drivers/net/tap.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 #include <linux/etherdevice.h>
   3 #include <linux/if_tap.h>
   4 #include <linux/if_vlan.h>
   5 #include <linux/interrupt.h>
   6 #include <linux/nsproxy.h>
   7 #include <linux/compat.h>
   8 #include <linux/if_tun.h>
   9 #include <linux/module.h>
  10 #include <linux/skbuff.h>
  11 #include <linux/cache.h>
  12 #include <linux/sched/signal.h>
  13 #include <linux/types.h>
  14 #include <linux/slab.h>
  15 #include <linux/wait.h>
  16 #include <linux/cdev.h>
  17 #include <linux/idr.h>
  18 #include <linux/fs.h>
  19 #include <linux/uio.h>
  20
  21 #include <net/net_namespace.h>
  22 #include <net/rtnetlink.h>
  23 #include <net/sock.h>
  24 #include <linux/virtio_net.h>
  25 #include <linux/skb_array.h>
  26
  27 #define TAP_IFFEATURES (IFF_VNET_HDR | IFF_MULTI_QUEUE)
  28
  29 #define TAP_VNET_LE 0x80000000
  30 #define TAP_VNET_BE 0x40000000
  31
  32 #ifdef CONFIG_TUN_VNET_CROSS_LE
  33 static inline bool tap_legacy_is_little_endian(struct tap_queue *q)
  34 {
  35         return q->flags & TAP_VNET_BE ? false :
  36                 virtio_legacy_is_little_endian();
  37 }
  38
  39 static long tap_get_vnet_be(struct tap_queue *q, int __user *sp)
  40 {
  41         int s = !!(q->flags & TAP_VNET_BE);
  42
  43         if (put_user(s, sp))
  44                 return -EFAULT;
  45
  46         return 0;
  47 }
  48
  49 static long tap_set_vnet_be(struct tap_queue *q, int __user *sp)
  50 {
  51         int s;
  52
  53         if (get_user(s, sp))
  54                 return -EFAULT;
  55
  56         if (s)
  57                 q->flags |= TAP_VNET_BE;
  58         else
  59                 q->flags &= ~TAP_VNET_BE;
  60
  61         return 0;
  62 }
  63 #else
  64 static inline bool tap_legacy_is_little_endian(struct tap_queue *q)
  65 {
  66         return virtio_legacy_is_little_endian();
  67 }
  68
  69 static long tap_get_vnet_be(struct tap_queue *q, int __user *argp)
  70 {
  71         return -EINVAL;
  72 }
  73
  74 static long tap_set_vnet_be(struct tap_queue *q, int __user *argp)
  75 {
  76         return -EINVAL;
  77 }
  78 #endif /* CONFIG_TUN_VNET_CROSS_LE */
  79
  80 static inline bool tap_is_little_endian(struct tap_queue *q)
  81 {
  82         return q->flags & TAP_VNET_LE ||
  83                 tap_legacy_is_little_endian(q);
  84 }
  85
  86 static inline u16 tap16_to_cpu(struct tap_queue *q, __virtio16 val)
  87 {
  88         return __virtio16_to_cpu(tap_is_little_endian(q), val);
  89 }
  90
  91 static inline __virtio16 cpu_to_tap16(struct tap_queue *q, u16 val)
  92 {
  93         return __cpu_to_virtio16(tap_is_little_endian(q), val);
  94 }
  95
  96 static struct proto tap_proto = {
  97         .name = "tap",
  98         .owner = THIS_MODULE,
  99         .obj_size = sizeof(struct tap_queue),
 100 };
 101
 102 #define TAP_NUM_DEVS (1U << MINORBITS)
 103
 104 static LIST_HEAD(major_list);
 105
 106 struct major_info {
 107         struct rcu_head rcu;
 108         dev_t major;
 109         struct idr minor_idr;
 110         spinlock_t minor_lock;
 111         const char *device_name;
 112         struct list_head next;
 113 };
 114
 115 #define GOODCOPY_LEN 128
 116
 117 static const struct proto_ops tap_socket_ops;
 118
 119 #define RX_OFFLOADS (NETIF_F_GRO | NETIF_F_LRO)
 120 #define TAP_FEATURES (NETIF_F_GSO | NETIF_F_SG | NETIF_F_FRAGLIST)
 121
 122 static struct tap_dev *tap_dev_get_rcu(const struct net_device *dev)
 123 {
 124         return rcu_dereference(dev->rx_handler_data);
 125 }
 126
 127 /*
 128  * RCU usage:
 129  * The tap_queue and the macvlan_dev are loosely coupled, the
 130  * pointers from one to the other can only be read while rcu_read_lock
 131  * or rtnl is held.
 132  *
 133  * Both the file and the macvlan_dev hold a reference on the tap_queue
 134  * through sock_hold(&q->sk). When the macvlan_dev goes away first,
 135  * q->vlan becomes inaccessible. When the files gets closed,
 136  * tap_get_queue() fails.
 137  *
 138  * There may still be references to the struct sock inside of the
 139  * queue from outbound SKBs, but these never reference back to the
 140  * file or the dev. The data structure is freed through __sk_free
 141  * when both our references and any pending SKBs are gone.
 142  */
 143
 144 static int tap_enable_queue(struct tap_dev *tap, struct file *file,
 145                             struct tap_queue *q)
 146 {
 147         int err = -EINVAL;
 148
 149         ASSERT_RTNL();
 150
 151         if (q->enabled)
 152                 goto out;
 153
 154         err = 0;
 155         rcu_assign_pointer(tap->taps[tap->numvtaps], q);
 156         q->queue_index = tap->numvtaps;
 157         q->enabled = true;
 158
 159         tap->numvtaps++;
 160 out:
 161         return err;
 162 }
 163
 164 /* Requires RTNL */
 165 static int tap_set_queue(struct tap_dev *tap, struct file *file,
 166                          struct tap_queue *q)
 167 {
 168         if (tap->numqueues == MAX_TAP_QUEUES)
 169                 return -EBUSY;
 170
 171         rcu_assign_pointer(q->tap, tap);
 172         rcu_assign_pointer(tap->taps[tap->numvtaps], q);
 173         sock_hold(&q->sk);
 174
 175         q->file = file;
 176         q->queue_index = tap->numvtaps;
 177         q->enabled = true;
 178         file->private_data = q;
 179         list_add_tail(&q->next, &tap->queue_list);
 180
 181         tap->numvtaps++;
 182         tap->numqueues++;
 183
 184         return 0;
 185 }
 186
 187 static int tap_disable_queue(struct tap_queue *q)
 188 {
 189         struct tap_dev *tap;
 190         struct tap_queue *nq;
 191
 192         ASSERT_RTNL();
 193         if (!q->enabled)
 194                 return -EINVAL;
 195
 196         tap = rtnl_dereference(q->tap);
 197
 198         if (tap) {
 199                 int index = q->queue_index;
 200                 BUG_ON(index >= tap->numvtaps);
 201                 nq = rtnl_dereference(tap->taps[tap->numvtaps - 1]);
 202                 nq->queue_index = index;
 203
 204                 rcu_assign_pointer(tap->taps[index], nq);
 205                 RCU_INIT_POINTER(tap->taps[tap->numvtaps - 1], NULL);
 206                 q->enabled = false;
 207
 208                 tap->numvtaps--;
 209         }
 210
 211         return 0;
 212 }
 213
 214 /*
 215  * The file owning the queue got closed, give up both
 216  * the reference that the files holds as well as the
 217  * one from the macvlan_dev if that still exists.
 218  *
 219  * Using the spinlock makes sure that we don't get
 220  * to the queue again after destroying it.
 221  */
 222 static void tap_put_queue(struct tap_queue *q)
 223 {
 224         struct tap_dev *tap;
 225
 226         rtnl_lock();
 227         tap = rtnl_dereference(q->tap);
 228
 229         if (tap) {
 230                 if (q->enabled)
 231                         BUG_ON(tap_disable_queue(q));
 232
 233                 tap->numqueues--;
 234                 RCU_INIT_POINTER(q->tap, NULL);
 235                 sock_put(&q->sk);
 236                 list_del_init(&q->next);
 237         }
 238
 239         rtnl_unlock();
 240
 241         synchronize_rcu();
 242         sock_put(&q->sk);
 243 }
 244
 245 /*
 246  * Select a queue based on the rxq of the device on which this packet
 247  * arrived. If the incoming device is not mq, calculate a flow hash
 248  * to select a queue. If all fails, find the first available queue.
 249  * Cache vlan->numvtaps since it can become zero during the execution
 250  * of this function.
 251  */
 252 static struct tap_queue *tap_get_queue(struct tap_dev *tap,
 253                                        struct sk_buff *skb)
 254 {
 255         struct tap_queue *queue = NULL;
 256         /* Access to taps array is protected by rcu, but access to numvtaps
 257          * isn't. Below we use it to lookup a queue, but treat it as a hint
 258          * and validate that the result isn't NULL - in case we are
 259          * racing against queue removal.
 260          */
 261         int numvtaps = READ_ONCE(tap->numvtaps);
 262         __u32 rxq;
 263
 264         if (!numvtaps)
 265                 goto out;
 266
 267         if (numvtaps == 1)
 268                 goto single;
 269
 270         /* Check if we can use flow to select a queue */
 271         rxq = skb_get_hash(skb);
 272         if (rxq) {
 273                 queue = rcu_dereference(tap->taps[rxq % numvtaps]);
 274                 goto out;
 275         }
 276
 277         if (likely(skb_rx_queue_recorded(skb))) {
 278                 rxq = skb_get_rx_queue(skb);
 279
 280                 while (unlikely(rxq >= numvtaps))
 281                         rxq -= numvtaps;
 282
 283                 queue = rcu_dereference(tap->taps[rxq]);
 284                 goto out;
 285         }
 286
 287 single:
 288         queue = rcu_dereference(tap->taps[0]);
 289 out:
 290         return queue;
 291 }
 292
 293 /*
 294  * The net_device is going away, give up the reference
 295  * that it holds on all queues and safely set the pointer
 296  * from the queues to NULL.
 297  */
 298 void tap_del_queues(struct tap_dev *tap)
 299 {
 300         struct tap_queue *q, *tmp;
 301
 302         ASSERT_RTNL();
 303         list_for_each_entry_safe(q, tmp, &tap->queue_list, next) {
 304                 list_del_init(&q->next);
 305                 RCU_INIT_POINTER(q->tap, NULL);
 306                 if (q->enabled)
 307                         tap->numvtaps--;
 308                 tap->numqueues--;
 309                 sock_put(&q->sk);
 310         }
 311         BUG_ON(tap->numvtaps);
 312         BUG_ON(tap->numqueues);
 313         /* guarantee that any future tap_set_queue will fail */
 314         tap->numvtaps = MAX_TAP_QUEUES;
 315 }
 316 EXPORT_SYMBOL_GPL(tap_del_queues);
 317
 318 rx_handler_result_t tap_handle_frame(struct sk_buff **pskb)
 319 {
 320         struct sk_buff *skb = *pskb;
 321         struct net_device *dev = skb->dev;
 322         struct tap_dev *tap;
 323         struct tap_queue *q;
 324         netdev_features_t features = TAP_FEATURES;
 325
 326         tap = tap_dev_get_rcu(dev);
 327         if (!tap)
 328                 return RX_HANDLER_PASS;
 329
 330         q = tap_get_queue(tap, skb);
 331         if (!q)
 332                 return RX_HANDLER_PASS;
 333
 334         skb_push(skb, ETH_HLEN);
 335
 336         /* Apply the forward feature mask so that we perform segmentation
 337          * according to users wishes.  This only works if VNET_HDR is
 338          * enabled.
 339          */
 340         if (q->flags & IFF_VNET_HDR)
 341                 features |= tap->tap_features;
 342         if (netif_needs_gso(skb, features)) {
 343                 struct sk_buff *segs = __skb_gso_segment(skb, features, false);
 344                 struct sk_buff *next;
 345
 346                 if (IS_ERR(segs))
 347                         goto drop;
 348
 349                 if (!segs) {
 350                         if (ptr_ring_produce(&q->ring, skb))
 351                                 goto drop;
 352                         goto wake_up;
 353                 }
 354
 355                 consume_skb(skb);
 356                 skb_list_walk_safe(segs, skb, next) {
 357                         skb_mark_not_on_list(skb);
 358                         if (ptr_ring_produce(&q->ring, skb)) {
 359                                 kfree_skb(skb);
 360                                 kfree_skb_list(next);
 361                                 break;
 362                         }
 363                 }
 364         } else {
 365                 /* If we receive a partial checksum and the tap side
 366                  * doesn't support checksum offload, compute the checksum.
 367                  * Note: it doesn't matter which checksum feature to
 368                  *        check, we either support them all or none.
 369                  */
 370                 if (skb->ip_summed == CHECKSUM_PARTIAL &&
 371                     !(features & NETIF_F_CSUM_MASK) &&
 372                     skb_checksum_help(skb))
 373                         goto drop;
 374                 if (ptr_ring_produce(&q->ring, skb))
 375                         goto drop;
 376         }
 377
 378 wake_up:
 379         wake_up_interruptible_poll(sk_sleep(&q->sk), EPOLLIN | EPOLLRDNORM | EPOLLRDBAND);
 380         return RX_HANDLER_CONSUMED;
 381
 382 drop:
 383         /* Count errors/drops only here, thus don't care about args. */
 384         if (tap->count_rx_dropped)
 385                 tap->count_rx_dropped(tap);
 386         kfree_skb(skb);
 387         return RX_HANDLER_CONSUMED;
 388 }
 389 EXPORT_SYMBOL_GPL(tap_handle_frame);
 390
 391 static struct major_info *tap_get_major(int major)
 392 {
 393         struct major_info *tap_major;
 394
 395         list_for_each_entry_rcu(tap_major, &major_list, next) {
 396                 if (tap_major->major == major)
 397                         return tap_major;
 398         }
 399
 400         return NULL;
 401 }
 402
 403 int tap_get_minor(dev_t major, struct tap_dev *tap)
 404 {
 405         int retval = -ENOMEM;
 406         struct major_info *tap_major;
 407
 408         rcu_read_lock();
 409         tap_major = tap_get_major(MAJOR(major));
 410         if (!tap_major) {
 411                 retval = -EINVAL;
 412                 goto unlock;
 413         }
 414
 415         spin_lock(&tap_major->minor_lock);
 416         retval = idr_alloc(&tap_major->minor_idr, tap, 1, TAP_NUM_DEVS, GFP_ATOMIC);
 417         if (retval >= 0) {
 418                 tap->minor = retval;
 419         } else if (retval == -ENOSPC) {
 420                 netdev_err(tap->dev, "Too many tap devices\n");
 421                 retval = -EINVAL;
 422         }
 423         spin_unlock(&tap_major->minor_lock);
 424
 425 unlock:
 426         rcu_read_unlock();
 427         return retval < 0 ? retval : 0;
 428 }
 429 EXPORT_SYMBOL_GPL(tap_get_minor);
 430
 431 void tap_free_minor(dev_t major, struct tap_dev *tap)
 432 {
 433         struct major_info *tap_major;
 434
 435         rcu_read_lock();
 436         tap_major = tap_get_major(MAJOR(major));
 437         if (!tap_major) {
 438                 goto unlock;
 439         }
 440
 441         spin_lock(&tap_major->minor_lock);
 442         if (tap->minor) {
 443                 idr_remove(&tap_major->minor_idr, tap->minor);
 444                 tap->minor = 0;
 445         }
 446         spin_unlock(&tap_major->minor_lock);
 447
 448 unlock:
 449         rcu_read_unlock();
 450 }
 451 EXPORT_SYMBOL_GPL(tap_free_minor);
 452
 453 static struct tap_dev *dev_get_by_tap_file(int major, int minor)
 454 {
 455         struct net_device *dev = NULL;
 456         struct tap_dev *tap;
 457         struct major_info *tap_major;
 458
 459         rcu_read_lock();
 460         tap_major = tap_get_major(major);
 461         if (!tap_major) {
 462                 tap = NULL;
 463                 goto unlock;
 464         }
 465
 466         spin_lock(&tap_major->minor_lock);
 467         tap = idr_find(&tap_major->minor_idr, minor);
 468         if (tap) {
 469                 dev = tap->dev;
 470                 dev_hold(dev);
 471         }
 472         spin_unlock(&tap_major->minor_lock);
 473
 474 unlock:
 475         rcu_read_unlock();
 476         return tap;
 477 }
 478
 479 static void tap_sock_write_space(struct sock *sk)
 480 {
 481         wait_queue_head_t *wqueue;
 482
 483         if (!sock_writeable(sk) ||
 484             !test_and_clear_bit(SOCKWQ_ASYNC_NOSPACE, &sk->sk_socket->flags))
 485                 return;
 486
 487         wqueue = sk_sleep(sk);
 488         if (wqueue && waitqueue_active(wqueue))
 489                 wake_up_interruptible_poll(wqueue, EPOLLOUT | EPOLLWRNORM | EPOLLWRBAND);
 490 }
 491
 492 static void tap_sock_destruct(struct sock *sk)
 493 {
 494         struct tap_queue *q = container_of(sk, struct tap_queue, sk);
 495
 496         ptr_ring_cleanup(&q->ring, __skb_array_destroy_skb);
 497 }
 498
 499 static int tap_open(struct inode *inode, struct file *file)
 500 {
 501         struct net *net = current->nsproxy->net_ns;
 502         struct tap_dev *tap;
 503         struct tap_queue *q;
 504         int err = -ENODEV;
 505
 506         rtnl_lock();
 507         tap = dev_get_by_tap_file(imajor(inode), iminor(inode));
 508         if (!tap)
 509                 goto err;
 510
 511         err = -ENOMEM;
 512         q = (struct tap_queue *)sk_alloc(net, AF_UNSPEC, GFP_KERNEL,
 513                                              &tap_proto, 0);
 514         if (!q)
 515                 goto err;
 516         if (ptr_ring_init(&q->ring, tap->dev->tx_queue_len, GFP_KERNEL)) {
 517                 sk_free(&q->sk);
 518                 goto err;
 519         }
 520
 521         init_waitqueue_head(&q->sock.wq.wait);
 522         q->sock.type = SOCK_RAW;
 523         q->sock.state = SS_CONNECTED;
 524         q->sock.file = file;
 525         q->sock.ops = &tap_socket_ops;
 526         sock_init_data(&q->sock, &q->sk);
 527         q->sk.sk_write_space = tap_sock_write_space;
 528         q->sk.sk_destruct = tap_sock_destruct;
 529         q->flags = IFF_VNET_HDR | IFF_NO_PI | IFF_TAP;
 530         q->vnet_hdr_sz = sizeof(struct virtio_net_hdr);
 531
 532         /*
 533          * so far only KVM virtio_net uses tap, enable zero copy between
 534          * guest kernel and host kernel when lower device supports zerocopy
 535          *
 536          * The macvlan supports zerocopy iff the lower device supports zero
 537          * copy so we don't have to look at the lower device directly.
 538          */
 539         if ((tap->dev->features & NETIF_F_HIGHDMA) && (tap->dev->features & NETIF_F_SG))
 540                 sock_set_flag(&q->sk, SOCK_ZEROCOPY);
 541
 542         err = tap_set_queue(tap, file, q);
 543         if (err) {
 544                 /* tap_sock_destruct() will take care of freeing ptr_ring */
 545                 goto err_put;
 546         }
 547
 548         dev_put(tap->dev);
 549
 550         rtnl_unlock();
 551         return err;
 552
 553 err_put:
 554         sock_put(&q->sk);
 555 err:
 556         if (tap)
 557                 dev_put(tap->dev);
 558
 559         rtnl_unlock();
 560         return err;
 561 }
 562
 563 static int tap_release(struct inode *inode, struct file *file)
 564 {
 565         struct tap_queue *q = file->private_data;
 566         tap_put_queue(q);
 567         return 0;
 568 }
 569
 570 static __poll_t tap_poll(struct file *file, poll_table *wait)
 571 {
 572         struct tap_queue *q = file->private_data;
 573         __poll_t mask = EPOLLERR;
 574
 575         if (!q)
 576                 goto out;
 577
 578         mask = 0;
 579         poll_wait(file, &q->sock.wq.wait, wait);
 580
 581         if (!ptr_ring_empty(&q->ring))
 582                 mask |= EPOLLIN | EPOLLRDNORM;
 583
 584         if (sock_writeable(&q->sk) ||
 585             (!test_and_set_bit(SOCKWQ_ASYNC_NOSPACE, &q->sock.flags) &&
 586              sock_writeable(&q->sk)))
 587                 mask |= EPOLLOUT | EPOLLWRNORM;
 588
 589 out:
 590         return mask;
 591 }
 592
 593 static inline struct sk_buff *tap_alloc_skb(struct sock *sk, size_t prepad,
 594                                             size_t len, size_t linear,
 595                                                 int noblock, int *err)
 596 {
 597         struct sk_buff *skb;
 598
 599         /* Under a page?  Don't bother with paged skb. */
 600         if (prepad + len < PAGE_SIZE || !linear)
 601                 linear = len;
 602
 603         skb = sock_alloc_send_pskb(sk, prepad + linear, len - linear, noblock,
 604                                    err, 0);
 605         if (!skb)
 606                 return NULL;
 607
 608         skb_reserve(skb, prepad);
 609         skb_put(skb, linear);
 610         skb->data_len = len - linear;
 611         skb->len += len - linear;
 612
 613         return skb;
 614 }
 615
 616 /* Neighbour code has some assumptions on HH_DATA_MOD alignment */
 617 #define TAP_RESERVE HH_DATA_OFF(ETH_HLEN)
 618
 619 /* Get packet from user space buffer */
 620 static ssize_t tap_get_user(struct tap_queue *q, void *msg_control,
 621                             struct iov_iter *from, int noblock)
 622 {
 623         int good_linear = SKB_MAX_HEAD(TAP_RESERVE);
 624         struct sk_buff *skb;
 625         struct tap_dev *tap;
 626         unsigned long total_len = iov_iter_count(from);
 627         unsigned long len = total_len;
 628         int err;
 629         struct virtio_net_hdr vnet_hdr = { 0 };
 630         int vnet_hdr_len = 0;
 631         int copylen = 0;
 632         int depth;
 633         bool zerocopy = false;
 634         size_t linear;
 635
 636         if (q->flags & IFF_VNET_HDR) {
 637                 vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
 638
 639                 err = -EINVAL;
 640                 if (len < vnet_hdr_len)
 641                         goto err;
 642                 len -= vnet_hdr_len;
 643
 644                 err = -EFAULT;
 645                 if (!copy_from_iter_full(&vnet_hdr, sizeof(vnet_hdr), from))
 646                         goto err;
 647                 iov_iter_advance(from, vnet_hdr_len - sizeof(vnet_hdr));
 648                 if ((vnet_hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) &&
 649                      tap16_to_cpu(q, vnet_hdr.csum_start) +
 650                      tap16_to_cpu(q, vnet_hdr.csum_offset) + 2 >
 651                              tap16_to_cpu(q, vnet_hdr.hdr_len))
 652                         vnet_hdr.hdr_len = cpu_to_tap16(q,
 653                                  tap16_to_cpu(q, vnet_hdr.csum_start) +
 654                                  tap16_to_cpu(q, vnet_hdr.csum_offset) + 2);
 655                 err = -EINVAL;
 656                 if (tap16_to_cpu(q, vnet_hdr.hdr_len) > len)
 657                         goto err;
 658         }
 659
 660         err = -EINVAL;
 661         if (unlikely(len < ETH_HLEN))
 662                 goto err;
 663
 664         if (msg_control && sock_flag(&q->sk, SOCK_ZEROCOPY)) {
 665                 struct iov_iter i;
 666
 667                 copylen = vnet_hdr.hdr_len ?
 668                         tap16_to_cpu(q, vnet_hdr.hdr_len) : GOODCOPY_LEN;
 669                 if (copylen > good_linear)
 670                         copylen = good_linear;
 671                 else if (copylen < ETH_HLEN)
 672                         copylen = ETH_HLEN;
 673                 linear = copylen;
 674                 i = *from;
 675                 iov_iter_advance(&i, copylen);
 676                 if (iov_iter_npages(&i, INT_MAX) <= MAX_SKB_FRAGS)
 677                         zerocopy = true;
 678         }
 679
 680         if (!zerocopy) {
 681                 copylen = len;
 682                 linear = tap16_to_cpu(q, vnet_hdr.hdr_len);
 683                 if (linear > good_linear)
 684                         linear = good_linear;
 685                 else if (linear < ETH_HLEN)
 686                         linear = ETH_HLEN;
 687         }
 688
 689         skb = tap_alloc_skb(&q->sk, TAP_RESERVE, copylen,
 690                             linear, noblock, &err);
 691         if (!skb)
 692                 goto err;
 693
 694         if (zerocopy)
 695                 err = zerocopy_sg_from_iter(skb, from);
 696         else
 697                 err = skb_copy_datagram_from_iter(skb, 0, from, len);
 698
 699         if (err)
 700                 goto err_kfree;
 701
 702         skb_set_network_header(skb, ETH_HLEN);
 703         skb_reset_mac_header(skb);
 704         skb->protocol = eth_hdr(skb)->h_proto;
 705
 706         if (vnet_hdr_len) {
 707                 err = virtio_net_hdr_to_skb(skb, &vnet_hdr,
 708                                             tap_is_little_endian(q));
 709                 if (err)
 710                         goto err_kfree;
 711         }
 712
 713         skb_probe_transport_header(skb);
 714
 715         /* Move network header to the right position for VLAN tagged packets */
 716         if (eth_type_vlan(skb->protocol) &&
 717             __vlan_get_protocol(skb, skb->protocol, &depth) != 0)
 718                 skb_set_network_header(skb, depth);
 719
 720         rcu_read_lock();
 721         tap = rcu_dereference(q->tap);
 722         /* copy skb_ubuf_info for callback when skb has no error */
 723         if (zerocopy) {
 724                 skb_zcopy_init(skb, msg_control);
 725         } else if (msg_control) {
 726                 struct ubuf_info *uarg = msg_control;
 727                 uarg->callback(NULL, uarg, false);
 728         }
 729
 730         if (tap) {
 731                 skb->dev = tap->dev;
 732                 dev_queue_xmit(skb);
 733         } else {
 734                 kfree_skb(skb);
 735         }
 736         rcu_read_unlock();
 737
 738         return total_len;
 739
 740 err_kfree:
 741         kfree_skb(skb);
 742
 743 err:
 744         rcu_read_lock();
 745         tap = rcu_dereference(q->tap);
 746         if (tap && tap->count_tx_dropped)
 747                 tap->count_tx_dropped(tap);
 748         rcu_read_unlock();
 749
 750         return err;
 751 }
 752
 753 static ssize_t tap_write_iter(struct kiocb *iocb, struct iov_iter *from)
 754 {
 755         struct file *file = iocb->ki_filp;
 756         struct tap_queue *q = file->private_data;
 757
 758         return tap_get_user(q, NULL, from, file->f_flags & O_NONBLOCK);
 759 }
 760
 761 /* Put packet to the user space buffer */
 762 static ssize_t tap_put_user(struct tap_queue *q,
 763                             const struct sk_buff *skb,
 764                             struct iov_iter *iter)
 765 {
 766         int ret;
 767         int vnet_hdr_len = 0;
 768         int vlan_offset = 0;
 769         int total;
 770
 771         if (q->flags & IFF_VNET_HDR) {
 772                 int vlan_hlen = skb_vlan_tag_present(skb) ? VLAN_HLEN : 0;
 773                 struct virtio_net_hdr vnet_hdr;
 774
 775                 vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
 776                 if (iov_iter_count(iter) < vnet_hdr_len)
 777                         return -EINVAL;
 778
 779                 if (virtio_net_hdr_from_skb(skb, &vnet_hdr,
 780                                             tap_is_little_endian(q), true,
 781                                             vlan_hlen))
 782                         BUG();
 783
 784                 if (copy_to_iter(&vnet_hdr, sizeof(vnet_hdr), iter) !=
 785                     sizeof(vnet_hdr))
 786                         return -EFAULT;
 787
 788                 iov_iter_advance(iter, vnet_hdr_len - sizeof(vnet_hdr));
 789         }
 790         total = vnet_hdr_len;
 791         total += skb->len;
 792
 793         if (skb_vlan_tag_present(skb)) {
 794                 struct {
 795                         __be16 h_vlan_proto;
 796                         __be16 h_vlan_TCI;
 797                 } veth;
 798                 veth.h_vlan_proto = skb->vlan_proto;
 799                 veth.h_vlan_TCI = htons(skb_vlan_tag_get(skb));
 800
 801                 vlan_offset = offsetof(struct vlan_ethhdr, h_vlan_proto);
 802                 total += VLAN_HLEN;
 803
 804                 ret = skb_copy_datagram_iter(skb, 0, iter, vlan_offset);
 805                 if (ret || !iov_iter_count(iter))
 806                         goto done;
 807
 808                 ret = copy_to_iter(&veth, sizeof(veth), iter);
 809                 if (ret != sizeof(veth) || !iov_iter_count(iter))
 810                         goto done;
 811         }
 812
 813         ret = skb_copy_datagram_iter(skb, vlan_offset, iter,
 814                                      skb->len - vlan_offset);
 815
 816 done:
 817         return ret ? ret : total;
 818 }
 819
 820 static ssize_t tap_do_read(struct tap_queue *q,
 821                            struct iov_iter *to,
 822                            int noblock, struct sk_buff *skb)
 823 {
 824         DEFINE_WAIT(wait);
 825         ssize_t ret = 0;
 826
 827         if (!iov_iter_count(to)) {
 828                 kfree_skb(skb);
 829                 return 0;
 830         }
 831
 832         if (skb)
 833                 goto put;
 834
 835         while (1) {
 836                 if (!noblock)
 837                         prepare_to_wait(sk_sleep(&q->sk), &wait,
 838                                         TASK_INTERRUPTIBLE);
 839
 840                 /* Read frames from the queue */
 841                 skb = ptr_ring_consume(&q->ring);
 842                 if (skb)
 843                         break;
 844                 if (noblock) {
 845                         ret = -EAGAIN;
 846                         break;
 847                 }
 848                 if (signal_pending(current)) {
 849                         ret = -ERESTARTSYS;
 850                         break;
 851                 }
 852                 /* Nothing to read, let's sleep */
 853                 schedule();
 854         }
 855         if (!noblock)
 856                 finish_wait(sk_sleep(&q->sk), &wait);
 857
 858 put:
 859         if (skb) {
 860                 ret = tap_put_user(q, skb, to);
 861                 if (unlikely(ret < 0))
 862                         kfree_skb(skb);
 863                 else
 864                         consume_skb(skb);
 865         }
 866         return ret;
 867 }
 868
 869 static ssize_t tap_read_iter(struct kiocb *iocb, struct iov_iter *to)
 870 {
 871         struct file *file = iocb->ki_filp;
 872         struct tap_queue *q = file->private_data;
 873         ssize_t len = iov_iter_count(to), ret;
 874
 875         ret = tap_do_read(q, to, file->f_flags & O_NONBLOCK, NULL);
 876         ret = min_t(ssize_t, ret, len);
 877         if (ret > 0)
 878                 iocb->ki_pos = ret;
 879         return ret;
 880 }
 881
 882 static struct tap_dev *tap_get_tap_dev(struct tap_queue *q)
 883 {
 884         struct tap_dev *tap;
 885
 886         ASSERT_RTNL();
 887         tap = rtnl_dereference(q->tap);
 888         if (tap)
 889                 dev_hold(tap->dev);
 890
 891         return tap;
 892 }
 893
 894 static void tap_put_tap_dev(struct tap_dev *tap)
 895 {
 896         dev_put(tap->dev);
 897 }
 898
 899 static int tap_ioctl_set_queue(struct file *file, unsigned int flags)
 900 {
 901         struct tap_queue *q = file->private_data;
 902         struct tap_dev *tap;
 903         int ret;
 904
 905         tap = tap_get_tap_dev(q);
 906         if (!tap)
 907                 return -EINVAL;
 908
 909         if (flags & IFF_ATTACH_QUEUE)
 910                 ret = tap_enable_queue(tap, file, q);
 911         else if (flags & IFF_DETACH_QUEUE)
 912                 ret = tap_disable_queue(q);
 913         else
 914                 ret = -EINVAL;
 915
 916         tap_put_tap_dev(tap);
 917         return ret;
 918 }
 919
 920 static int set_offload(struct tap_queue *q, unsigned long arg)
 921 {
 922         struct tap_dev *tap;
 923         netdev_features_t features;
 924         netdev_features_t feature_mask = 0;
 925
 926         tap = rtnl_dereference(q->tap);
 927         if (!tap)
 928                 return -ENOLINK;
 929
 930         features = tap->dev->features;
 931
 932         if (arg & TUN_F_CSUM) {
 933                 feature_mask = NETIF_F_HW_CSUM;
 934
 935                 if (arg & (TUN_F_TSO4 | TUN_F_TSO6)) {
 936                         if (arg & TUN_F_TSO_ECN)
 937                                 feature_mask |= NETIF_F_TSO_ECN;
 938                         if (arg & TUN_F_TSO4)
 939                                 feature_mask |= NETIF_F_TSO;
 940                         if (arg & TUN_F_TSO6)
 941                                 feature_mask |= NETIF_F_TSO6;
 942                 }
 943         }
 944
 945         /* tun/tap driver inverts the usage for TSO offloads, where
 946          * setting the TSO bit means that the userspace wants to
 947          * accept TSO frames and turning it off means that user space
 948          * does not support TSO.
 949          * For tap, we have to invert it to mean the same thing.
 950          * When user space turns off TSO, we turn off GSO/LRO so that
 951          * user-space will not receive TSO frames.
 952          */
 953         if (feature_mask & (NETIF_F_TSO | NETIF_F_TSO6))
 954                 features |= RX_OFFLOADS;
 955         else
 956                 features &= ~RX_OFFLOADS;
 957
 958         /* tap_features are the same as features on tun/tap and
 959          * reflect user expectations.
 960          */
 961         tap->tap_features = feature_mask;
 962         if (tap->update_features)
 963                 tap->update_features(tap, features);
 964
 965         return 0;
 966 }
 967
 968 /*
 969  * provide compatibility with generic tun/tap interface
 970  */
 971 static long tap_ioctl(struct file *file, unsigned int cmd,
 972                       unsigned long arg)
 973 {
 974         struct tap_queue *q = file->private_data;
 975         struct tap_dev *tap;
 976         void __user *argp = (void __user *)arg;
 977         struct ifreq __user *ifr = argp;
 978         unsigned int __user *up = argp;
 979         unsigned short u;
 980         int __user *sp = argp;
 981         struct sockaddr sa;
 982         int s;
 983         int ret;
 984
 985         switch (cmd) {
 986         case TUNSETIFF:
 987                 /* ignore the name, just look at flags */
 988                 if (get_user(u, &ifr->ifr_flags))
 989                         return -EFAULT;
 990
 991                 ret = 0;
 992                 if ((u & ~TAP_IFFEATURES) != (IFF_NO_PI | IFF_TAP))
 993                         ret = -EINVAL;
 994                 else
 995                         q->flags = (q->flags & ~TAP_IFFEATURES) | u;
 996
 997                 return ret;
 998
 999         case TUNGETIFF:
1000                 rtnl_lock();
1001                 tap = tap_get_tap_dev(q);
1002                 if (!tap) {
1003                         rtnl_unlock();
1004                         return -ENOLINK;
1005                 }
1006
1007                 ret = 0;
1008                 u = q->flags;
1009                 if (copy_to_user(&ifr->ifr_name, tap->dev->name, IFNAMSIZ) ||
1010                     put_user(u, &ifr->ifr_flags))
1011                         ret = -EFAULT;
1012                 tap_put_tap_dev(tap);
1013                 rtnl_unlock();
1014                 return ret;
1015
1016         case TUNSETQUEUE:
1017                 if (get_user(u, &ifr->ifr_flags))
1018                         return -EFAULT;
1019                 rtnl_lock();
1020                 ret = tap_ioctl_set_queue(file, u);
1021                 rtnl_unlock();
1022                 return ret;
1023
1024         case TUNGETFEATURES:
1025                 if (put_user(IFF_TAP | IFF_NO_PI | TAP_IFFEATURES, up))
1026                         return -EFAULT;
1027                 return 0;
1028
1029         case TUNSETSNDBUF:
1030                 if (get_user(s, sp))
1031                         return -EFAULT;
1032                 if (s <= 0)
1033                         return -EINVAL;
1034
1035                 q->sk.sk_sndbuf = s;
1036                 return 0;
1037
1038         case TUNGETVNETHDRSZ:
1039                 s = q->vnet_hdr_sz;
1040                 if (put_user(s, sp))
1041                         return -EFAULT;
1042                 return 0;
1043
1044         case TUNSETVNETHDRSZ:
1045                 if (get_user(s, sp))
1046                         return -EFAULT;
1047                 if (s < (int)sizeof(struct virtio_net_hdr))
1048                         return -EINVAL;
1049
1050                 q->vnet_hdr_sz = s;
1051                 return 0;
1052
1053         case TUNGETVNETLE:
1054                 s = !!(q->flags & TAP_VNET_LE);
1055                 if (put_user(s, sp))
1056                         return -EFAULT;
1057                 return 0;
1058
1059         case TUNSETVNETLE:
1060                 if (get_user(s, sp))
1061                         return -EFAULT;
1062                 if (s)
1063                         q->flags |= TAP_VNET_LE;
1064                 else
1065                         q->flags &= ~TAP_VNET_LE;
1066                 return 0;
1067
1068         case TUNGETVNETBE:
1069                 return tap_get_vnet_be(q, sp);
1070
1071         case TUNSETVNETBE:
1072                 return tap_set_vnet_be(q, sp);
1073
1074         case TUNSETOFFLOAD:
1075                 /* let the user check for future flags */
1076                 if (arg & ~(TUN_F_CSUM | TUN_F_TSO4 | TUN_F_TSO6 |
1077                             TUN_F_TSO_ECN | TUN_F_UFO))
1078                         return -EINVAL;
1079
1080                 rtnl_lock();
1081                 ret = set_offload(q, arg);
1082                 rtnl_unlock();
1083                 return ret;
1084
1085         case SIOCGIFHWADDR:
1086                 rtnl_lock();
1087                 tap = tap_get_tap_dev(q);
1088                 if (!tap) {
1089                         rtnl_unlock();
1090                         return -ENOLINK;
1091                 }
1092                 ret = 0;
1093                 dev_get_mac_address(&sa, dev_net(tap->dev), tap->dev->name);
1094                 if (copy_to_user(&ifr->ifr_name, tap->dev->name, IFNAMSIZ) ||
1095                     copy_to_user(&ifr->ifr_hwaddr, &sa, sizeof(sa)))
1096                         ret = -EFAULT;
1097                 tap_put_tap_dev(tap);
1098                 rtnl_unlock();
1099                 return ret;
1100
1101         case SIOCSIFHWADDR:
1102                 if (copy_from_user(&sa, &ifr->ifr_hwaddr, sizeof(sa)))
1103                         return -EFAULT;
1104                 rtnl_lock();
1105                 tap = tap_get_tap_dev(q);
1106                 if (!tap) {
1107                         rtnl_unlock();
1108                         return -ENOLINK;
1109                 }
1110                 ret = dev_set_mac_address_user(tap->dev, &sa, NULL);
1111                 tap_put_tap_dev(tap);
1112                 rtnl_unlock();
1113                 return ret;
1114
1115         default:
1116                 return -EINVAL;
1117         }
1118 }
1119
1120 static const struct file_operations tap_fops = {
1121         .owner          = THIS_MODULE,
1122         .open           = tap_open,
1123         .release        = tap_release,
1124         .read_iter      = tap_read_iter,
1125         .write_iter     = tap_write_iter,
1126         .poll           = tap_poll,
1127         .llseek         = no_llseek,
1128         .unlocked_ioctl = tap_ioctl,
1129         .compat_ioctl   = compat_ptr_ioctl,
1130 };
1131
1132 static int tap_get_user_xdp(struct tap_queue *q, struct xdp_buff *xdp)
1133 {
1134         struct tun_xdp_hdr *hdr = xdp->data_hard_start;
1135         struct virtio_net_hdr *gso = &hdr->gso;
1136         int buflen = hdr->buflen;
1137         int vnet_hdr_len = 0;
1138         struct tap_dev *tap;
1139         struct sk_buff *skb;
1140         int err, depth;
1141
1142         if (q->flags & IFF_VNET_HDR)
1143                 vnet_hdr_len = READ_ONCE(q->vnet_hdr_sz);
1144
1145         skb = build_skb(xdp->data_hard_start, buflen);
1146         if (!skb) {
1147                 err = -ENOMEM;
1148                 goto err;
1149         }
1150
1151         skb_reserve(skb, xdp->data - xdp->data_hard_start);
1152         skb_put(skb, xdp->data_end - xdp->data);
1153
1154         skb_set_network_header(skb, ETH_HLEN);
1155         skb_reset_mac_header(skb);
1156         skb->protocol = eth_hdr(skb)->h_proto;
1157
1158         if (vnet_hdr_len) {
1159                 err = virtio_net_hdr_to_skb(skb, gso, tap_is_little_endian(q));
1160                 if (err)
1161                         goto err_kfree;
1162         }
1163
1164         /* Move network header to the right position for VLAN tagged packets */
1165         if (eth_type_vlan(skb->protocol) &&
1166             __vlan_get_protocol(skb, skb->protocol, &depth) != 0)
1167                 skb_set_network_header(skb, depth);
1168
1169         rcu_read_lock();
1170         tap = rcu_dereference(q->tap);
1171         if (tap) {
1172                 skb->dev = tap->dev;
1173                 skb_probe_transport_header(skb);
1174                 dev_queue_xmit(skb);
1175         } else {
1176                 kfree_skb(skb);
1177         }
1178         rcu_read_unlock();
1179
1180         return 0;
1181
1182 err_kfree:
1183         kfree_skb(skb);
1184 err:
1185         rcu_read_lock();
1186         tap = rcu_dereference(q->tap);
1187         if (tap && tap->count_tx_dropped)
1188                 tap->count_tx_dropped(tap);
1189         rcu_read_unlock();
1190         return err;
1191 }
1192
1193 static int tap_sendmsg(struct socket *sock, struct msghdr *m,
1194                        size_t total_len)
1195 {
1196         struct tap_queue *q = container_of(sock, struct tap_queue, sock);
1197         struct tun_msg_ctl *ctl = m->msg_control;
1198         struct xdp_buff *xdp;
1199         int i;
1200
1201         if (ctl && (ctl->type == TUN_MSG_PTR)) {
1202                 for (i = 0; i < ctl->num; i++) {
1203                         xdp = &((struct xdp_buff *)ctl->ptr)[i];
1204                         tap_get_user_xdp(q, xdp);
1205                 }
1206                 return 0;
1207         }
1208
1209         return tap_get_user(q, ctl ? ctl->ptr : NULL, &m->msg_iter,
1210                             m->msg_flags & MSG_DONTWAIT);
1211 }
1212
1213 static int tap_recvmsg(struct socket *sock, struct msghdr *m,
1214                        size_t total_len, int flags)
1215 {
1216         struct tap_queue *q = container_of(sock, struct tap_queue, sock);
1217         struct sk_buff *skb = m->msg_control;
1218         int ret;
1219         if (flags & ~(MSG_DONTWAIT|MSG_TRUNC)) {
1220                 kfree_skb(skb);
1221                 return -EINVAL;
1222         }
1223         ret = tap_do_read(q, &m->msg_iter, flags & MSG_DONTWAIT, skb);
1224         if (ret > total_len) {
1225                 m->msg_flags |= MSG_TRUNC;
1226                 ret = flags & MSG_TRUNC ? ret : total_len;
1227         }
1228         return ret;
1229 }
1230
1231 static int tap_peek_len(struct socket *sock)
1232 {
1233         struct tap_queue *q = container_of(sock, struct tap_queue,
1234                                                sock);
1235         return PTR_RING_PEEK_CALL(&q->ring, __skb_array_len_with_tag);
1236 }
1237
1238 /* Ops structure to mimic raw sockets with tun */
1239 static const struct proto_ops tap_socket_ops = {
1240         .sendmsg = tap_sendmsg,
1241         .recvmsg = tap_recvmsg,
1242         .peek_len = tap_peek_len,
1243 };
1244
1245 /* Get an underlying socket object from tun file.  Returns error unless file is
1246  * attached to a device.  The returned object works like a packet socket, it
1247  * can be used for sock_sendmsg/sock_recvmsg.  The caller is responsible for
1248  * holding a reference to the file for as long as the socket is in use. */
1249 struct socket *tap_get_socket(struct file *file)
1250 {
1251         struct tap_queue *q;
1252         if (file->f_op != &tap_fops)
1253                 return ERR_PTR(-EINVAL);
1254         q = file->private_data;
1255         if (!q)
1256                 return ERR_PTR(-EBADFD);
1257         return &q->sock;
1258 }
1259 EXPORT_SYMBOL_GPL(tap_get_socket);
1260
1261 struct ptr_ring *tap_get_ptr_ring(struct file *file)
1262 {
1263         struct tap_queue *q;
1264
1265         if (file->f_op != &tap_fops)
1266                 return ERR_PTR(-EINVAL);
1267         q = file->private_data;
1268         if (!q)
1269                 return ERR_PTR(-EBADFD);
1270         return &q->ring;
1271 }
1272 EXPORT_SYMBOL_GPL(tap_get_ptr_ring);
1273
1274 int tap_queue_resize(struct tap_dev *tap)
1275 {
1276         struct net_device *dev = tap->dev;
1277         struct tap_queue *q;
1278         struct ptr_ring **rings;
1279         int n = tap->numqueues;
1280         int ret, i = 0;
1281
1282         rings = kmalloc_array(n, sizeof(*rings), GFP_KERNEL);
1283         if (!rings)
1284                 return -ENOMEM;
1285
1286         list_for_each_entry(q, &tap->queue_list, next)
1287                 rings[i++] = &q->ring;
1288
1289         ret = ptr_ring_resize_multiple(rings, n,
1290                                        dev->tx_queue_len, GFP_KERNEL,
1291                                        __skb_array_destroy_skb);
1292
1293         kfree(rings);
1294         return ret;
1295 }
1296 EXPORT_SYMBOL_GPL(tap_queue_resize);
1297
1298 static int tap_list_add(dev_t major, const char *device_name)
1299 {
1300         struct major_info *tap_major;
1301
1302         tap_major = kzalloc(sizeof(*tap_major), GFP_ATOMIC);
1303         if (!tap_major)
1304                 return -ENOMEM;
1305
1306         tap_major->major = MAJOR(major);
1307
1308         idr_init(&tap_major->minor_idr);
1309         spin_lock_init(&tap_major->minor_lock);
1310
1311         tap_major->device_name = device_name;
1312
1313         list_add_tail_rcu(&tap_major->next, &major_list);
1314         return 0;
1315 }
1316
1317 int tap_create_cdev(struct cdev *tap_cdev, dev_t *tap_major,
1318                     const char *device_name, struct module *module)
1319 {
1320         int err;
1321
1322         err = alloc_chrdev_region(tap_major, 0, TAP_NUM_DEVS, device_name);
1323         if (err)
1324                 goto out1;
1325
1326         cdev_init(tap_cdev, &tap_fops);
1327         tap_cdev->owner = module;
1328         err = cdev_add(tap_cdev, *tap_major, TAP_NUM_DEVS);
1329         if (err)
1330                 goto out2;
1331
1332         err =  tap_list_add(*tap_major, device_name);
1333         if (err)
1334                 goto out3;
1335
1336         return 0;
1337
1338 out3:
1339         cdev_del(tap_cdev);
1340 out2:
1341         unregister_chrdev_region(*tap_major, TAP_NUM_DEVS);
1342 out1:
1343         return err;
1344 }
1345 EXPORT_SYMBOL_GPL(tap_create_cdev);
1346
1347 void tap_destroy_cdev(dev_t major, struct cdev *tap_cdev)
1348 {
1349         struct major_info *tap_major, *tmp;
1350
1351         cdev_del(tap_cdev);
1352         unregister_chrdev_region(major, TAP_NUM_DEVS);
1353         list_for_each_entry_safe(tap_major, tmp, &major_list, next) {
1354                 if (tap_major->major == MAJOR(major)) {
1355                         idr_destroy(&tap_major->minor_idr);
1356                         list_del_rcu(&tap_major->next);
1357                         kfree_rcu(tap_major, rcu);
1358                 }
1359         }
1360 }
1361 EXPORT_SYMBOL_GPL(tap_destroy_cdev);
1362
1363 MODULE_AUTHOR("Arnd Bergmann <arnd@arndb.de>");
1364 MODULE_AUTHOR("Sainath Grandhi <sainath.grandhi@intel.com>");
1365 MODULE_LICENSE("GPL");