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 <[email protected]>
   9  *            Magnus Karlsson <[email protected]>
  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 (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
 122                 return -EINVAL;
 123
 124         if (!xskq_peek_addr(xs->umem->fq, &addr) ||
 125             len > xs->umem->chunk_size_nohr) {
 126                 xs->rx_dropped++;
 127                 return -ENOSPC;
 128         }
 129
 130         addr += xs->umem->headroom;
 131
 132         buffer = xdp_umem_get_data(xs->umem, addr);
 133         memcpy(buffer, xdp->data, len);
 134         err = xskq_produce_batch_desc(xs->rx, addr, len);
 135         if (!err) {
 136                 xskq_discard_addr(xs->umem->fq);
 137                 xsk_flush(xs);
 138                 return 0;
 139         }
 140
 141         xs->rx_dropped++;
 142         return err;
 143 }
 144
 145 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries)
 146 {
 147         xskq_produce_flush_addr_n(umem->cq, nb_entries);
 148 }
 149 EXPORT_SYMBOL(xsk_umem_complete_tx);
 150
 151 void xsk_umem_consume_tx_done(struct xdp_umem *umem)
 152 {
 153         struct xdp_sock *xs;
 154
 155         rcu_read_lock();
 156         list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
 157                 xs->sk.sk_write_space(&xs->sk);
 158         }
 159         rcu_read_unlock();
 160 }
 161 EXPORT_SYMBOL(xsk_umem_consume_tx_done);
 162
 163 bool xsk_umem_consume_tx(struct xdp_umem *umem, dma_addr_t *dma, u32 *len)
 164 {
 165         struct xdp_desc desc;
 166         struct xdp_sock *xs;
 167
 168         rcu_read_lock();
 169         list_for_each_entry_rcu(xs, &umem->xsk_list, list) {
 170                 if (!xskq_peek_desc(xs->tx, &desc))
 171                         continue;
 172
 173                 if (xskq_produce_addr_lazy(umem->cq, desc.addr))
 174                         goto out;
 175
 176                 *dma = xdp_umem_get_dma(umem, desc.addr);
 177                 *len = desc.len;
 178
 179                 xskq_discard_desc(xs->tx);
 180                 rcu_read_unlock();
 181                 return true;
 182         }
 183
 184 out:
 185         rcu_read_unlock();
 186         return false;
 187 }
 188 EXPORT_SYMBOL(xsk_umem_consume_tx);
 189
 190 static int xsk_zc_xmit(struct sock *sk)
 191 {
 192         struct xdp_sock *xs = xdp_sk(sk);
 193         struct net_device *dev = xs->dev;
 194
 195         return dev->netdev_ops->ndo_xsk_async_xmit(dev, xs->queue_id);
 196 }
 197
 198 static void xsk_destruct_skb(struct sk_buff *skb)
 199 {
 200         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 201         struct xdp_sock *xs = xdp_sk(skb->sk);
 202
 203         WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr));
 204
 205         sock_wfree(skb);
 206 }
 207
 208 static int xsk_generic_xmit(struct sock *sk, struct msghdr *m,
 209                             size_t total_len)
 210 {
 211         u32 max_batch = TX_BATCH_SIZE;
 212         struct xdp_sock *xs = xdp_sk(sk);
 213         bool sent_frame = false;
 214         struct xdp_desc desc;
 215         struct sk_buff *skb;
 216         int err = 0;
 217
 218         if (unlikely(!xs->tx))
 219                 return -ENOBUFS;
 220
 221         mutex_lock(&xs->mutex);
 222
 223         while (xskq_peek_desc(xs->tx, &desc)) {
 224                 char *buffer;
 225                 u64 addr;
 226                 u32 len;
 227
 228                 if (max_batch-- == 0) {
 229                         err = -EAGAIN;
 230                         goto out;
 231                 }
 232
 233                 if (xskq_reserve_addr(xs->umem->cq)) {
 234                         err = -EAGAIN;
 235                         goto out;
 236                 }
 237
 238                 len = desc.len;
 239                 if (unlikely(len > xs->dev->mtu)) {
 240                         err = -EMSGSIZE;
 241                         goto out;
 242                 }
 243
 244                 if (xs->queue_id >= xs->dev->real_num_tx_queues) {
 245                         err = -ENXIO;
 246                         goto out;
 247                 }
 248
 249                 skb = sock_alloc_send_skb(sk, len, 1, &err);
 250                 if (unlikely(!skb)) {
 251                         err = -EAGAIN;
 252                         goto out;
 253                 }
 254
 255                 skb_put(skb, len);
 256                 addr = desc.addr;
 257                 buffer = xdp_umem_get_data(xs->umem, addr);
 258                 err = skb_store_bits(skb, 0, buffer, len);
 259                 if (unlikely(err)) {
 260                         kfree_skb(skb);
 261                         goto out;
 262                 }
 263
 264                 skb->dev = xs->dev;
 265                 skb->priority = sk->sk_priority;
 266                 skb->mark = sk->sk_mark;
 267                 skb_shinfo(skb)->destructor_arg = (void *)(long)addr;
 268                 skb->destructor = xsk_destruct_skb;
 269
 270                 err = dev_direct_xmit(skb, xs->queue_id);
 271                 /* Ignore NET_XMIT_CN as packet might have been sent */
 272                 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) {
 273                         err = -EAGAIN;
 274                         /* SKB consumed by dev_direct_xmit() */
 275                         goto out;
 276                 }
 277
 278                 sent_frame = true;
 279                 xskq_discard_desc(xs->tx);
 280         }
 281
 282 out:
 283         if (sent_frame)
 284                 sk->sk_write_space(sk);
 285
 286         mutex_unlock(&xs->mutex);
 287         return err;
 288 }
 289
 290 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 291 {
 292         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 293         struct sock *sk = sock->sk;
 294         struct xdp_sock *xs = xdp_sk(sk);
 295
 296         if (unlikely(!xs->dev))
 297                 return -ENXIO;
 298         if (unlikely(!(xs->dev->flags & IFF_UP)))
 299                 return -ENETDOWN;
 300         if (need_wait)
 301                 return -EOPNOTSUPP;
 302
 303         return (xs->zc) ? xsk_zc_xmit(sk) : xsk_generic_xmit(sk, m, total_len);
 304 }
 305
 306 static __poll_t xsk_poll_mask(struct socket *sock, __poll_t events)
 307 {
 308         __poll_t mask = datagram_poll_mask(sock, events);
 309         struct sock *sk = sock->sk;
 310         struct xdp_sock *xs = xdp_sk(sk);
 311
 312         if (xs->rx && !xskq_empty_desc(xs->rx))
 313                 mask |= POLLIN | POLLRDNORM;
 314         if (xs->tx && !xskq_full_desc(xs->tx))
 315                 mask |= POLLOUT | POLLWRNORM;
 316
 317         return mask;
 318 }
 319
 320 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 321                           bool umem_queue)
 322 {
 323         struct xsk_queue *q;
 324
 325         if (entries == 0 || *queue || !is_power_of_2(entries))
 326                 return -EINVAL;
 327
 328         q = xskq_create(entries, umem_queue);
 329         if (!q)
 330                 return -ENOMEM;
 331
 332         /* Make sure queue is ready before it can be seen by others */
 333         smp_wmb();
 334         *queue = q;
 335         return 0;
 336 }
 337
 338 static int xsk_release(struct socket *sock)
 339 {
 340         struct sock *sk = sock->sk;
 341         struct xdp_sock *xs = xdp_sk(sk);
 342         struct net *net;
 343
 344         if (!sk)
 345                 return 0;
 346
 347         net = sock_net(sk);
 348
 349         local_bh_disable();
 350         sock_prot_inuse_add(net, sk->sk_prot, -1);
 351         local_bh_enable();
 352
 353         if (xs->dev) {
 354                 /* Wait for driver to stop using the xdp socket. */
 355                 synchronize_net();
 356                 dev_put(xs->dev);
 357                 xs->dev = NULL;
 358         }
 359
 360         sock_orphan(sk);
 361         sock->sk = NULL;
 362
 363         sk_refcnt_debug_release(sk);
 364         sock_put(sk);
 365
 366         return 0;
 367 }
 368
 369 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 370 {
 371         struct socket *sock;
 372         int err;
 373
 374         sock = sockfd_lookup(fd, &err);
 375         if (!sock)
 376                 return ERR_PTR(-ENOTSOCK);
 377
 378         if (sock->sk->sk_family != PF_XDP) {
 379                 sockfd_put(sock);
 380                 return ERR_PTR(-ENOPROTOOPT);
 381         }
 382
 383         return sock;
 384 }
 385
 386 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 387 {
 388         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 389         struct sock *sk = sock->sk;
 390         struct xdp_sock *xs = xdp_sk(sk);
 391         struct net_device *dev;
 392         u32 flags, qid;
 393         int err = 0;
 394
 395         if (addr_len < sizeof(struct sockaddr_xdp))
 396                 return -EINVAL;
 397         if (sxdp->sxdp_family != AF_XDP)
 398                 return -EINVAL;
 399
 400         mutex_lock(&xs->mutex);
 401         if (xs->dev) {
 402                 err = -EBUSY;
 403                 goto out_release;
 404         }
 405
 406         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 407         if (!dev) {
 408                 err = -ENODEV;
 409                 goto out_release;
 410         }
 411
 412         if (!xs->rx && !xs->tx) {
 413                 err = -EINVAL;
 414                 goto out_unlock;
 415         }
 416
 417         qid = sxdp->sxdp_queue_id;
 418
 419         if ((xs->rx && qid >= dev->real_num_rx_queues) ||
 420             (xs->tx && qid >= dev->real_num_tx_queues)) {
 421                 err = -EINVAL;
 422                 goto out_unlock;
 423         }
 424
 425         flags = sxdp->sxdp_flags;
 426
 427         if (flags & XDP_SHARED_UMEM) {
 428                 struct xdp_sock *umem_xs;
 429                 struct socket *sock;
 430
 431                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY)) {
 432                         /* Cannot specify flags for shared sockets. */
 433                         err = -EINVAL;
 434                         goto out_unlock;
 435                 }
 436
 437                 if (xs->umem) {
 438                         /* We have already our own. */
 439                         err = -EINVAL;
 440                         goto out_unlock;
 441                 }
 442
 443                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 444                 if (IS_ERR(sock)) {
 445                         err = PTR_ERR(sock);
 446                         goto out_unlock;
 447                 }
 448
 449                 umem_xs = xdp_sk(sock->sk);
 450                 if (!umem_xs->umem) {
 451                         /* No umem to inherit. */
 452                         err = -EBADF;
 453                         sockfd_put(sock);
 454                         goto out_unlock;
 455                 } else if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
 456                         err = -EINVAL;
 457                         sockfd_put(sock);
 458                         goto out_unlock;
 459                 }
 460
 461                 xdp_get_umem(umem_xs->umem);
 462                 xs->umem = umem_xs->umem;
 463                 sockfd_put(sock);
 464         } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
 465                 err = -EINVAL;
 466                 goto out_unlock;
 467         } else {
 468                 /* This xsk has its own umem. */
 469                 xskq_set_umem(xs->umem->fq, &xs->umem->props);
 470                 xskq_set_umem(xs->umem->cq, &xs->umem->props);
 471
 472                 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
 473                 if (err)
 474                         goto out_unlock;
 475         }
 476
 477         xs->dev = dev;
 478         xs->zc = xs->umem->zc;
 479         xs->queue_id = qid;
 480         xskq_set_umem(xs->rx, &xs->umem->props);
 481         xskq_set_umem(xs->tx, &xs->umem->props);
 482         xdp_add_sk_umem(xs->umem, xs);
 483
 484 out_unlock:
 485         if (err)
 486                 dev_put(dev);
 487 out_release:
 488         mutex_unlock(&xs->mutex);
 489         return err;
 490 }
 491
 492 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 493                           char __user *optval, unsigned int optlen)
 494 {
 495         struct sock *sk = sock->sk;
 496         struct xdp_sock *xs = xdp_sk(sk);
 497         int err;
 498
 499         if (level != SOL_XDP)
 500                 return -ENOPROTOOPT;
 501
 502         switch (optname) {
 503         case XDP_RX_RING:
 504         case XDP_TX_RING:
 505         {
 506                 struct xsk_queue **q;
 507                 int entries;
 508
 509                 if (optlen < sizeof(entries))
 510                         return -EINVAL;
 511                 if (copy_from_user(&entries, optval, sizeof(entries)))
 512                         return -EFAULT;
 513
 514                 mutex_lock(&xs->mutex);
 515                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
 516                 err = xsk_init_queue(entries, q, false);
 517                 mutex_unlock(&xs->mutex);
 518                 return err;
 519         }
 520         case XDP_UMEM_REG:
 521         {
 522                 struct xdp_umem_reg mr;
 523                 struct xdp_umem *umem;
 524
 525                 if (copy_from_user(&mr, optval, sizeof(mr)))
 526                         return -EFAULT;
 527
 528                 mutex_lock(&xs->mutex);
 529                 if (xs->umem) {
 530                         mutex_unlock(&xs->mutex);
 531                         return -EBUSY;
 532                 }
 533
 534                 umem = xdp_umem_create(&mr);
 535                 if (IS_ERR(umem)) {
 536                         mutex_unlock(&xs->mutex);
 537                         return PTR_ERR(umem);
 538                 }
 539
 540                 /* Make sure umem is ready before it can be seen by others */
 541                 smp_wmb();
 542                 xs->umem = umem;
 543                 mutex_unlock(&xs->mutex);
 544                 return 0;
 545         }
 546         case XDP_UMEM_FILL_RING:
 547         case XDP_UMEM_COMPLETION_RING:
 548         {
 549                 struct xsk_queue **q;
 550                 int entries;
 551
 552                 if (copy_from_user(&entries, optval, sizeof(entries)))
 553                         return -EFAULT;
 554
 555                 mutex_lock(&xs->mutex);
 556                 if (!xs->umem) {
 557                         mutex_unlock(&xs->mutex);
 558                         return -EINVAL;
 559                 }
 560
 561                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
 562                         &xs->umem->cq;
 563                 err = xsk_init_queue(entries, q, true);
 564                 mutex_unlock(&xs->mutex);
 565                 return err;
 566         }
 567         default:
 568                 break;
 569         }
 570
 571         return -ENOPROTOOPT;
 572 }
 573
 574 static int xsk_getsockopt(struct socket *sock, int level, int optname,
 575                           char __user *optval, int __user *optlen)
 576 {
 577         struct sock *sk = sock->sk;
 578         struct xdp_sock *xs = xdp_sk(sk);
 579         int len;
 580
 581         if (level != SOL_XDP)
 582                 return -ENOPROTOOPT;
 583
 584         if (get_user(len, optlen))
 585                 return -EFAULT;
 586         if (len < 0)
 587                 return -EINVAL;
 588
 589         switch (optname) {
 590         case XDP_STATISTICS:
 591         {
 592                 struct xdp_statistics stats;
 593
 594                 if (len < sizeof(stats))
 595                         return -EINVAL;
 596
 597                 mutex_lock(&xs->mutex);
 598                 stats.rx_dropped = xs->rx_dropped;
 599                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
 600                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
 601                 mutex_unlock(&xs->mutex);
 602
 603                 if (copy_to_user(optval, &stats, sizeof(stats)))
 604                         return -EFAULT;
 605                 if (put_user(sizeof(stats), optlen))
 606                         return -EFAULT;
 607
 608                 return 0;
 609         }
 610         case XDP_MMAP_OFFSETS:
 611         {
 612                 struct xdp_mmap_offsets off;
 613
 614                 if (len < sizeof(off))
 615                         return -EINVAL;
 616
 617                 off.rx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 618                 off.rx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 619                 off.rx.desc     = offsetof(struct xdp_rxtx_ring, desc);
 620                 off.tx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 621                 off.tx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 622                 off.tx.desc     = offsetof(struct xdp_rxtx_ring, desc);
 623
 624                 off.fr.producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 625                 off.fr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 626                 off.fr.desc     = offsetof(struct xdp_umem_ring, desc);
 627                 off.cr.producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 628                 off.cr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 629                 off.cr.desc     = offsetof(struct xdp_umem_ring, desc);
 630
 631                 len = sizeof(off);
 632                 if (copy_to_user(optval, &off, len))
 633                         return -EFAULT;
 634                 if (put_user(len, optlen))
 635                         return -EFAULT;
 636
 637                 return 0;
 638         }
 639         default:
 640                 break;
 641         }
 642
 643         return -EOPNOTSUPP;
 644 }
 645
 646 static int xsk_mmap(struct file *file, struct socket *sock,
 647                     struct vm_area_struct *vma)
 648 {
 649         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
 650         unsigned long size = vma->vm_end - vma->vm_start;
 651         struct xdp_sock *xs = xdp_sk(sock->sk);
 652         struct xsk_queue *q = NULL;
 653         struct xdp_umem *umem;
 654         unsigned long pfn;
 655         struct page *qpg;
 656
 657         if (offset == XDP_PGOFF_RX_RING) {
 658                 q = READ_ONCE(xs->rx);
 659         } else if (offset == XDP_PGOFF_TX_RING) {
 660                 q = READ_ONCE(xs->tx);
 661         } else {
 662                 umem = READ_ONCE(xs->umem);
 663                 if (!umem)
 664                         return -EINVAL;
 665
 666                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
 667                         q = READ_ONCE(umem->fq);
 668                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
 669                         q = READ_ONCE(umem->cq);
 670         }
 671
 672         if (!q)
 673                 return -EINVAL;
 674
 675         qpg = virt_to_head_page(q->ring);
 676         if (size > (PAGE_SIZE << compound_order(qpg)))
 677                 return -EINVAL;
 678
 679         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
 680         return remap_pfn_range(vma, vma->vm_start, pfn,
 681                                size, vma->vm_page_prot);
 682 }
 683
 684 static struct proto xsk_proto = {
 685         .name =         "XDP",
 686         .owner =        THIS_MODULE,
 687         .obj_size =     sizeof(struct xdp_sock),
 688 };
 689
 690 static const struct proto_ops xsk_proto_ops = {
 691         .family         = PF_XDP,
 692         .owner          = THIS_MODULE,
 693         .release        = xsk_release,
 694         .bind           = xsk_bind,
 695         .connect        = sock_no_connect,
 696         .socketpair     = sock_no_socketpair,
 697         .accept         = sock_no_accept,
 698         .getname        = sock_no_getname,
 699         .poll_mask      = xsk_poll_mask,
 700         .ioctl          = sock_no_ioctl,
 701         .listen         = sock_no_listen,
 702         .shutdown       = sock_no_shutdown,
 703         .setsockopt     = xsk_setsockopt,
 704         .getsockopt     = xsk_getsockopt,
 705         .sendmsg        = xsk_sendmsg,
 706         .recvmsg        = sock_no_recvmsg,
 707         .mmap           = xsk_mmap,
 708         .sendpage       = sock_no_sendpage,
 709 };
 710
 711 static void xsk_destruct(struct sock *sk)
 712 {
 713         struct xdp_sock *xs = xdp_sk(sk);
 714
 715         if (!sock_flag(sk, SOCK_DEAD))
 716                 return;
 717
 718         xskq_destroy(xs->rx);
 719         xskq_destroy(xs->tx);
 720         xdp_del_sk_umem(xs->umem, xs);
 721         xdp_put_umem(xs->umem);
 722
 723         sk_refcnt_debug_dec(sk);
 724 }
 725
 726 static int xsk_create(struct net *net, struct socket *sock, int protocol,
 727                       int kern)
 728 {
 729         struct sock *sk;
 730         struct xdp_sock *xs;
 731
 732         if (!ns_capable(net->user_ns, CAP_NET_RAW))
 733                 return -EPERM;
 734         if (sock->type != SOCK_RAW)
 735                 return -ESOCKTNOSUPPORT;
 736
 737         if (protocol)
 738                 return -EPROTONOSUPPORT;
 739
 740         sock->state = SS_UNCONNECTED;
 741
 742         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
 743         if (!sk)
 744                 return -ENOBUFS;
 745
 746         sock->ops = &xsk_proto_ops;
 747
 748         sock_init_data(sock, sk);
 749
 750         sk->sk_family = PF_XDP;
 751
 752         sk->sk_destruct = xsk_destruct;
 753         sk_refcnt_debug_inc(sk);
 754
 755         xs = xdp_sk(sk);
 756         mutex_init(&xs->mutex);
 757
 758         local_bh_disable();
 759         sock_prot_inuse_add(net, &xsk_proto, 1);
 760         local_bh_enable();
 761
 762         return 0;
 763 }
 764
 765 static const struct net_proto_family xsk_family_ops = {
 766         .family = PF_XDP,
 767         .create = xsk_create,
 768         .owner  = THIS_MODULE,
 769 };
 770
 771 static int __init xsk_init(void)
 772 {
 773         int err;
 774
 775         err = proto_register(&xsk_proto, 0 /* no slab */);
 776         if (err)
 777                 goto out;
 778
 779         err = sock_register(&xsk_family_ops);
 780         if (err)
 781                 goto out_proto;
 782
 783         return 0;
 784
 785 out_proto:
 786         proto_unregister(&xsk_proto);
 787 out:
 788         return err;
 789 }
 790
 791 fs_initcall(xsk_init);