH)
1014                         vhost_net_signal_used(nvq);
1015                 if (unlikely(vq_log))
1016                         vhost_log_write(vq, vq_log, log, vhost_len,
1017                                         vq->iov, in);
1018                 total_len += vhost_len;
1019         } while (likely(!vhost_exceeds_weight(vq, ++recv_pkts, total_len)));
1020
1021         if (unlikely(busyloop_intr))
1022                 vhost_poll_queue(&vq->poll);
1023         else if (!sock_len)
1024                 vhost_net_enable_vq(net, vq);
1025 out:
1026         vhost_net_signal_used(nvq);
1027         mutex_unlock(&vq->mutex);
1028 }
1029
1030 static void handle_tx_kick(struct vhost_work *work)
1031 {
1032         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1033                                                   poll.work);
1034         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1035
1036         handle_tx(net);
1037 }
1038
1039 static void handle_rx_kick(struct vhost_work *work)
1040 {
1041         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
1042                                                   poll.work);
1043         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
1044
1045         handle_rx(net);
1046 }
1047
1048 static void handle_tx_net(struct vhost_work *work)
1049 {
1050         struct vhost_net *net = container_of(work, struct vhost_net,
1051                                              poll[VHOST_NET_VQ_TX].work);
1052         handle_tx(net);
1053 }
1054
1055 static void handle_rx_net(struct vhost_work *work)
1056 {
1057         struct vhost_net *net = container_of(work, struct vhost_net,
1058                                              poll[VHOST_NET_VQ_RX].work);
1059         handle_rx(net);
1060 }
1061
1062 static int vhost_net_open(struct inode *inode, struct file *f)
1063 {
1064         struct vhost_net *n;
1065         struct vhost_dev *dev;
1066         struct vhost_virtqueue **vqs;
1067         void **queue;
1068         int i;
1069
1070         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
1071         if (!n)
1072                 return -ENOMEM;
1073         vqs = kmalloc_array(VHOST_NET_VQ_MAX, sizeof(*vqs), GFP_KERNEL);
1074         if (!vqs) {
1075                 kvfree(n);
1076                 return -ENOMEM;
1077         }
1078
1079         queue = kmalloc_array(VHOST_NET_BATCH, sizeof(void *),
1080                               GFP_KERNEL);
1081         if (!queue) {
1082                 kfree(vqs);
1083                 kvfree(n);
1084                 return -ENOMEM;
1085         }
1086         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
1087
1088         dev = &n->dev;
1089         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
1090         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
1091         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
1092         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
1093         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
1094                 n->vqs[i].ubufs = NULL;
1095                 n->vqs[i].ubuf_info = NULL;
1096                 n->vqs[i].upend_idx = 0;
1097                 n->vqs[i].done_idx = 0;
1098                 n->vqs[i].vhost_hlen = 0;
1099                 n->vqs[i].sock_hlen = 0;
1100                 n->vqs[i].rx_ring = NULL;
1101                 vhost_net_buf_init(&n->vqs[i].rxq);
1102         }
1103         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX,
1104                        UIO_MAXIOV + VHOST_NET_BATCH,
1105                        VHOST_NET_PKT_WEIGHT, VHOST_NET_WEIGHT);
1106
1107         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
1108         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
1109
1110         f->private_data = n;
1111
1112         return 0;
1113 }
1114
1115 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
1116                                         struct vhost_virtqueue *vq)
1117 {
1118         struct socket *sock;
1119         struct vhost_net_virtqueue *nvq =
1120                 container_of(vq, struct vhost_net_virtqueue, vq);
1121
1122         mutex_lock(&vq->mutex);
1123         sock = vq->private_data;
1124         vhost_net_disable_vq(n, vq);
1125         vq->private_data = NULL;
1126         vhost_net_buf_unproduce(nvq);
1127         nvq->rx_ring = NULL;
1128         mutex_unlock(&vq->mutex);
1129         return sock;
1130 }
1131
1132 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
1133                            struct socket **rx_sock)
1134 {
1135         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
1136         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
1137 }
1138
1139 static void vhost_net_flush_vq(struct vhost_net *n, int index)
1140 {
1141         vhost_poll_flush(n->poll + index);
1142         vhost_poll_flush(&n->vqs[index].vq.poll);
1143 }
1144
1145 static void vhost_net_flush(struct vhost_net *n)
1146 {
1147         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1148         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1149         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1150                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1151                 n->tx_flush = true;
1152                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1153                 /* Wait for all lower device DMAs done. */
1154                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1155                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1156                 n->tx_flush = false;
1157                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1158                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1159         }
1160 }
1161
1162 static int vhost_net_release(struct inode *inode, struct file *f)
1163 {
1164         struct vhost_net *n = f->private_data;
1165         struct socket *tx_sock;
1166         struct socket *rx_sock;
1167
1168         vhost_net_stop(n, &tx_sock, &rx_sock);
1169         vhost_net_flush(n);
1170         vhost_dev_stop(&n->dev);
1171         vhost_dev_cleanup(&n->dev);
1172         vhost_net_vq_reset(n);
1173         if (tx_sock)
1174                 sockfd_put(tx_sock);
1175         if (rx_sock)
1176                 sockfd_put(rx_sock);
1177         /* Make sure no callbacks are outstanding */
1178         synchronize_rcu_bh();
1179         /* We do an extra flush before freeing memory,
1180          * since jobs can re-queue themselves. */
1181         vhost_net_flush(n);
1182         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1183         kfree(n->dev.vqs);
1184         kvfree(n);
1185         return 0;
1186 }
1187
1188 static struct socket *get_raw_socket(int fd)
1189 {
1190         int r;
1191         struct socket *sock = sockfd_lookup(fd, &r);
1192
1193         if (!sock)
1194                 return ERR_PTR(-ENOTSOCK);
1195
1196         /* Parameter checking */
1197         if (sock->sk->sk_type != SOCK_RAW) {
1198                 r = -ESOCKTNOSUPPORT;
1199                 goto err;
1200         }
1201
1202         if (sock->sk->sk_family != AF_PACKET) {
1203                 r = -EPFNOSUPPORT;
1204                 goto err;
1205         }
1206         return sock;
1207 err:
1208         sockfd_put(sock);
1209         return ERR_PTR(r);
1210 }
1211
1212 static struct ptr_ring *get_tap_ptr_ring(struct file *file)
1213 {
1214         struct ptr_ring *ring;
1215         ring = tun_get_tx_ring(file);
1216         if (!IS_ERR(ring))
1217                 goto out;
1218         ring = tap_get_ptr_ring(file);
1219         if (!IS_ERR(ring))
1220                 goto out;
1221         ring = NULL;
1222 out:
1223         return ring;
1224 }
1225
1226 static struct socket *get_tap_socket(int fd)
1227 {
1228         struct file *file = fget(fd);
1229         struct socket *sock;
1230
1231         if (!file)
1232                 return ERR_PTR(-EBADF);
1233         sock = tun_get_socket(file);
1234         if (!IS_ERR(sock))
1235                 return sock;
1236         sock = tap_get_socket(file);
1237         if (IS_ERR(sock))
1238                 fput(file);
1239         return sock;
1240 }
1241
1242 static struct socket *get_socket(int fd)
1243 {
1244         struct socket *sock;
1245
1246         /* special case to disable backend */
1247         if (fd == -1)
1248                 return NULL;
1249         sock = get_raw_socket(fd);
1250         if (!IS_ERR(sock))
1251                 return sock;
1252         sock = get_tap_socket(fd);
1253         if (!IS_ERR(sock))
1254                 return sock;
1255         return ERR_PTR(-ENOTSOCK);
1256 }
1257
1258 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1259 {
1260         struct socket *sock, *oldsock;
1261         struct vhost_virtqueue *vq;
1262         struct vhost_net_virtqueue *nvq;
1263         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1264         int r;
1265
1266         mutex_lock(&n->dev.mutex);
1267         r = vhost_dev_check_owner(&n->dev);
1268         if (r)
1269                 goto err;
1270
1271         if (index >= VHOST_NET_VQ_MAX) {
1272                 r = -ENOBUFS;
1273                 goto err;
1274         }
1275         vq = &n->vqs[index].vq;
1276         nvq = &n->vqs[index];
1277         mutex_lock(&vq->mutex);
1278
1279         /* Verify that ring has been setup correctly. */
1280         if (!vhost_vq_access_ok(vq)) {
1281                 r = -EFAULT;
1282                 goto err_vq;
1283         }
1284         sock = get_socket(fd);
1285         if (IS_ERR(sock)) {
1286                 r = PTR_ERR(sock);
1287                 goto err_vq;
1288         }
1289
1290         /* start polling new socket */
1291         oldsock = vq->private_data;
1292         if (sock != oldsock) {
1293                 ubufs = vhost_net_ubuf_alloc(vq,
1294                                              sock && vhost_sock_zcopy(sock));
1295                 if (IS_ERR(ubufs)) {
1296                         r = PTR_ERR(ubufs);
1297                         goto err_ubufs;
1298                 }
1299
1300                 vhost_net_disable_vq(n, vq);
1301                 vq->private_data = sock;
1302                 vhost_net_buf_unproduce(nvq);
1303                 r = vhost_vq_init_access(vq);
1304                 if (r)
1305                         goto err_used;
1306                 r = vhost_net_enable_vq(n, vq);
1307                 if (r)
1308                         goto err_used;
1309                 if (index == VHOST_NET_VQ_RX) {
1310                         if (sock)
1311                                 nvq->rx_ring = get_tap_ptr_ring(sock->file);
1312                         else
1313                                 nvq->rx_ring = NULL;
1314                 }
1315
1316                 oldubufs = nvq->ubufs;
1317                 nvq->ubufs = ubufs;
1318
1319                 n->tx_packets = 0;
1320                 n->tx_zcopy_err = 0;
1321                 n->tx_flush = false;
1322         }
1323
1324         mutex_unlock(&vq->mutex);
1325
1326         if (oldubufs) {
1327                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1328                 mutex_lock(&vq->mutex);
1329                 vhost_zerocopy_signal_used(n, vq);
1330                 mutex_unlock(&vq->mutex);
1331         }
1332
1333         if (oldsock) {
1334                 vhost_net_flush_vq(n, index);
1335                 sockfd_put(oldsock);
1336         }
1337
1338         mutex_unlock(&n->dev.mutex);
1339         return 0;
1340
1341 err_used:
1342         vq->private_data = oldsock;
1343         vhost_net_enable_vq(n, vq);
1344         if (ubufs)
1345                 vhost_net_ubuf_put_wait_and_free(ubufs);
1346 err_ubufs:
1347         if (sock)
1348                 sockfd_put(sock);
1349 err_vq:
1350         mutex_unlock(&vq->mutex);
1351 err:
1352         mutex_unlock(&n->dev.mutex);
1353         return r;
1354 }
1355
1356 static long vhost_net_reset_owner(struct vhost_net *n)
1357 {
1358         struct socket *tx_sock = NULL;
1359         struct socket *rx_sock = NULL;
1360         long err;
1361         struct vhost_umem *umem;
1362
1363         mutex_lock(&n->dev.mutex);
1364         err = vhost_dev_check_owner(&n->dev);
1365         if (err)
1366                 goto done;
1367         umem = vhost_dev_reset_owner_prepare();
1368         if (!umem) {
1369                 err = -ENOMEM;
1370                 goto done;
1371         }
1372         vhost_net_stop(n, &tx_sock, &rx_sock);
1373         vhost_net_flush(n);
1374         vhost_dev_stop(&n->dev);
1375         vhost_dev_reset_owner(&n->dev, umem);
1376         vhost_net_vq_reset(n);
1377 done:
1378         mutex_unlock(&n->dev.mutex);
1379         if (tx_sock)
1380                 sockfd_put(tx_sock);
1381         if (rx_sock)
1382                 sockfd_put(rx_sock);
1383         return err;
1384 }
1385
1386 static int vhost_net_set_backend_features(struct vhost_net *n, u64 features)
1387 {
1388         int i;
1389
1390         mutex_lock(&n->dev.mutex);
1391         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1392                 mutex_lock(&n->vqs[i].vq.mutex);
1393                 n->vqs[i].vq.acked_backend_features = features;
1394                 mutex_unlock(&n->vqs[i].vq.mutex);
1395         }
1396         mutex_unlock(&n->dev.mutex);
1397
1398         return 0;
1399 }
1400
1401 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1402 {
1403         size_t vhost_hlen, sock_hlen, hdr_len;
1404         int i;
1405
1406         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1407                                (1ULL << VIRTIO_F_VERSION_1))) ?
1408                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1409                         sizeof(struct virtio_net_hdr);
1410         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1411                 /* vhost provides vnet_hdr */
1412                 vhost_hlen = hdr_len;
1413                 sock_hlen = 0;
1414         } else {
1415                 /* socket provides vnet_hdr */
1416                 vhost_hlen = 0;
1417                 sock_hlen = hdr_len;
1418         }
1419         mutex_lock(&n->dev.mutex);
1420         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1421             !vhost_log_access_ok(&n->dev))
1422                 goto out_unlock;
1423
1424         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1425                 if (vhost_init_device_iotlb(&n->dev, true))
1426                         goto out_unlock;
1427         }
1428
1429         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1430                 mutex_lock(&n->vqs[i].vq.mutex);
1431                 n->vqs[i].vq.acked_features = features;
1432                 n->vqs[i].vhost_hlen = vhost_hlen;
1433                 n->vqs[i].sock_hlen = sock_hlen;
1434                 mutex_unlock(&n->vqs[i].vq.mutex);
1435         }
1436         mutex_unlock(&n->dev.mutex);
1437         return 0;
1438
1439 out_unlock:
1440         mutex_unlock(&n->dev.mutex);
1441         return -EFAULT;
1442 }
1443
1444 static long vhost_net_set_owner(struct vhost_net *n)
1445 {
1446         int r;
1447
1448         mutex_lock(&n->dev.mutex);
1449         if (vhost_dev_has_owner(&n->dev)) {
1450                 r = -EBUSY;
1451                 goto out;
1452         }
1453         r = vhost_net_set_ubuf_info(n);
1454         if (r)
1455                 goto out;
1456         r = vhost_dev_set_owner(&n->dev);
1457         if (r)
1458                 vhost_net_clear_ubuf_info(n);
1459         vhost_net_flush(n);
1460 out:
1461         mutex_unlock(&n->dev.mutex);
1462         return r;
1463 }
1464
1465 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1466                             unsigned long arg)
1467 {
1468         struct vhost_net *n = f->private_data;
1469         void __user *argp = (void __user *)arg;
1470         u64 __user *featurep = argp;
1471         struct vhost_vring_file backend;
1472         u64 features;
1473         int r;
1474
1475         switch (ioctl) {
1476         case VHOST_NET_SET_BACKEND:
1477                 if (copy_from_user(&backend, argp, sizeof backend))
1478                         return -EFAULT;
1479                 return vhost_net_set_backend(n, backend.index, backend.fd);
1480         case VHOST_GET_FEATURES:
1481                 features = VHOST_NET_FEATURES;
1482                 if (copy_to_user(featurep, &features, sizeof features))
1483                         return -EFAULT;
1484                 return 0;
1485         case VHOST_SET_FEATURES:
1486                 if (copy_from_user(&features, featurep, sizeof features))
1487                         return -EFAULT;
1488                 if (features & ~VHOST_NET_FEATURES)
1489                         return -EOPNOTSUPP;
1490                 return vhost_net_set_features(n, features);
1491         case VHOST_GET_BACKEND_FEATURES:
1492                 features = VHOST_NET_BACKEND_FEATURES;
1493                 if (copy_to_user(featurep, &features, sizeof(features)))
1494                         return -EFAULT;
1495                 return 0;
1496         case VHOST_SET_BACKEND_FEATURES:
1497                 if (copy_from_user(&features, featurep, sizeof(features)))
1498                         return -EFAULT;
1499                 if (features & ~VHOST_NET_BACKEND_FEATURES)
1500                         return -EOPNOTSUPP;
1501                 return vhost_net_set_backend_features(n, features);
1502         case VHOST_RESET_OWNER:
1503                 return vhost_net_reset_owner(n);
1504         case VHOST_SET_OWNER:
1505                 return vhost_net_set_owner(n);
1506         default:
1507                 mutex_lock(&n->dev.mutex);
1508                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1509                 if (r == -ENOIOCTLCMD)
1510                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1511                 else
1512                         vhost_net_flush(n);
1513                 mutex_unlock(&n->dev.mutex);
1514                 return r;
1515         }
1516 }
1517
1518 #ifdef CONFIG_COMPAT
1519 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1520                                    unsigned long arg)
1521 {
1522         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1523 }
1524 #endif
1525
1526 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1527 {
1528         struct file *file = iocb->ki_filp;
1529         struct vhost_net *n = file->private_data;
1530         struct vhost_dev *dev = &n->dev;
1531         int noblock = file->f_flags & O_NONBLOCK;
1532
1533         return vhost_chr_read_iter(dev, to, noblock);
1534 }
1535
1536 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1537                                         struct iov_iter *from)
1538 {
1539         struct file *file = iocb->ki_filp;
1540         struct vhost_net *n = file->private_data;
1541         struct vhost_dev *dev = &n->dev;
1542
1543         return vhost_chr_write_iter(dev, from);
1544 }
1545
1546 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1547 {
1548         struct vhost_net *n = file->private_data;
1549         struct vhost_dev *dev = &n->dev;
1550
1551         return vhost_chr_poll(file, dev, wait);
1552 }
1553
1554 static const struct file_operations vhost_net_fops = {
1555         .owner          = THIS_MODULE,
1556         .release        = vhost_net_release,
1557         .read_iter      = vhost_net_chr_read_iter,
1558         .write_iter     = vhost_net_chr_write_iter,
1559         .poll           = vhost_net_chr_poll,
1560         .unlocked_ioctl = vhost_net_ioctl,
1561 #ifdef CONFIG_COMPAT
1562         .compat_ioctl   = vhost_net_compat_ioctl,
1563 #endif
1564         .open           = vhost_net_open,
1565         .llseek         = noop_llseek,
1566 };
1567
1568 static struct miscdevice vhost_net_misc = {
1569         .minor = VHOST_NET_MINOR,
1570         .name = "vhost-net",
1571         .fops = &vhost_net_fops,
1572 };
1573
1574 static int vhost_net_init(void)
1575 {
1576         if (experimental_zcopytx)
1577                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1578         return misc_register(&vhost_net_misc);
1579 }
1580 module_init(vhost_net_init);
1581
1582 static void vhost_net_exit(void)
1583 {
1584         misc_deregister(&vhost_net_misc);
1585 }
1586 module_exit(vhost_net_exit);
1587
1588 MODULE_VERSION("0.0.1");
1589 MODULE_LICENSE("GPL v2");
1590 MODULE_AUTHOR("Michael S. Tsirkin");
1591 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1592 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1593 MODULE_ALIAS("devname:vhost-net");