net/xdp/xsk.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* XDP sockets
   3  *
   4  * AF_XDP sockets allows a channel between XDP programs and userspace
   5  * applications.
   6  * Copyright(c) 2018 Intel Corporation.
   7  *
   8  * Author(s): Björn Töpel <bjorn.topel@intel.com>
   9  *            Magnus Karlsson <magnus.karlsson@intel.com>
  10  */
  11
  12 #define pr_fmt(fmt) "AF_XDP: %s: " fmt, __func__
  13
  14 #include <linux/if_xdp.h>
  15 #include <linux/init.h>
  16 #include <linux/sched/mm.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/task.h>
  19 #include <linux/socket.h>
  20 #include <linux/file.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/net.h>
  23 #include <linux/netdevice.h>
  24 #include <linux/rculist.h>
  25 #include <net/xdp_sock.h>
  26 #include <net/xdp.h>
  27
  28 #include "xsk_queue.h"
  29 #include "xdp_umem.h"
  30
  31 #define TX_BATCH_SIZE 16
  32
  33 static struct xdp_sock *xdp_sk(struct sock *sk)
  34 {
  35         return (struct xdp_sock *)sk;
  36 }
  37
  38 bool xsk_is_setup_for_bpf_map(struct xdp_sock *xs)
  39 {
  40         return READ_ONCE(xs->rx) &&  READ_ONCE(xs->umem) &&
  41                 READ_ONCE(xs->umem->fq);
  42 }
  43
  44 u64 *xsk_umem_peek_addr(struct xdp_umem *umem, u64 *addr)
  45 {
  46         return xskq_peek_addr(umem->fq, addr);
  47 }
  48 EXPORT_SYMBOL(xsk_umem_peek_addr);
  49
  50 void xsk_umem_discard_addr(struct xdp_umem *umem)
  51 {
  52         xskq_discard_addr(umem->fq);
  53 }
  54 EXPORT_SYMBOL(xsk_umem_discard_addr);
  55
  56 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
  57 {
  58         void *buffer;
  59         u64 addr;
  60         int err;
  61
  62         if (!xskq_peek_addr(xs->umem->fq, &addr) ||
  63             len > xs->umem->chunk_size_nohr) {
  64                 xs->rx_dropped++;
  65                 return -ENOSPC;
  66         }
  67
  68         addr += xs->umem->headroom;
  69
  70         buffer = xdp_umem_get_data(xs->umem, addr);
  71         memcpy(buffer, xdp->data, len);
  72         err = xskq_produce_batch_desc(xs->rx, addr, len);
  73         if (!err) {
  74                 xskq_discard_addr(xs->umem->fq);
  75                 xdp_return_buff(xdp);
  76                 return 0;
  77         }
  78
  79         xs->rx_dropped++;
  80         return err;
  81 }
  82
  83 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
  84 {
  85         int err = xskq_produce_batch_desc(xs->rx, (u64)xdp->handle, len);
  86
  87         if (err) {
  88                 xdp_return_buff(xdp);
  89                 xs->rx_dropped++;
  90         }
  91
  92         return err;
  93 }
  94
  95 int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
  96 {
  97         u32 len;
  98
  99         if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
 100                 return -EINVAL;
 101
 102         len = xdp->data_end - xdp->data;
 103
 104         return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ?
 105                 __xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len);
 106 }
 107
 108 void xsk_flush(struct xdp_sock *xs)
 109 {
 110         xskq_produce_flush_desc(xs->rx);
 111         xs->sk.sk_data_ready(&xs->sk);
 112 }
 113
 114 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 115 {
 116         u32 len = xdp->data_end - xdp->data;
 117         void *buffer;
 118         u64 addr;
 119         int err;
 120
 121         if (!xskq_peek_addr(xs->umem->fq, &addr) ||
 122             len > xs->umem->chunk_size_nohr) {
 123                 xs->rx_dropped++;
 124                 return -ENOSPC;
 125         }
 126
 127         addr += xs->umem->headroom;
 128
 129         buffer = xdp_umem_get_data(xs->umem, addr);
 130         memcpy(buffer, xdp->data, len);
 131         err = xskq_produce_batch_desc(xs->rx, addr, len);
 132         if (!err) {
 133                 xskq_discard_addr(xs->umem->fq);
 134                 xsk_flush(xs);
 135                 return 0;
 136         }
 137
 138         xs->rx_dropped++;
 139         return err;
 140 }
 141
 142 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
 143 {
 144         xskq_produce_flush_addr_n(umem->cq, nb_entries);
 145 }
 146 EXPORT_SYMBOL(xsk_umem_complete_tx);
 147
 148 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
 149 {
 150         struct xdp_sock *xs;
 151
 152         rcu_read_lock();
 153         list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
 154                 xs->sk.sk_write_space(&xs->sk);
 155         }
 156         rcu_read_unlock();
 157 }
 158 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
 159
 160 bool xsk_umem_consume_tx(struct xdp_umem *umem, dma_addr_t *dma, u32 *len)
 161 {
 162         struct xdp_desc desc;
 163         struct xdp_sock *xs;
 164
 165         rcu_read_lock();
 166         list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
 167                 if (!xskq_peek_desc(xs->tx, &desc))
 168                         continue;
 169
 170                 if (xskq_produce_addr_lazy(umem->cq, desc.addr))
 171                         goto out;
 172
 173                 *dma = xdp_umem_get_dma(umem, desc.addr);
 174                 *len = desc.len;
 175
 176                 xskq_discard_desc(xs->tx);
 177                 rcu_read_unlock();
 178                 return true;
 179         }
 180
 181 out:
 182         rcu_read_unlock();
 183         return false;
 184 }
 185 EXPORT_SYMBOL(xsk_umem_consume_tx);
 186
 187 static int xsk_zc_xmit(struct sock *sk)
 188 {
 189         struct xdp_sock *xs = xdp_sk(sk);
 190         struct net_device *dev = xs->dev;
 191
 192         return dev->netdev_ops->ndo_xsk_async_xmit(dev, xs->queue_id);
 193 }
 194
 195 static void xsk_destruct_skb(struct sk_buff *skb)
 196 {
 197         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 198         struct xdp_sock *xs = xdp_sk(skb->sk);
 199
 200         WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr));
 201
 202         sock_wfree(skb);
 203 }
 204
 205 static int xsk_generic_xmit(struct sock *sk, struct msghdr *m,
 206                             size_t total_len)
 207 {
 208         u32 max_batch = TX_BATCH_SIZE;
 209         struct xdp_sock *xs = xdp_sk(sk);
 210         bool sent_frame = false;
 211         struct xdp_desc desc;
 212         struct sk_buff *skb;
 213         int err = 0;
 214
 215         if (unlikely(!xs->tx))
 216                 return -ENOBUFS;
 217
 218         mutex_lock(&xs->mutex);
 219
 220         while (xskq_peek_desc(xs->tx, &desc)) {
 221                 char *buffer;
 222                 u64 addr;
 223                 u32 len;
 224
 225                 if (max_batch-- == 0) {
 226                         err = -EAGAIN;
 227                         goto out;
 228                 }
 229
 230                 if (xskq_reserve_addr(xs->umem->cq)) {
 231                         err = -EAGAIN;
 232                         goto out;
 233                 }
 234
 235                 len = desc.len;
 236                 if (unlikely(len > xs->dev->mtu)) {
 237                         err = -EMSGSIZE;
 238                         goto out;
 239                 }
 240
 241                 if (xs->queue_id >= xs->dev->real_num_tx_queues) {
 242                         err = -ENXIO;
 243                         goto out;
 244                 }
 245
 246                 skb = sock_alloc_send_skb(sk, len, 1, &err);
 247                 if (unlikely(!skb)) {
 248                         err = -EAGAIN;
 249                         goto out;
 250                 }
 251
 252                 skb_put(skb, len);
 253                 addr = desc.addr;
 254                 buffer = xdp_umem_get_data(xs->umem, addr);
 255                 err = skb_store_bits(skb, 0, buffer, len);
 256                 if (unlikely(err)) {
 257                         kfree_skb(skb);
 258                         goto out;
 259                 }
 260
 261                 skb->dev = xs->dev;
 262                 skb->priority = sk->sk_priority;
 263                 skb->mark = sk->sk_mark;
 264                 skb_shinfo(skb)->destructor_arg = (void *)(long)addr;
 265                 skb->destructor = xsk_destruct_skb;
 266
 267                 err = dev_direct_xmit(skb, xs->queue_id);
 268                 /* Ignore NET_XMIT_CN as packet might have been sent */
 269                 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
 270                         err = -EAGAIN;
 271                         /* SKB consumed by dev_direct_xmit() */
 272                         goto out;
 273                 }
 274
 275                 sent_frame = true;
 276                 xskq_discard_desc(xs->tx);
 277         }
 278
 279 out:
 280         if (sent_frame)
 281                 sk->sk_write_space(sk);
 282
 283         mutex_unlock(&xs->mutex);
 284         return err;
 285 }
 286
 287 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 288 {
 289         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 290         struct sock *sk = sock->sk;
 291         struct xdp_sock *xs = xdp_sk(sk);
 292
 293         if (unlikely(!xs->dev))
 294                 return -ENXIO;
 295         if (unlikely(!(xs->dev->flags & IFF_UP)))
 296                 return -ENETDOWN;
 297         if (need_wait)
 298                 return -EOPNOTSUPP;
 299
 300         return (xs->zc) ? xsk_zc_xmit(sk) : xsk_generic_xmit(sk, m, total_len);
 301 }
 302
 303 static unsigned int xsk_poll(struct file *file, struct socket *sock,
 304                              struct poll_table_struct *wait)
 305 {
 306         unsigned int mask = datagram_poll(file, sock, wait);
 307         struct sock *sk = sock->sk;
 308         struct xdp_sock *xs = xdp_sk(sk);
 309
 310         if (xs->rx && !xskq_empty_desc(xs->rx))
 311                 mask |= POLLIN | POLLRDNORM;
 312         if (xs->tx && !xskq_full_desc(xs->tx))
 313                 mask |= POLLOUT | POLLWRNORM;
 314
 315         return mask;
 316 }
 317
 318 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 319                           bool umem_queue)
 320 {
 321         struct xsk_queue *q;
 322
 323         if (entries == 0 || *queue || !is_power_of_2(entries))
 324                 return -EINVAL;
 325
 326         q = xskq_create(entries, umem_queue);
 327         if (!q)
 328                 return -ENOMEM;
 329
 330         /* Make sure queue is ready before it can be seen by others */
 331         smp_wmb();
 332         *queue = q;
 333         return 0;
 334 }
 335
 336 static int xsk_release(struct socket *sock)
 337 {
 338         struct sock *sk = sock->sk;
 339         struct xdp_sock *xs = xdp_sk(sk);
 340         struct net *net;
 341
 342         if (!sk)
 343                 return 0;
 344
 345         net = sock_net(sk);
 346
 347         local_bh_disable();
 348         sock_prot_inuse_add(net, sk->sk_prot, -1);
 349         local_bh_enable();
 350
 351         if (xs->dev) {
 352                 /* Wait for driver to stop using the xdp socket. */
 353                 synchronize_net();
 354                 dev_put(xs->dev);
 355                 xs->dev = NULL;
 356         }
 357
 358         sock_orphan(sk);
 359         sock->sk = NULL;
 360
 361         sk_refcnt_debug_release(sk);
 362         sock_put(sk);
 363
 364         return 0;
 365 }
 366
 367 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 368 {
 369         struct socket *sock;
 370         int err;
 371
 372         sock = sockfd_lookup(fd, &err);
 373         if (!sock)
 374                 return ERR_PTR(-ENOTSOCK);
 375
 376         if (sock->sk->sk_family != PF_XDP) {
 377                 sockfd_put(sock);
 378                 return ERR_PTR(-ENOPROTOOPT);
 379         }
 380
 381         return sock;
 382 }
 383
 384 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 385 {
 386         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 387         struct sock *sk = sock->sk;
 388         struct xdp_sock *xs = xdp_sk(sk);
 389         struct net_device *dev;
 390         u32 flags, qid;
 391         int err = 0;
 392
 393         if (addr_len < sizeof(struct sockaddr_xdp))
 394                 return -EINVAL;
 395         if (sxdp->sxdp_family != AF_XDP)
 396                 return -EINVAL;
 397
 398         mutex_lock(&xs->mutex);
 399         if (xs->dev) {
 400                 err = -EBUSY;
 401                 goto out_release;
 402         }
 403
 404         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 405         if (!dev) {
 406                 err = -ENODEV;
 407                 goto out_release;
 408         }
 409
 410         if (!xs->rx && !xs->tx) {
 411                 err = -EINVAL;
 412                 goto out_unlock;
 413         }
 414
 415         qid = sxdp->sxdp_queue_id;
 416
 417         if ((xs->rx && qid >= dev->real_num_rx_queues) ||
 418             (xs->tx && qid >= dev->real_num_tx_queues)) {
 419                 err = -EINVAL;
 420                 goto out_unlock;
 421         }
 422
 423         flags = sxdp->sxdp_flags;
 424
 425         if (flags & XDP_SHARED_UMEM) {
 426                 struct xdp_sock *umem_xs;
 427                 struct socket *sock;
 428
 429                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY)) {
 430                         /* Cannot specify flags for shared sockets. */
 431                         err = -EINVAL;
 432                         goto out_unlock;
 433                 }
 434
 435                 if (xs->umem) {
 436                         /* We have already our own. */
 437                         err = -EINVAL;
 438                         goto out_unlock;
 439                 }
 440
 441                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 442                 if (IS_ERR(sock)) {
 443                         err = PTR_ERR(sock);
 444                         goto out_unlock;
 445                 }
 446
 447                 umem_xs = xdp_sk(sock->sk);
 448                 if (!umem_xs->umem) {
 449                         /* No umem to inherit. */
 450                         err = -EBADF;
 451                         sockfd_put(sock);
 452                         goto out_unlock;
 453                 } else if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
 454                         err = -EINVAL;
 455                         sockfd_put(sock);
 456                         goto out_unlock;
 457                 }
 458
 459                 xdp_get_umem(umem_xs->umem);
 460                 xs->umem = umem_xs->umem;
 461                 sockfd_put(sock);
 462         } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
 463                 err = -EINVAL;
 464                 goto out_unlock;
 465         } else {
 466                 /* This xsk has its own umem. */
 467                 xskq_set_umem(xs->umem->fq, &xs->umem->props);
 468                 xskq_set_umem(xs->umem->cq, &xs->umem->props);
 469
 470                 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
 471                 if (err)
 472                         goto out_unlock;
 473         }
 474
 475         xs->dev = dev;
 476         xs->zc = xs->umem->zc;
 477         xs->queue_id = qid;
 478         xskq_set_umem(xs->rx, &xs->umem->props);
 479         xskq_set_umem(xs->tx, &xs->umem->props);
 480         xdp_add_sk_umem(xs->umem, xs);
 481
 482 out_unlock:
 483         if (err)
 484                 dev_put(dev);
 485 out_release:
 486         mutex_unlock(&xs->mutex);
 487         return err;
 488 }
 489
 490 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 491                           char __user *optval, unsigned int optlen)
 492 {
 493         struct sock *sk = sock->sk;
 494         struct xdp_sock *xs = xdp_sk(sk);
 495         int err;
 496
 497         if (level != SOL_XDP)
 498                 return -ENOPROTOOPT;
 499
 500         switch (optname) {
 501         case XDP_RX_RING:
 502         case XDP_TX_RING:
 503         {
 504                 struct xsk_queue **q;
 505                 int entries;
 506
 507                 if (optlen < sizeof(entries))
 508                         return -EINVAL;
 509                 if (copy_from_user(&entries, optval, sizeof(entries)))
 510                         return -EFAULT;
 511
 512                 mutex_lock(&xs->mutex);
 513                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
 514                 err = xsk_init_queue(entries, q, false);
 515                 mutex_unlock(&xs->mutex);
 516                 return err;
 517         }
 518         case XDP_UMEM_REG:
 519         {
 520                 struct xdp_umem_reg mr;
 521                 struct xdp_umem *umem;
 522
 523                 if (copy_from_user(&mr, optval, sizeof(mr)))
 524                         return -EFAULT;
 525
 526                 mutex_lock(&xs->mutex);
 527                 if (xs->umem) {
 528                         mutex_unlock(&xs->mutex);
 529                         return -EBUSY;
 530                 }
 531
 532                 umem = xdp_umem_create(&mr);
 533                 if (IS_ERR(umem)) {
 534                         mutex_unlock(&xs->mutex);
 535                         return PTR_ERR(umem);
 536                 }
 537
 538                 /* Make sure umem is ready before it can be seen by others */
 539                 smp_wmb();
 540                 xs->umem = umem;
 541                 mutex_unlock(&xs->mutex);
 542                 return 0;
 543         }
 544         case XDP_UMEM_FILL_RING:
 545         case XDP_UMEM_COMPLETION_RING:
 546         {
 547                 struct xsk_queue **q;
 548                 int entries;
 549
 550                 if (copy_from_user(&entries, optval, sizeof(entries)))
 551                         return -EFAULT;
 552
 553                 mutex_lock(&xs->mutex);
 554                 if (!xs->umem) {
 555                         mutex_unlock(&xs->mutex);
 556                         return -EINVAL;
 557                 }
 558
 559                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
 560                         &xs->umem->cq;
 561                 err = xsk_init_queue(entries, q, true);
 562                 mutex_unlock(&xs->mutex);
 563                 return err;
 564         }
 565         default:
 566                 break;
 567         }
 568
 569         return -ENOPROTOOPT;
 570 }
 571
 572 static int xsk_getsockopt(struct socket *sock, int level, int optname,
 573                           char __user *optval, int __user *optlen)
 574 {
 575         struct sock *sk = sock->sk;
 576         struct xdp_sock *xs = xdp_sk(sk);
 577         int len;
 578
 579         if (level != SOL_XDP)
 580                 return -ENOPROTOOPT;
 581
 582         if (get_user(len, optlen))
 583                 return -EFAULT;
 584         if (len < 0)
 585                 return -EINVAL;
 586
 587         switch (optname) {
 588         case XDP_STATISTICS:
 589         {
 590                 struct xdp_statistics stats;
 591
 592                 if (len < sizeof(stats))
 593                         return -EINVAL;
 594
 595                 mutex_lock(&xs->mutex);
 596                 stats.rx_dropped = xs->rx_dropped;
 597                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
 598                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
 599                 mutex_unlock(&xs->mutex);
 600
 601                 if (copy_to_user(optval, &stats, sizeof(stats)))
 602                         return -EFAULT;
 603                 if (put_user(sizeof(stats), optlen))
 604                         return -EFAULT;
 605
 606                 return 0;
 607         }
 608         case XDP_MMAP_OFFSETS:
 609         {
 610                 struct xdp_mmap_offsets off;
 611
 612                 if (len < sizeof(off))
 613                         return -EINVAL;
 614
 615                 off.rx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 616                 off.rx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 617                 off.rx.desc     = offsetof(struct xdp_rxtx_ring, desc);
 618                 off.tx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 619                 off.tx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 620                 off.tx.desc     = offsetof(struct xdp_rxtx_ring, desc);
 621
 622                 off.fr.producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 623                 off.fr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 624                 off.fr.desc     = offsetof(struct xdp_umem_ring, desc);
 625                 off.cr.producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 626                 off.cr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 627                 off.cr.desc     = offsetof(struct xdp_umem_ring, desc);
 628
 629                 len = sizeof(off);
 630                 if (copy_to_user(optval, &off, len))
 631                         return -EFAULT;
 632                 if (put_user(len, optlen))
 633                         return -EFAULT;
 634
 635                 return 0;
 636         }
 637         default:
 638                 break;
 639         }
 640
 641         return -EOPNOTSUPP;
 642 }
 643
 644 static int xsk_mmap(struct file *file, struct socket *sock,
 645                     struct vm_area_struct *vma)
 646 {
 647         unsigned long offset = vma->vm_pgoff << PAGE_SHIFT;
 648         unsigned long size = vma->vm_end - vma->vm_start;
 649         struct xdp_sock *xs = xdp_sk(sock->sk);
 650         struct xsk_queue *q = NULL;
 651         struct xdp_umem *umem;
 652         unsigned long pfn;
 653         struct page *qpg;
 654
 655         if (offset == XDP_PGOFF_RX_RING) {
 656                 q = READ_ONCE(xs->rx);
 657         } else if (offset == XDP_PGOFF_TX_RING) {
 658                 q = READ_ONCE(xs->tx);
 659         } else {
 660                 umem = READ_ONCE(xs->umem);
 661                 if (!umem)
 662                         return -EINVAL;
 663
 664                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
 665                         q = READ_ONCE(umem->fq);
 666                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
 667                         q = READ_ONCE(umem->cq);
 668         }
 669
 670         if (!q)
 671                 return -EINVAL;
 672
 673         qpg = virt_to_head_page(q->ring);
 674         if (size > (PAGE_SIZE << compound_order(qpg)))
 675                 return -EINVAL;
 676
 677         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
 678         return remap_pfn_range(vma, vma->vm_start, pfn,
 679                                size, vma->vm_page_prot);
 680 }
 681
 682 static struct proto xsk_proto = {
 683         .name =         "XDP",
 684         .owner =        THIS_MODULE,
 685         .obj_size =     sizeof(struct xdp_sock),
 686 };
 687
 688 static const struct proto_ops xsk_proto_ops = {
 689         .family         = PF_XDP,
 690         .owner          = THIS_MODULE,
 691         .release        = xsk_release,
 692         .bind           = xsk_bind,
 693         .connect        = sock_no_connect,
 694         .socketpair     = sock_no_socketpair,
 695         .accept         = sock_no_accept,
 696         .getname        = sock_no_getname,
 697         .poll           = xsk_poll,
 698         .ioctl          = sock_no_ioctl,
 699         .listen         = sock_no_listen,
 700         .shutdown       = sock_no_shutdown,
 701         .setsockopt     = xsk_setsockopt,
 702         .getsockopt     = xsk_getsockopt,
 703         .sendmsg        = xsk_sendmsg,
 704         .recvmsg        = sock_no_recvmsg,
 705         .mmap           = xsk_mmap,
 706         .sendpage       = sock_no_sendpage,
 707 };
 708
 709 static void xsk_destruct(struct sock *sk)
 710 {
 711         struct xdp_sock *xs = xdp_sk(sk);
 712
 713         if (!sock_flag(sk, SOCK_DEAD))
 714                 return;
 715
 716         xskq_destroy(xs->rx);
 717         xskq_destroy(xs->tx);
 718         xdp_del_sk_umem(xs->umem, xs);
 719         xdp_put_umem(xs->umem);
 720
 721         sk_refcnt_debug_dec(sk);
 722 }
 723
 724 static int xsk_create(struct net *net, struct socket *sock, int protocol,
 725                       int kern)
 726 {
 727         struct sock *sk;
 728         struct xdp_sock *xs;
 729
 730         if (!ns_capable(net->user_ns, CAP_NET_RAW))
 731                 return -EPERM;
 732         if (sock->type != SOCK_RAW)
 733                 return -ESOCKTNOSUPPORT;
 734
 735         if (protocol)
 736                 return -EPROTONOSUPPORT;
 737
 738         sock->state = SS_UNCONNECTED;
 739
 740         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
 741         if (!sk)
 742                 return -ENOBUFS;
 743
 744         sock->ops = &xsk_proto_ops;
 745
 746         sock_init_data(sock, sk);
 747
 748         sk->sk_family = PF_XDP;
 749
 750         sk->sk_destruct = xsk_destruct;
 751         sk_refcnt_debug_inc(sk);
 752
 753         xs = xdp_sk(sk);
 754         mutex_init(&xs->mutex);
 755
 756         local_bh_disable();
 757         sock_prot_inuse_add(net, &xsk_proto, 1);
 758         local_bh_enable();
 759
 760         return 0;
 761 }
 762
 763 static const struct net_proto_family xsk_family_ops = {
 764         .family = PF_XDP,
 765         .create = xsk_create,
 766         .owner  = THIS_MODULE,
 767 };
 768
 769 static int __init xsk_init(void)
 770 {
 771         int err;
 772
 773         err = proto_register(&xsk_proto, 0 /* no slab */);
 774         if (err)
 775                 goto out;
 776
 777         err = sock_register(&xsk_family_ops);
 778         if (err)
 779                 goto out_proto;
 780
 781         return 0;
 782
 783 out_proto:
 784         proto_unregister(&xsk_proto);
 785 out:
 786         return err;
 787 }
 788
 789 fs_initcall(xsk_init);