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