drivers/vhost/net.c

   1 /* Copyright (C) 2009 Red Hat, Inc.
   2  * Author: Michael S. Tsirkin <[email protected]>
   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
  36 #include "vhost.h"
  37
  38 static int experimental_zcopytx = 1;
  39 module_param(experimental_zcopytx, int, 0444);
  40 MODULE_PARM_DESC(experimental_zcopytx, "Enable Zero Copy TX;"
  41                                        " 1 -Enable; 0 - Disable");
  42
  43 /* Max number of bytes transferred before requeueing the job.
  44  * Using this limit prevents one virtqueue from starving others. */
  45 #define VHOST_NET_WEIGHT 0x80000
  46
  47 /* Max number of packets transferred before requeueing the job.
  48  * Using this limit prevents one virtqueue from starving rx. */
  49 #define VHOST_NET_PKT_WEIGHT(vq) ((vq)->num * 2)
  50
  51 /* MAX number of TX used buffers for outstanding zerocopy */
  52 #define VHOST_MAX_PEND 128
  53 #define VHOST_GOODCOPY_LEN 256
  54
  55 /*
  56  * For transmit, used buffer len is unused; we override it to track buffer
  57  * status internally; used for zerocopy tx only.
  58  */
  59 /* Lower device DMA failed */
  60 #define VHOST_DMA_FAILED_LEN    ((__force __virtio32)3)
  61 /* Lower device DMA done */
  62 #define VHOST_DMA_DONE_LEN      ((__force __virtio32)2)
  63 /* Lower device DMA in progress */
  64 #define VHOST_DMA_IN_PROGRESS   ((__force __virtio32)1)
  65 /* Buffer unused */
  66 #define VHOST_DMA_CLEAR_LEN     ((__force __virtio32)0)
  67
  68 #define VHOST_DMA_IS_DONE(len) ((__force u32)(len) >= (__force u32)VHOST_DMA_DONE_LEN)
  69
  70 enum {
  71         VHOST_NET_FEATURES = VHOST_FEATURES |
  72                          (1ULL << VHOST_NET_F_VIRTIO_NET_HDR) |
  73                          (1ULL << VIRTIO_NET_F_MRG_RXBUF) |
  74                          (1ULL << VIRTIO_F_IOMMU_PLATFORM)
  75 };
  76
  77 enum {
  78         VHOST_NET_VQ_RX = 0,
  79         VHOST_NET_VQ_TX = 1,
  80         VHOST_NET_VQ_MAX = 2,
  81 };
  82
  83 struct vhost_net_ubuf_ref {
  84         /* refcount follows semantics similar to kref:
  85          *  0: object is released
  86          *  1: no outstanding ubufs
  87          * >1: outstanding ubufs
  88          */
  89         atomic_t refcount;
  90         wait_queue_head_t wait;
  91         struct vhost_virtqueue *vq;
  92 };
  93
  94 #define VHOST_RX_BATCH 64
  95 struct vhost_net_buf {
  96         void **queue;
  97         int tail;
  98         int head;
  99 };
 100
 101 struct vhost_net_virtqueue {
 102         struct vhost_virtqueue vq;
 103         size_t vhost_hlen;
 104         size_t sock_hlen;
 105         /* vhost zerocopy support fields below: */
 106         /* last used idx for outstanding DMA zerocopy buffers */
 107         int upend_idx;
 108         /* first used idx for DMA done zerocopy buffers */
 109         int done_idx;
 110         /* an array of userspace buffers info */
 111         struct ubuf_info *ubuf_info;
 112         /* Reference counting for outstanding ubufs.
 113          * Protected by vq mutex. Writers must also take device mutex. */
 114         struct vhost_net_ubuf_ref *ubufs;
 115         struct ptr_ring *rx_ring;
 116         struct vhost_net_buf rxq;
 117 };
 118
 119 struct vhost_net {
 120         struct vhost_dev dev;
 121         struct vhost_net_virtqueue vqs[VHOST_NET_VQ_MAX];
 122         struct vhost_poll poll[VHOST_NET_VQ_MAX];
 123         /* Number of TX recently submitted.
 124          * Protected by tx vq lock. */
 125         unsigned tx_packets;
 126         /* Number of times zerocopy TX recently failed.
 127          * Protected by tx vq lock. */
 128         unsigned tx_zcopy_err;
 129         /* Flush in progress. Protected by tx vq lock. */
 130         bool tx_flush;
 131 };
 132
 133 static unsigned vhost_net_zcopy_mask __read_mostly;
 134
 135 static void *vhost_net_buf_get_ptr(struct vhost_net_buf *rxq)
 136 {
 137         if (rxq->tail != rxq->head)
 138                 return rxq->queue[rxq->head];
 139         else
 140                 return NULL;
 141 }
 142
 143 static int vhost_net_buf_get_size(struct vhost_net_buf *rxq)
 144 {
 145         return rxq->tail - rxq->head;
 146 }
 147
 148 static int vhost_net_buf_is_empty(struct vhost_net_buf *rxq)
 149 {
 150         return rxq->tail == rxq->head;
 151 }
 152
 153 static void *vhost_net_buf_consume(struct vhost_net_buf *rxq)
 154 {
 155         void *ret = vhost_net_buf_get_ptr(rxq);
 156         ++rxq->head;
 157         return ret;
 158 }
 159
 160 static int vhost_net_buf_produce(struct vhost_net_virtqueue *nvq)
 161 {
 162         struct vhost_net_buf *rxq = &nvq->rxq;
 163
 164         rxq->head = 0;
 165         rxq->tail = ptr_ring_consume_batched(nvq->rx_ring, rxq->queue,
 166                                               VHOST_RX_BATCH);
 167         return rxq->tail;
 168 }
 169
 170 static void vhost_net_buf_unproduce(struct vhost_net_virtqueue *nvq)
 171 {
 172         struct vhost_net_buf *rxq = &nvq->rxq;
 173
 174         if (nvq->rx_ring && !vhost_net_buf_is_empty(rxq)) {
 175                 ptr_ring_unconsume(nvq->rx_ring, rxq->queue + rxq->head,
 176                                    vhost_net_buf_get_size(rxq),
 177                                    tun_ptr_free);
 178                 rxq->head = rxq->tail = 0;
 179         }
 180 }
 181
 182 static int vhost_net_buf_peek_len(void *ptr)
 183 {
 184         if (tun_is_xdp_buff(ptr)) {
 185                 struct xdp_buff *xdp = tun_ptr_to_xdp(ptr);
 186
 187                 return xdp->data_end - xdp->data;
 188         }
 189
 190         return __skb_array_len_with_tag(ptr);
 191 }
 192
 193 static int vhost_net_buf_peek(struct vhost_net_virtqueue *nvq)
 194 {
 195         struct vhost_net_buf *rxq = &nvq->rxq;
 196
 197         if (!vhost_net_buf_is_empty(rxq))
 198                 goto out;
 199
 200         if (!vhost_net_buf_produce(nvq))
 201                 return 0;
 202
 203 out:
 204         return vhost_net_buf_peek_len(vhost_net_buf_get_ptr(rxq));
 205 }
 206
 207 static void vhost_net_buf_init(struct vhost_net_buf *rxq)
 208 {
 209         rxq->head = rxq->tail = 0;
 210 }
 211
 212 static void vhost_net_enable_zcopy(int vq)
 213 {
 214         vhost_net_zcopy_mask |= 0x1 << vq;
 215 }
 216
 217 static struct vhost_net_ubuf_ref *
 218 vhost_net_ubuf_alloc(struct vhost_virtqueue *vq, bool zcopy)
 219 {
 220         struct vhost_net_ubuf_ref *ubufs;
 221         /* No zero copy backend? Nothing to count. */
 222         if (!zcopy)
 223                 return NULL;
 224         ubufs = kmalloc(sizeof(*ubufs), GFP_KERNEL);
 225         if (!ubufs)
 226                 return ERR_PTR(-ENOMEM);
 227         atomic_set(&ubufs->refcount, 1);
 228         init_waitqueue_head(&ubufs->wait);
 229         ubufs->vq = vq;
 230         return ubufs;
 231 }
 232
 233 static int vhost_net_ubuf_put(struct vhost_net_ubuf_ref *ubufs)
 234 {
 235         int r = atomic_sub_return(1, &ubufs->refcount);
 236         if (unlikely(!r))
 237                 wake_up(&ubufs->wait);
 238         return r;
 239 }
 240
 241 static void vhost_net_ubuf_put_and_wait(struct vhost_net_ubuf_ref *ubufs)
 242 {
 243         vhost_net_ubuf_put(ubufs);
 244         wait_event(ubufs->wait, !atomic_read(&ubufs->refcount));
 245 }
 246
 247 static void vhost_net_ubuf_put_wait_and_free(struct vhost_net_ubuf_ref *ubufs)
 248 {
 249         vhost_net_ubuf_put_and_wait(ubufs);
 250         kfree(ubufs);
 251 }
 252
 253 static void vhost_net_clear_ubuf_info(struct vhost_net *n)
 254 {
 255         int i;
 256
 257         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 258                 kfree(n->vqs[i].ubuf_info);
 259                 n->vqs[i].ubuf_info = NULL;
 260         }
 261 }
 262
 263 static int vhost_net_set_ubuf_info(struct vhost_net *n)
 264 {
 265         bool zcopy;
 266         int i;
 267
 268         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
 269                 zcopy = vhost_net_zcopy_mask & (0x1 << i);
 270                 if (!zcopy)
 271                         continue;
 272                 n->vqs[i].ubuf_info = kmalloc(sizeof(*n->vqs[i].ubuf_info) *
 273                                               UIO_MAXIOV, GFP_KERNEL);
 274                 if  (!n->vqs[i].ubuf_info)
 275                         goto err;
 276         }
 277         return 0;
 278
 279 err:
 280         vhost_net_clear_ubuf_info(n);
 281         return -ENOMEM;
 282 }
 283
 284 static void vhost_net_vq_reset(struct vhost_net *n)
 285 {
 286         int i;
 287
 288         vhost_net_clear_ubuf_info(n);
 289
 290         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 291                 n->vqs[i].done_idx = 0;
 292                 n->vqs[i].upend_idx = 0;
 293                 n->vqs[i].ubufs = NULL;
 294                 n->vqs[i].vhost_hlen = 0;
 295                 n->vqs[i].sock_hlen = 0;
 296                 vhost_net_buf_init(&n->vqs[i].rxq);
 297         }
 298
 299 }
 300
 301 static void vhost_net_tx_packet(struct vhost_net *net)
 302 {
 303         ++net->tx_packets;
 304         if (net->tx_packets < 1024)
 305                 return;
 306         net->tx_packets = 0;
 307         net->tx_zcopy_err = 0;
 308 }
 309
 310 static void vhost_net_tx_err(struct vhost_net *net)
 311 {
 312         ++net->tx_zcopy_err;
 313 }
 314
 315 static bool vhost_net_tx_select_zcopy(struct vhost_net *net)
 316 {
 317         /* TX flush waits for outstanding DMAs to be done.
 318          * Don't start new DMAs.
 319          */
 320         return !net->tx_flush &&
 321                 net->tx_packets / 64 >= net->tx_zcopy_err;
 322 }
 323
 324 static bool vhost_sock_zcopy(struct socket *sock)
 325 {
 326         return unlikely(experimental_zcopytx) &&
 327                 sock_flag(sock->sk, SOCK_ZEROCOPY);
 328 }
 329
 330 /* In case of DMA done not in order in lower device driver for some reason.
 331  * upend_idx is used to track end of used idx, done_idx is used to track head
 332  * of used idx. Once lower device DMA done contiguously, we will signal KVM
 333  * guest used idx.
 334  */
 335 static void vhost_zerocopy_signal_used(struct vhost_net *net,
 336                                        struct vhost_virtqueue *vq)
 337 {
 338         struct vhost_net_virtqueue *nvq =
 339                 container_of(vq, struct vhost_net_virtqueue, vq);
 340         int i, add;
 341         int j = 0;
 342
 343         for (i = nvq->done_idx; i != nvq->upend_idx; i = (i + 1) % UIO_MAXIOV) {
 344                 if (vq->heads[i].len == VHOST_DMA_FAILED_LEN)
 345                         vhost_net_tx_err(net);
 346                 if (VHOST_DMA_IS_DONE(vq->heads[i].len)) {
 347                         vq->heads[i].len = VHOST_DMA_CLEAR_LEN;
 348                         ++j;
 349                 } else
 350                         break;
 351         }
 352         while (j) {
 353                 add = min(UIO_MAXIOV - nvq->done_idx, j);
 354                 vhost_add_used_and_signal_n(vq->dev, vq,
 355                                             &vq->heads[nvq->done_idx], add);
 356                 nvq->done_idx = (nvq->done_idx + add) % UIO_MAXIOV;
 357                 j -= add;
 358         }
 359 }
 360
 361 static void vhost_zerocopy_callback(struct ubuf_info *ubuf, bool success)
 362 {
 363         struct vhost_net_ubuf_ref *ubufs = ubuf->ctx;
 364         struct vhost_virtqueue *vq = ubufs->vq;
 365         int cnt;
 366
 367         rcu_read_lock_bh();
 368
 369         /* set len to mark this desc buffers done DMA */
 370         vq->heads[ubuf->desc].len = success ?
 371                 VHOST_DMA_DONE_LEN : VHOST_DMA_FAILED_LEN;
 372         cnt = vhost_net_ubuf_put(ubufs);
 373
 374         /*
 375          * Trigger polling thread if guest stopped submitting new buffers:
 376          * in this case, the refcount after decrement will eventually reach 1.
 377          * We also trigger polling periodically after each 16 packets
 378          * (the value 16 here is more or less arbitrary, it's tuned to trigger
 379          * less than 10% of times).
 380          */
 381         if (cnt <= 1 || !(cnt % 16))
 382                 vhost_poll_queue(&vq->poll);
 383
 384         rcu_read_unlock_bh();
 385 }
 386
 387 static inline unsigned long busy_clock(void)
 388 {
 389         return local_clock() >> 10;
 390 }
 391
 392 static bool vhost_can_busy_poll(struct vhost_dev *dev,
 393                                 unsigned long endtime)
 394 {
 395         return likely(!need_resched()) &&
 396                likely(!time_after(busy_clock(), endtime)) &&
 397                likely(!signal_pending(current)) &&
 398                !vhost_has_work(dev);
 399 }
 400
 401 static void vhost_net_disable_vq(struct vhost_net *n,
 402                                  struct vhost_virtqueue *vq)
 403 {
 404         struct vhost_net_virtqueue *nvq =
 405                 container_of(vq, struct vhost_net_virtqueue, vq);
 406         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 407         if (!vq->private_data)
 408                 return;
 409         vhost_poll_stop(poll);
 410 }
 411
 412 static int vhost_net_enable_vq(struct vhost_net *n,
 413                                 struct vhost_virtqueue *vq)
 414 {
 415         struct vhost_net_virtqueue *nvq =
 416                 container_of(vq, struct vhost_net_virtqueue, vq);
 417         struct vhost_poll *poll = n->poll + (nvq - n->vqs);
 418         struct socket *sock;
 419
 420         sock = vq->private_data;
 421         if (!sock)
 422                 return 0;
 423
 424         return vhost_poll_start(poll, sock->file);
 425 }
 426
 427 static int vhost_net_tx_get_vq_desc(struct vhost_net *net,
 428                                     struct vhost_virtqueue *vq,
 429                                     struct iovec iov[], unsigned int iov_size,
 430                                     unsigned int *out_num, unsigned int *in_num)
 431 {
 432         unsigned long uninitialized_var(endtime);
 433         int r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 434                                   out_num, in_num, NULL, NULL);
 435
 436         if (r == vq->num && vq->busyloop_timeout) {
 437                 preempt_disable();
 438                 endtime = busy_clock() + vq->busyloop_timeout;
 439                 while (vhost_can_busy_poll(vq->dev, endtime) &&
 440                        vhost_vq_avail_empty(vq->dev, vq))
 441                         cpu_relax();
 442                 preempt_enable();
 443                 r = vhost_get_vq_desc(vq, vq->iov, ARRAY_SIZE(vq->iov),
 444                                       out_num, in_num, NULL, NULL);
 445         }
 446
 447         return r;
 448 }
 449
 450 static bool vhost_exceeds_maxpend(struct vhost_net *net)
 451 {
 452         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 453         struct vhost_virtqueue *vq = &nvq->vq;
 454
 455         return (nvq->upend_idx + UIO_MAXIOV - nvq->done_idx) % UIO_MAXIOV >
 456                min_t(unsigned int, VHOST_MAX_PEND, vq->num >> 2);
 457 }
 458
 459 /* Expects to be always run from workqueue - which acts as
 460  * read-size critical section for our kind of RCU. */
 461 static void handle_tx(struct vhost_net *net)
 462 {
 463         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 464         struct vhost_virtqueue *vq = &nvq->vq;
 465         unsigned out, in;
 466         int head;
 467         struct msghdr msg = {
 468                 .msg_name = NULL,
 469                 .msg_namelen = 0,
 470                 .msg_control = NULL,
 471                 .msg_controllen = 0,
 472                 .msg_flags = MSG_DONTWAIT,
 473         };
 474         size_t len, total_len = 0;
 475         int err;
 476         size_t hdr_size;
 477         struct socket *sock;
 478         struct vhost_net_ubuf_ref *uninitialized_var(ubufs);
 479         bool zcopy, zcopy_used;
 480         int sent_pkts = 0;
 481
 482         mutex_lock(&vq->mutex);
 483         sock = vq->private_data;
 484         if (!sock)
 485                 goto out;
 486
 487         if (!vq_iotlb_prefetch(vq))
 488                 goto out;
 489
 490         vhost_disable_notify(&net->dev, vq);
 491         vhost_net_disable_vq(net, vq);
 492
 493         hdr_size = nvq->vhost_hlen;
 494         zcopy = nvq->ubufs;
 495
 496         for (;;) {
 497                 /* Release DMAs done buffers first */
 498                 if (zcopy)
 499                         vhost_zerocopy_signal_used(net, vq);
 500
 501
 502                 head = vhost_net_tx_get_vq_desc(net, vq, vq->iov,
 503                                                 ARRAY_SIZE(vq->iov),
 504                                                 &out, &in);
 505                 /* On error, stop handling until the next kick. */
 506                 if (unlikely(head < 0))
 507                         break;
 508                 /* Nothing new?  Wait for eventfd to tell us they refilled. */
 509                 if (head == vq->num) {
 510                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 511                                 vhost_disable_notify(&net->dev, vq);
 512                                 continue;
 513                         }
 514                         break;
 515                 }
 516                 if (in) {
 517                         vq_err(vq, "Unexpected descriptor format for TX: "
 518                                "out %d, int %d\n", out, in);
 519                         break;
 520                 }
 521                 /* Skip header. TODO: support TSO. */
 522                 len = iov_length(vq->iov, out);
 523                 iov_iter_init(&msg.msg_iter, WRITE, vq->iov, out, len);
 524                 iov_iter_advance(&msg.msg_iter, hdr_size);
 525                 /* Sanity check */
 526                 if (!msg_data_left(&msg)) {
 527                         vq_err(vq, "Unexpected header len for TX: "
 528                                "%zd expected %zd\n",
 529                                len, hdr_size);
 530                         break;
 531                 }
 532                 len = msg_data_left(&msg);
 533
 534                 zcopy_used = zcopy && len >= VHOST_GOODCOPY_LEN
 535                                    && !vhost_exceeds_maxpend(net)
 536                                    && vhost_net_tx_select_zcopy(net);
 537
 538                 /* use msg_control to pass vhost zerocopy ubuf info to skb */
 539                 if (zcopy_used) {
 540                         struct ubuf_info *ubuf;
 541                         ubuf = nvq->ubuf_info + nvq->upend_idx;
 542
 543                         vq->heads[nvq->upend_idx].id = cpu_to_vhost32(vq, head);
 544                         vq->heads[nvq->upend_idx].len = VHOST_DMA_IN_PROGRESS;
 545                         ubuf->callback = vhost_zerocopy_callback;
 546                         ubuf->ctx = nvq->ubufs;
 547                         ubuf->desc = nvq->upend_idx;
 548                         refcount_set(&ubuf->refcnt, 1);
 549                         msg.msg_control = ubuf;
 550                         msg.msg_controllen = sizeof(ubuf);
 551                         ubufs = nvq->ubufs;
 552                         atomic_inc(&ubufs->refcount);
 553                         nvq->upend_idx = (nvq->upend_idx + 1) % UIO_MAXIOV;
 554                 } else {
 555                         msg.msg_control = NULL;
 556                         ubufs = NULL;
 557                 }
 558
 559                 total_len += len;
 560                 if (total_len < VHOST_NET_WEIGHT &&
 561                     !vhost_vq_avail_empty(&net->dev, vq) &&
 562                     likely(!vhost_exceeds_maxpend(net))) {
 563                         msg.msg_flags |= MSG_MORE;
 564                 } else {
 565                         msg.msg_flags &= ~MSG_MORE;
 566                 }
 567
 568                 /* TODO: Check specific error and bomb out unless ENOBUFS? */
 569                 err = sock->ops->sendmsg(sock, &msg, len);
 570                 if (unlikely(err < 0)) {
 571                         if (zcopy_used) {
 572                                 vhost_net_ubuf_put(ubufs);
 573                                 nvq->upend_idx = ((unsigned)nvq->upend_idx - 1)
 574                                         % UIO_MAXIOV;
 575                         }
 576                         vhost_discard_vq_desc(vq, 1);
 577                         vhost_net_enable_vq(net, vq);
 578                         break;
 579                 }
 580                 if (err != len)
 581                         pr_debug("Truncated TX packet: "
 582                                  " len %d != %zd\n", err, len);
 583                 if (!zcopy_used)
 584                         vhost_add_used_and_signal(&net->dev, vq, head, 0);
 585                 else
 586                         vhost_zerocopy_signal_used(net, vq);
 587                 vhost_net_tx_packet(net);
 588                 if (unlikely(total_len >= VHOST_NET_WEIGHT) ||
 589                     unlikely(++sent_pkts >= VHOST_NET_PKT_WEIGHT(vq))) {
 590                         vhost_poll_queue(&vq->poll);
 591                         break;
 592                 }
 593         }
 594 out:
 595         mutex_unlock(&vq->mutex);
 596 }
 597
 598 static int peek_head_len(struct vhost_net_virtqueue *rvq, struct sock *sk)
 599 {
 600         struct sk_buff *head;
 601         int len = 0;
 602         unsigned long flags;
 603
 604         if (rvq->rx_ring)
 605                 return vhost_net_buf_peek(rvq);
 606
 607         spin_lock_irqsave(&sk->sk_receive_queue.lock, flags);
 608         head = skb_peek(&sk->sk_receive_queue);
 609         if (likely(head)) {
 610                 len = head->len;
 611                 if (skb_vlan_tag_present(head))
 612                         len += VLAN_HLEN;
 613         }
 614
 615         spin_unlock_irqrestore(&sk->sk_receive_queue.lock, flags);
 616         return len;
 617 }
 618
 619 static int sk_has_rx_data(struct sock *sk)
 620 {
 621         struct socket *sock = sk->sk_socket;
 622
 623         if (sock->ops->peek_len)
 624                 return sock->ops->peek_len(sock);
 625
 626         return skb_queue_empty(&sk->sk_receive_queue);
 627 }
 628
 629 static int vhost_net_rx_peek_head_len(struct vhost_net *net, struct sock *sk)
 630 {
 631         struct vhost_net_virtqueue *rvq = &net->vqs[VHOST_NET_VQ_RX];
 632         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_TX];
 633         struct vhost_virtqueue *vq = &nvq->vq;
 634         unsigned long uninitialized_var(endtime);
 635         int len = peek_head_len(rvq, sk);
 636
 637         if (!len && vq->busyloop_timeout) {
 638                 /* Both tx vq and rx socket were polled here */
 639                 mutex_lock_nested(&vq->mutex, 1);
 640                 vhost_disable_notify(&net->dev, vq);
 641
 642                 preempt_disable();
 643                 endtime = busy_clock() + vq->busyloop_timeout;
 644
 645                 while (vhost_can_busy_poll(&net->dev, endtime) &&
 646                        !sk_has_rx_data(sk) &&
 647                        vhost_vq_avail_empty(&net->dev, vq))
 648                         cpu_relax();
 649
 650                 preempt_enable();
 651
 652                 if (!vhost_vq_avail_empty(&net->dev, vq))
 653                         vhost_poll_queue(&vq->poll);
 654                 else if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 655                         vhost_disable_notify(&net->dev, vq);
 656                         vhost_poll_queue(&vq->poll);
 657                 }
 658
 659                 mutex_unlock(&vq->mutex);
 660
 661                 len = peek_head_len(rvq, sk);
 662         }
 663
 664         return len;
 665 }
 666
 667 /* This is a multi-buffer version of vhost_get_desc, that works if
 668  *      vq has read descriptors only.
 669  * @vq          - the relevant virtqueue
 670  * @datalen     - data length we'll be reading
 671  * @iovcount    - returned count of io vectors we fill
 672  * @log         - vhost log
 673  * @log_num     - log offset
 674  * @quota       - headcount quota, 1 for big buffer
 675  *      returns number of buffer heads allocated, negative on error
 676  */
 677 static int get_rx_bufs(struct vhost_virtqueue *vq,
 678                        struct vring_used_elem *heads,
 679                        int datalen,
 680                        unsigned *iovcount,
 681                        struct vhost_log *log,
 682                        unsigned *log_num,
 683                        unsigned int quota)
 684 {
 685         unsigned int out, in;
 686         int seg = 0;
 687         int headcount = 0;
 688         unsigned d;
 689         int r, nlogs = 0;
 690         /* len is always initialized before use since we are always called with
 691          * datalen > 0.
 692          */
 693         u32 uninitialized_var(len);
 694
 695         while (datalen > 0 && headcount < quota) {
 696                 if (unlikely(seg >= UIO_MAXIOV)) {
 697                         r = -ENOBUFS;
 698                         goto err;
 699                 }
 700                 r = vhost_get_vq_desc(vq, vq->iov + seg,
 701                                       ARRAY_SIZE(vq->iov) - seg, &out,
 702                                       &in, log, log_num);
 703                 if (unlikely(r < 0))
 704                         goto err;
 705
 706                 d = r;
 707                 if (d == vq->num) {
 708                         r = 0;
 709                         goto err;
 710                 }
 711                 if (unlikely(out || in <= 0)) {
 712                         vq_err(vq, "unexpected descriptor format for RX: "
 713                                 "out %d, in %d\n", out, in);
 714                         r = -EINVAL;
 715                         goto err;
 716                 }
 717                 if (unlikely(log)) {
 718                         nlogs += *log_num;
 719                         log += *log_num;
 720                 }
 721                 heads[headcount].id = cpu_to_vhost32(vq, d);
 722                 len = iov_length(vq->iov + seg, in);
 723                 heads[headcount].len = cpu_to_vhost32(vq, len);
 724                 datalen -= len;
 725                 ++headcount;
 726                 seg += in;
 727         }
 728         heads[headcount - 1].len = cpu_to_vhost32(vq, len + datalen);
 729         *iovcount = seg;
 730         if (unlikely(log))
 731                 *log_num = nlogs;
 732
 733         /* Detect overrun */
 734         if (unlikely(datalen > 0)) {
 735                 r = UIO_MAXIOV + 1;
 736                 goto err;
 737         }
 738         return headcount;
 739 err:
 740         vhost_discard_vq_desc(vq, headcount);
 741         return r;
 742 }
 743
 744 /* Expects to be always run from workqueue - which acts as
 745  * read-size critical section for our kind of RCU. */
 746 static void handle_rx(struct vhost_net *net)
 747 {
 748         struct vhost_net_virtqueue *nvq = &net->vqs[VHOST_NET_VQ_RX];
 749         struct vhost_virtqueue *vq = &nvq->vq;
 750         unsigned uninitialized_var(in), log;
 751         struct vhost_log *vq_log;
 752         struct msghdr msg = {
 753                 .msg_name = NULL,
 754                 .msg_namelen = 0,
 755                 .msg_control = NULL, /* FIXME: get and handle RX aux data. */
 756                 .msg_controllen = 0,
 757                 .msg_flags = MSG_DONTWAIT,
 758         };
 759         struct virtio_net_hdr hdr = {
 760                 .flags = 0,
 761                 .gso_type = VIRTIO_NET_HDR_GSO_NONE
 762         };
 763         size_t total_len = 0;
 764         int err, mergeable;
 765         s16 headcount, nheads = 0;
 766         size_t vhost_hlen, sock_hlen;
 767         size_t vhost_len, sock_len;
 768         struct socket *sock;
 769         struct iov_iter fixup;
 770         __virtio16 num_buffers;
 771
 772         mutex_lock_nested(&vq->mutex, 0);
 773         sock = vq->private_data;
 774         if (!sock)
 775                 goto out;
 776
 777         if (!vq_iotlb_prefetch(vq))
 778                 goto out;
 779
 780         vhost_disable_notify(&net->dev, vq);
 781         vhost_net_disable_vq(net, vq);
 782
 783         vhost_hlen = nvq->vhost_hlen;
 784         sock_hlen = nvq->sock_hlen;
 785
 786         vq_log = unlikely(vhost_has_feature(vq, VHOST_F_LOG_ALL)) ?
 787                 vq->log : NULL;
 788         mergeable = vhost_has_feature(vq, VIRTIO_NET_F_MRG_RXBUF);
 789
 790         while ((sock_len = vhost_net_rx_peek_head_len(net, sock->sk))) {
 791                 sock_len += sock_hlen;
 792                 vhost_len = sock_len + vhost_hlen;
 793                 headcount = get_rx_bufs(vq, vq->heads + nheads, vhost_len,
 794                                         &in, vq_log, &log,
 795                                         likely(mergeable) ? UIO_MAXIOV : 1);
 796                 /* On error, stop handling until the next kick. */
 797                 if (unlikely(headcount < 0))
 798                         goto out;
 799                 /* OK, now we need to know about added descriptors. */
 800                 if (!headcount) {
 801                         if (unlikely(vhost_enable_notify(&net->dev, vq))) {
 802                                 /* They have slipped one in as we were
 803                                  * doing that: check again. */
 804                                 vhost_disable_notify(&net->dev, vq);
 805                                 continue;
 806                         }
 807                         /* Nothing new?  Wait for eventfd to tell us
 808                          * they refilled. */
 809                         goto out;
 810                 }
 811                 if (nvq->rx_ring)
 812                         msg.msg_control = vhost_net_buf_consume(&nvq->rxq);
 813                 /* On overrun, truncate and discard */
 814                 if (unlikely(headcount > UIO_MAXIOV)) {
 815                         iov_iter_init(&msg.msg_iter, READ, vq->iov, 1, 1);
 816                         err = sock->ops->recvmsg(sock, &msg,
 817                                                  1, MSG_DONTWAIT | MSG_TRUNC);
 818                         pr_debug("Discarded rx packet: len %zd\n", sock_len);
 819                         continue;
 820                 }
 821                 /* We don't need to be notified again. */
 822                 iov_iter_init(&msg.msg_iter, READ, vq->iov, in, vhost_len);
 823                 fixup = msg.msg_iter;
 824                 if (unlikely((vhost_hlen))) {
 825                         /* We will supply the header ourselves
 826                          * TODO: support TSO.
 827                          */
 828                         iov_iter_advance(&msg.msg_iter, vhost_hlen);
 829                 }
 830                 err = sock->ops->recvmsg(sock, &msg,
 831                                          sock_len, MSG_DONTWAIT | MSG_TRUNC);
 832                 /* Userspace might have consumed the packet meanwhile:
 833                  * it's not supposed to do this usually, but might be hard
 834                  * to prevent. Discard data we got (if any) and keep going. */
 835                 if (unlikely(err != sock_len)) {
 836                         pr_debug("Discarded rx packet: "
 837                                  " len %d, expected %zd\n", err, sock_len);
 838                         vhost_discard_vq_desc(vq, headcount);
 839                         continue;
 840                 }
 841                 /* Supply virtio_net_hdr if VHOST_NET_F_VIRTIO_NET_HDR */
 842                 if (unlikely(vhost_hlen)) {
 843                         if (copy_to_iter(&hdr, sizeof(hdr),
 844                                          &fixup) != sizeof(hdr)) {
 845                                 vq_err(vq, "Unable to write vnet_hdr "
 846                                        "at addr %p\n", vq->iov->iov_base);
 847                                 goto out;
 848                         }
 849                 } else {
 850                         /* Header came from socket; we'll need to patch
 851                          * ->num_buffers over if VIRTIO_NET_F_MRG_RXBUF
 852                          */
 853                         iov_iter_advance(&fixup, sizeof(hdr));
 854                 }
 855                 /* TODO: Should check and handle checksum. */
 856
 857                 num_buffers = cpu_to_vhost16(vq, headcount);
 858                 if (likely(mergeable) &&
 859                     copy_to_iter(&num_buffers, sizeof num_buffers,
 860                                  &fixup) != sizeof num_buffers) {
 861                         vq_err(vq, "Failed num_buffers write");
 862                         vhost_discard_vq_desc(vq, headcount);
 863                         goto out;
 864                 }
 865                 nheads += headcount;
 866                 if (nheads > VHOST_RX_BATCH) {
 867                         vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 868                                                     nheads);
 869                         nheads = 0;
 870                 }
 871                 if (unlikely(vq_log))
 872                         vhost_log_write(vq, vq_log, log, vhost_len);
 873                 total_len += vhost_len;
 874                 if (unlikely(total_len >= VHOST_NET_WEIGHT)) {
 875                         vhost_poll_queue(&vq->poll);
 876                         goto out;
 877                 }
 878         }
 879         vhost_net_enable_vq(net, vq);
 880 out:
 881         if (nheads)
 882                 vhost_add_used_and_signal_n(&net->dev, vq, vq->heads,
 883                                             nheads);
 884         mutex_unlock(&vq->mutex);
 885 }
 886
 887 static void handle_tx_kick(struct vhost_work *work)
 888 {
 889         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 890                                                   poll.work);
 891         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 892
 893         handle_tx(net);
 894 }
 895
 896 static void handle_rx_kick(struct vhost_work *work)
 897 {
 898         struct vhost_virtqueue *vq = container_of(work, struct vhost_virtqueue,
 899                                                   poll.work);
 900         struct vhost_net *net = container_of(vq->dev, struct vhost_net, dev);
 901
 902         handle_rx(net);
 903 }
 904
 905 static void handle_tx_net(struct vhost_work *work)
 906 {
 907         struct vhost_net *net = container_of(work, struct vhost_net,
 908                                              poll[VHOST_NET_VQ_TX].work);
 909         handle_tx(net);
 910 }
 911
 912 static void handle_rx_net(struct vhost_work *work)
 913 {
 914         struct vhost_net *net = container_of(work, struct vhost_net,
 915                                              poll[VHOST_NET_VQ_RX].work);
 916         handle_rx(net);
 917 }
 918
 919 static int vhost_net_open(struct inode *inode, struct file *f)
 920 {
 921         struct vhost_net *n;
 922         struct vhost_dev *dev;
 923         struct vhost_virtqueue **vqs;
 924         void **queue;
 925         int i;
 926
 927         n = kvmalloc(sizeof *n, GFP_KERNEL | __GFP_RETRY_MAYFAIL);
 928         if (!n)
 929                 return -ENOMEM;
 930         vqs = kmalloc(VHOST_NET_VQ_MAX * sizeof(*vqs), GFP_KERNEL);
 931         if (!vqs) {
 932                 kvfree(n);
 933                 return -ENOMEM;
 934         }
 935
 936         queue = kmalloc_array(VHOST_RX_BATCH, sizeof(void *),
 937                               GFP_KERNEL);
 938         if (!queue) {
 939                 kfree(vqs);
 940                 kvfree(n);
 941                 return -ENOMEM;
 942         }
 943         n->vqs[VHOST_NET_VQ_RX].rxq.queue = queue;
 944
 945         dev = &n->dev;
 946         vqs[VHOST_NET_VQ_TX] = &n->vqs[VHOST_NET_VQ_TX].vq;
 947         vqs[VHOST_NET_VQ_RX] = &n->vqs[VHOST_NET_VQ_RX].vq;
 948         n->vqs[VHOST_NET_VQ_TX].vq.handle_kick = handle_tx_kick;
 949         n->vqs[VHOST_NET_VQ_RX].vq.handle_kick = handle_rx_kick;
 950         for (i = 0; i < VHOST_NET_VQ_MAX; i++) {
 951                 n->vqs[i].ubufs = NULL;
 952                 n->vqs[i].ubuf_info = NULL;
 953                 n->vqs[i].upend_idx = 0;
 954                 n->vqs[i].done_idx = 0;
 955                 n->vqs[i].vhost_hlen = 0;
 956                 n->vqs[i].sock_hlen = 0;
 957                 n->vqs[i].rx_ring = NULL;
 958                 vhost_net_buf_init(&n->vqs[i].rxq);
 959         }
 960         vhost_dev_init(dev, vqs, VHOST_NET_VQ_MAX);
 961
 962         vhost_poll_init(n->poll + VHOST_NET_VQ_TX, handle_tx_net, EPOLLOUT, dev);
 963         vhost_poll_init(n->poll + VHOST_NET_VQ_RX, handle_rx_net, EPOLLIN, dev);
 964
 965         f->private_data = n;
 966
 967         return 0;
 968 }
 969
 970 static struct socket *vhost_net_stop_vq(struct vhost_net *n,
 971                                         struct vhost_virtqueue *vq)
 972 {
 973         struct socket *sock;
 974         struct vhost_net_virtqueue *nvq =
 975                 container_of(vq, struct vhost_net_virtqueue, vq);
 976
 977         mutex_lock(&vq->mutex);
 978         sock = vq->private_data;
 979         vhost_net_disable_vq(n, vq);
 980         vq->private_data = NULL;
 981         vhost_net_buf_unproduce(nvq);
 982         nvq->rx_ring = NULL;
 983         mutex_unlock(&vq->mutex);
 984         return sock;
 985 }
 986
 987 static void vhost_net_stop(struct vhost_net *n, struct socket **tx_sock,
 988                            struct socket **rx_sock)
 989 {
 990         *tx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_TX].vq);
 991         *rx_sock = vhost_net_stop_vq(n, &n->vqs[VHOST_NET_VQ_RX].vq);
 992 }
 993
 994 static void vhost_net_flush_vq(struct vhost_net *n, int index)
 995 {
 996         vhost_poll_flush(n->poll + index);
 997         vhost_poll_flush(&n->vqs[index].vq.poll);
 998 }
 999
1000 static void vhost_net_flush(struct vhost_net *n)
1001 {
1002         vhost_net_flush_vq(n, VHOST_NET_VQ_TX);
1003         vhost_net_flush_vq(n, VHOST_NET_VQ_RX);
1004         if (n->vqs[VHOST_NET_VQ_TX].ubufs) {
1005                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1006                 n->tx_flush = true;
1007                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1008                 /* Wait for all lower device DMAs done. */
1009                 vhost_net_ubuf_put_and_wait(n->vqs[VHOST_NET_VQ_TX].ubufs);
1010                 mutex_lock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1011                 n->tx_flush = false;
1012                 atomic_set(&n->vqs[VHOST_NET_VQ_TX].ubufs->refcount, 1);
1013                 mutex_unlock(&n->vqs[VHOST_NET_VQ_TX].vq.mutex);
1014         }
1015 }
1016
1017 static int vhost_net_release(struct inode *inode, struct file *f)
1018 {
1019         struct vhost_net *n = f->private_data;
1020         struct socket *tx_sock;
1021         struct socket *rx_sock;
1022
1023         vhost_net_stop(n, &tx_sock, &rx_sock);
1024         vhost_net_flush(n);
1025         vhost_dev_stop(&n->dev);
1026         vhost_dev_cleanup(&n->dev);
1027         vhost_net_vq_reset(n);
1028         if (tx_sock)
1029                 sockfd_put(tx_sock);
1030         if (rx_sock)
1031                 sockfd_put(rx_sock);
1032         /* Make sure no callbacks are outstanding */
1033         synchronize_rcu_bh();
1034         /* We do an extra flush before freeing memory,
1035          * since jobs can re-queue themselves. */
1036         vhost_net_flush(n);
1037         kfree(n->vqs[VHOST_NET_VQ_RX].rxq.queue);
1038         kfree(n->dev.vqs);
1039         kvfree(n);
1040         return 0;
1041 }
1042
1043 static struct socket *get_raw_socket(int fd)
1044 {
1045         struct {
1046                 struct sockaddr_ll sa;
1047                 char  buf[MAX_ADDR_LEN];
1048         } uaddr;
1049         int r;
1050         struct socket *sock = sockfd_lookup(fd, &r);
1051
1052         if (!sock)
1053                 return ERR_PTR(-ENOTSOCK);
1054
1055         /* Parameter checking */
1056         if (sock->sk->sk_type != SOCK_RAW) {
1057                 r = -ESOCKTNOSUPPORT;
1058                 goto err;
1059         }
1060
1061         r = sock->ops->getname(sock, (struct sockaddr *)&uaddr.sa, 0);
1062         if (r < 0)
1063                 goto err;
1064
1065         if (uaddr.sa.sll_family != AF_PACKET) {
1066                 r = -EPFNOSUPPORT;
1067                 goto err;
1068         }
1069         return sock;
1070 err:
1071         sockfd_put(sock);
1072         return ERR_PTR(r);
1073 }
1074
1075 static struct ptr_ring *get_tap_ptr_ring(int fd)
1076 {
1077         struct ptr_ring *ring;
1078         struct file *file = fget(fd);
1079
1080         if (!file)
1081                 return NULL;
1082         ring = tun_get_tx_ring(file);
1083         if (!IS_ERR(ring))
1084                 goto out;
1085         ring = tap_get_ptr_ring(file);
1086         if (!IS_ERR(ring))
1087                 goto out;
1088         ring = NULL;
1089 out:
1090         fput(file);
1091         return ring;
1092 }
1093
1094 static struct socket *get_tap_socket(int fd)
1095 {
1096         struct file *file = fget(fd);
1097         struct socket *sock;
1098
1099         if (!file)
1100                 return ERR_PTR(-EBADF);
1101         sock = tun_get_socket(file);
1102         if (!IS_ERR(sock))
1103                 return sock;
1104         sock = tap_get_socket(file);
1105         if (IS_ERR(sock))
1106                 fput(file);
1107         return sock;
1108 }
1109
1110 static struct socket *get_socket(int fd)
1111 {
1112         struct socket *sock;
1113
1114         /* special case to disable backend */
1115         if (fd == -1)
1116                 return NULL;
1117         sock = get_raw_socket(fd);
1118         if (!IS_ERR(sock))
1119                 return sock;
1120         sock = get_tap_socket(fd);
1121         if (!IS_ERR(sock))
1122                 return sock;
1123         return ERR_PTR(-ENOTSOCK);
1124 }
1125
1126 static long vhost_net_set_backend(struct vhost_net *n, unsigned index, int fd)
1127 {
1128         struct socket *sock, *oldsock;
1129         struct vhost_virtqueue *vq;
1130         struct vhost_net_virtqueue *nvq;
1131         struct vhost_net_ubuf_ref *ubufs, *oldubufs = NULL;
1132         int r;
1133
1134         mutex_lock(&n->dev.mutex);
1135         r = vhost_dev_check_owner(&n->dev);
1136         if (r)
1137                 goto err;
1138
1139         if (index >= VHOST_NET_VQ_MAX) {
1140                 r = -ENOBUFS;
1141                 goto err;
1142         }
1143         vq = &n->vqs[index].vq;
1144         nvq = &n->vqs[index];
1145         mutex_lock(&vq->mutex);
1146
1147         /* Verify that ring has been setup correctly. */
1148         if (!vhost_vq_access_ok(vq)) {
1149                 r = -EFAULT;
1150                 goto err_vq;
1151         }
1152         sock = get_socket(fd);
1153         if (IS_ERR(sock)) {
1154                 r = PTR_ERR(sock);
1155                 goto err_vq;
1156         }
1157
1158         /* start polling new socket */
1159         oldsock = vq->private_data;
1160         if (sock != oldsock) {
1161                 ubufs = vhost_net_ubuf_alloc(vq,
1162                                              sock && vhost_sock_zcopy(sock));
1163                 if (IS_ERR(ubufs)) {
1164                         r = PTR_ERR(ubufs);
1165                         goto err_ubufs;
1166                 }
1167
1168                 vhost_net_disable_vq(n, vq);
1169                 vq->private_data = sock;
1170                 vhost_net_buf_unproduce(nvq);
1171                 r = vhost_vq_init_access(vq);
1172                 if (r)
1173                         goto err_used;
1174                 r = vhost_net_enable_vq(n, vq);
1175                 if (r)
1176                         goto err_used;
1177                 if (index == VHOST_NET_VQ_RX)
1178                         nvq->rx_ring = get_tap_ptr_ring(fd);
1179
1180                 oldubufs = nvq->ubufs;
1181                 nvq->ubufs = ubufs;
1182
1183                 n->tx_packets = 0;
1184                 n->tx_zcopy_err = 0;
1185                 n->tx_flush = false;
1186         }
1187
1188         mutex_unlock(&vq->mutex);
1189
1190         if (oldubufs) {
1191                 vhost_net_ubuf_put_wait_and_free(oldubufs);
1192                 mutex_lock(&vq->mutex);
1193                 vhost_zerocopy_signal_used(n, vq);
1194                 mutex_unlock(&vq->mutex);
1195         }
1196
1197         if (oldsock) {
1198                 vhost_net_flush_vq(n, index);
1199                 sockfd_put(oldsock);
1200         }
1201
1202         mutex_unlock(&n->dev.mutex);
1203         return 0;
1204
1205 err_used:
1206         vq->private_data = oldsock;
1207         vhost_net_enable_vq(n, vq);
1208         if (ubufs)
1209                 vhost_net_ubuf_put_wait_and_free(ubufs);
1210 err_ubufs:
1211         sockfd_put(sock);
1212 err_vq:
1213         mutex_unlock(&vq->mutex);
1214 err:
1215         mutex_unlock(&n->dev.mutex);
1216         return r;
1217 }
1218
1219 static long vhost_net_reset_owner(struct vhost_net *n)
1220 {
1221         struct socket *tx_sock = NULL;
1222         struct socket *rx_sock = NULL;
1223         long err;
1224         struct vhost_umem *umem;
1225
1226         mutex_lock(&n->dev.mutex);
1227         err = vhost_dev_check_owner(&n->dev);
1228         if (err)
1229                 goto done;
1230         umem = vhost_dev_reset_owner_prepare();
1231         if (!umem) {
1232                 err = -ENOMEM;
1233                 goto done;
1234         }
1235         vhost_net_stop(n, &tx_sock, &rx_sock);
1236         vhost_net_flush(n);
1237         vhost_dev_stop(&n->dev);
1238         vhost_dev_reset_owner(&n->dev, umem);
1239         vhost_net_vq_reset(n);
1240 done:
1241         mutex_unlock(&n->dev.mutex);
1242         if (tx_sock)
1243                 sockfd_put(tx_sock);
1244         if (rx_sock)
1245                 sockfd_put(rx_sock);
1246         return err;
1247 }
1248
1249 static int vhost_net_set_features(struct vhost_net *n, u64 features)
1250 {
1251         size_t vhost_hlen, sock_hlen, hdr_len;
1252         int i;
1253
1254         hdr_len = (features & ((1ULL << VIRTIO_NET_F_MRG_RXBUF) |
1255                                (1ULL << VIRTIO_F_VERSION_1))) ?
1256                         sizeof(struct virtio_net_hdr_mrg_rxbuf) :
1257                         sizeof(struct virtio_net_hdr);
1258         if (features & (1 << VHOST_NET_F_VIRTIO_NET_HDR)) {
1259                 /* vhost provides vnet_hdr */
1260                 vhost_hlen = hdr_len;
1261                 sock_hlen = 0;
1262         } else {
1263                 /* socket provides vnet_hdr */
1264                 vhost_hlen = 0;
1265                 sock_hlen = hdr_len;
1266         }
1267         mutex_lock(&n->dev.mutex);
1268         if ((features & (1 << VHOST_F_LOG_ALL)) &&
1269             !vhost_log_access_ok(&n->dev))
1270                 goto out_unlock;
1271
1272         if ((features & (1ULL << VIRTIO_F_IOMMU_PLATFORM))) {
1273                 if (vhost_init_device_iotlb(&n->dev, true))
1274                         goto out_unlock;
1275         }
1276
1277         for (i = 0; i < VHOST_NET_VQ_MAX; ++i) {
1278                 mutex_lock(&n->vqs[i].vq.mutex);
1279                 n->vqs[i].vq.acked_features = features;
1280                 n->vqs[i].vhost_hlen = vhost_hlen;
1281                 n->vqs[i].sock_hlen = sock_hlen;
1282                 mutex_unlock(&n->vqs[i].vq.mutex);
1283         }
1284         mutex_unlock(&n->dev.mutex);
1285         return 0;
1286
1287 out_unlock:
1288         mutex_unlock(&n->dev.mutex);
1289         return -EFAULT;
1290 }
1291
1292 static long vhost_net_set_owner(struct vhost_net *n)
1293 {
1294         int r;
1295
1296         mutex_lock(&n->dev.mutex);
1297         if (vhost_dev_has_owner(&n->dev)) {
1298                 r = -EBUSY;
1299                 goto out;
1300         }
1301         r = vhost_net_set_ubuf_info(n);
1302         if (r)
1303                 goto out;
1304         r = vhost_dev_set_owner(&n->dev);
1305         if (r)
1306                 vhost_net_clear_ubuf_info(n);
1307         vhost_net_flush(n);
1308 out:
1309         mutex_unlock(&n->dev.mutex);
1310         return r;
1311 }
1312
1313 static long vhost_net_ioctl(struct file *f, unsigned int ioctl,
1314                             unsigned long arg)
1315 {
1316         struct vhost_net *n = f->private_data;
1317         void __user *argp = (void __user *)arg;
1318         u64 __user *featurep = argp;
1319         struct vhost_vring_file backend;
1320         u64 features;
1321         int r;
1322
1323         switch (ioctl) {
1324         case VHOST_NET_SET_BACKEND:
1325                 if (copy_from_user(&backend, argp, sizeof backend))
1326                         return -EFAULT;
1327                 return vhost_net_set_backend(n, backend.index, backend.fd);
1328         case VHOST_GET_FEATURES:
1329                 features = VHOST_NET_FEATURES;
1330                 if (copy_to_user(featurep, &features, sizeof features))
1331                         return -EFAULT;
1332                 return 0;
1333         case VHOST_SET_FEATURES:
1334                 if (copy_from_user(&features, featurep, sizeof features))
1335                         return -EFAULT;
1336                 if (features & ~VHOST_NET_FEATURES)
1337                         return -EOPNOTSUPP;
1338                 return vhost_net_set_features(n, features);
1339         case VHOST_RESET_OWNER:
1340                 return vhost_net_reset_owner(n);
1341         case VHOST_SET_OWNER:
1342                 return vhost_net_set_owner(n);
1343         default:
1344                 mutex_lock(&n->dev.mutex);
1345                 r = vhost_dev_ioctl(&n->dev, ioctl, argp);
1346                 if (r == -ENOIOCTLCMD)
1347                         r = vhost_vring_ioctl(&n->dev, ioctl, argp);
1348                 else
1349                         vhost_net_flush(n);
1350                 mutex_unlock(&n->dev.mutex);
1351                 return r;
1352         }
1353 }
1354
1355 #ifdef CONFIG_COMPAT
1356 static long vhost_net_compat_ioctl(struct file *f, unsigned int ioctl,
1357                                    unsigned long arg)
1358 {
1359         return vhost_net_ioctl(f, ioctl, (unsigned long)compat_ptr(arg));
1360 }
1361 #endif
1362
1363 static ssize_t vhost_net_chr_read_iter(struct kiocb *iocb, struct iov_iter *to)
1364 {
1365         struct file *file = iocb->ki_filp;
1366         struct vhost_net *n = file->private_data;
1367         struct vhost_dev *dev = &n->dev;
1368         int noblock = file->f_flags & O_NONBLOCK;
1369
1370         return vhost_chr_read_iter(dev, to, noblock);
1371 }
1372
1373 static ssize_t vhost_net_chr_write_iter(struct kiocb *iocb,
1374                                         struct iov_iter *from)
1375 {
1376         struct file *file = iocb->ki_filp;
1377         struct vhost_net *n = file->private_data;
1378         struct vhost_dev *dev = &n->dev;
1379
1380         return vhost_chr_write_iter(dev, from);
1381 }
1382
1383 static __poll_t vhost_net_chr_poll(struct file *file, poll_table *wait)
1384 {
1385         struct vhost_net *n = file->private_data;
1386         struct vhost_dev *dev = &n->dev;
1387
1388         return vhost_chr_poll(file, dev, wait);
1389 }
1390
1391 static const struct file_operations vhost_net_fops = {
1392         .owner          = THIS_MODULE,
1393         .release        = vhost_net_release,
1394         .read_iter      = vhost_net_chr_read_iter,
1395         .write_iter     = vhost_net_chr_write_iter,
1396         .poll           = vhost_net_chr_poll,
1397         .unlocked_ioctl = vhost_net_ioctl,
1398 #ifdef CONFIG_COMPAT
1399         .compat_ioctl   = vhost_net_compat_ioctl,
1400 #endif
1401         .open           = vhost_net_open,
1402         .llseek         = noop_llseek,
1403 };
1404
1405 static struct miscdevice vhost_net_misc = {
1406         .minor = VHOST_NET_MINOR,
1407         .name = "vhost-net",
1408         .fops = &vhost_net_fops,
1409 };
1410
1411 static int vhost_net_init(void)
1412 {
1413         if (experimental_zcopytx)
1414                 vhost_net_enable_zcopy(VHOST_NET_VQ_TX);
1415         return misc_register(&vhost_net_misc);
1416 }
1417 module_init(vhost_net_init);
1418
1419 static void vhost_net_exit(void)
1420 {
1421         misc_deregister(&vhost_net_misc);
1422 }
1423 module_exit(vhost_net_exit);
1424
1425 MODULE_VERSION("0.0.1");
1426 MODULE_LICENSE("GPL v2");
1427 MODULE_AUTHOR("Michael S. Tsirkin");
1428 MODULE_DESCRIPTION("Host kernel accelerator for virtio net");
1429 MODULE_ALIAS_MISCDEV(VHOST_NET_MINOR);
1430 MODULE_ALIAS("devname:vhost-net");