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 (!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 __poll_t xsk_poll_mask(struct socket *sock, __poll_t events)
 304 {
 305         __poll_t mask = datagram_poll_mask(sock, events);
 306         struct sock *sk = sock->sk;
 307         struct xdp_sock *xs = xdp_sk(sk);
 308
 309         if (xs->rx && !xskq_empty_desc(xs->rx))
 310                 mask |= POLLIN | POLLRDNORM;
 311         if (xs->tx && !xskq_full_desc(xs->tx))
 312                 mask |= POLLOUT | POLLWRNORM;
 313
 314         return mask;
 315 }
 316
 317 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 318                           bool umem_queue)
 319 {
 320         struct xsk_queue *q;
 321
 322         if (entries == 0 || *queue || !is_power_of_2(entries))
 323                 return -EINVAL;
 324
 325         q = xskq_create(entries, umem_queue);
 326         if (!q)
 327                 return -ENOMEM;
 328
 329         /* Make sure queue is ready before it can be seen by others */
 330         smp_wmb();
 331         *queue = q;
 332         return 0;
 333 }
 334
 335 static int xsk_release(struct socket *sock)
 336 {
 337         struct sock *sk = sock->sk;
 338         struct xdp_sock *xs = xdp_sk(sk);
 339         struct net *net;
 340
 341         if (!sk)
 342                 return 0;
 343
 344         net = sock_net(sk);
 345
 346         local_bh_disable();
 347         sock_prot_inuse_add(net, sk->sk_prot, -1);
 348         local_bh_enable();
 349
 350         if (xs->dev) {
 351                 /* Wait for driver to stop using the xdp socket. */
 352                 synchronize_net();
 353                 dev_put(xs->dev);
 354                 xs->dev = NULL;
 355         }
 356
 357         sock_orphan(sk);
 358         sock->sk = NULL;
 359
 360         sk_refcnt_debug_release(sk);
 361         sock_put(sk);
 362
 363         return 0;
 364 }
 365
 366 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 367 {
 368         struct socket *sock;
 369         int err;
 370
 371         sock = sockfd_lookup(fd, &err);
 372         if (!sock)
 373                 return ERR_PTR(-ENOTSOCK);
 374
 375         if (sock->sk->sk_family != PF_XDP) {
 376                 sockfd_put(sock);
 377                 return ERR_PTR(-ENOPROTOOPT);
 378         }
 379
 380         return sock;
 381 }
 382
 383 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 384 {
 385         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 386         struct sock *sk = sock->sk;
 387         struct xdp_sock *xs = xdp_sk(sk);
 388         struct net_device *dev;
 389         u32 flags, qid;
 390         int err = 0;
 391
 392         if (addr_len < sizeof(struct sockaddr_xdp))
 393                 return -EINVAL;
 394         if (sxdp->sxdp_family != AF_XDP)
 395                 return -EINVAL;
 396
 397         mutex_lock(&xs->mutex);
 398         if (xs->dev) {
 399                 err = -EBUSY;
 400                 goto out_release;
 401         }
 402
 403         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 404         if (!dev) {
 405                 err = -ENODEV;
 406                 goto out_release;
 407         }
 408
 409         if (!xs->rx && !xs->tx) {
 410                 err = -EINVAL;
 411                 goto out_unlock;
 412         }
 413
 414         qid = sxdp->sxdp_queue_id;
 415
 416         if ((xs->rx && qid >= dev->real_num_rx_queues) ||
 417             (xs->tx && qid >= dev->real_num_tx_queues)) {
 418                 err = -EINVAL;
 419                 goto out_unlock;
 420         }
 421
 422         flags = sxdp->sxdp_flags;
 423
 424         if (flags & XDP_SHARED_UMEM) {
 425                 struct xdp_sock *umem_xs;
 426                 struct socket *sock;
 427
 428                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY)) {
 429                         /* Cannot specify flags for shared sockets. */
 430                         err = -EINVAL;
 431                         goto out_unlock;
 432                 }
 433
 434                 if (xs->umem) {
 435                         /* We have already our own. */
 436                         err = -EINVAL;
 437                         goto out_unlock;
 438                 }
 439
 440                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 441                 if (IS_ERR(sock)) {
 442                         err = PTR_ERR(sock);
 443                         goto out_unlock;
 444                 }
 445
 446                 umem_xs = xdp_sk(sock->sk);
 447                 if (!umem_xs->umem) {
 448                         /* No umem to inherit. */
 449                         err = -EBADF;
 450                         sockfd_put(sock);
 451                         goto out_unlock;
 452                 } else if (umem_xs->dev != dev || umem_xs->queue_id != qid) {
 453                         err = -EINVAL;
 454                         sockfd_put(sock);
 455                         goto out_unlock;
 456                 }
 457
 458                 xdp_get_umem(umem_xs->umem);
 459                 xs->umem = umem_xs->umem;
 460                 sockfd_put(sock);
 461         } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) {
 462                 err = -EINVAL;
 463                 goto out_unlock;
 464         } else {
 465                 /* This xsk has its own umem. */
 466                 xskq_set_umem(xs->umem->fq, &xs->umem->props);
 467                 xskq_set_umem(xs->umem->cq, &xs->umem->props);
 468
 469                 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags);
 470                 if (err)
 471                         goto out_unlock;
 472         }
 473
 474         xs->dev = dev;
 475         xs->zc = xs->umem->zc;
 476         xs->queue_id = qid;
 477         xskq_set_umem(xs->rx, &xs->umem->props);
 478         xskq_set_umem(xs->tx, &xs->umem->props);
 479         xdp_add_sk_umem(xs->umem, xs);
 480
 481 out_unlock:
 482         if (err)
 483                 dev_put(dev);
 484 out_release:
 485         mutex_unlock(&xs->mutex);
 486         return err;
 487 }
 488
 489 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 490                           char __user *optval, unsigned int optlen)
 491 {
 492         struct sock *sk = sock->sk;
 493         struct xdp_sock *xs = xdp_sk(sk);
 494         int err;
 495
 496         if (level != SOL_XDP)
 497                 return -ENOPROTOOPT;
 498
 499         switch (optname) {
 500         case XDP_RX_RING:
 501         case XDP_TX_RING:
 502         {
 503                 struct xsk_queue **q;
 504                 int entries;
 505
 506                 if (optlen < sizeof(entries))
 507                         return -EINVAL;
 508                 if (copy_from_user(&entries, optval, sizeof(entries)))
 509                         return -EFAULT;
 510
 511                 mutex_lock(&xs->mutex);
 512                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
 513                 err = xsk_init_queue(entries, q, false);
 514                 mutex_unlock(&xs->mutex);
 515                 return err;
 516         }
 517         case XDP_UMEM_REG:
 518         {
 519                 struct xdp_umem_reg mr;
 520                 struct xdp_umem *umem;
 521
 522                 if (copy_from_user(&mr, optval, sizeof(mr)))
 523                         return -EFAULT;
 524
 525                 mutex_lock(&xs->mutex);
 526                 if (xs->umem) {
 527                         mutex_unlock(&xs->mutex);
 528                         return -EBUSY;
 529                 }
 530
 531                 umem = xdp_umem_create(&mr);
 532                 if (IS_ERR(umem)) {
 533                         mutex_unlock(&xs->mutex);
 534                         return PTR_ERR(umem);
 535                 }
 536
 537                 /* Make sure umem is ready before it can be seen by others */
 538                 smp_wmb();
 539                 xs->umem = umem;
 540                 mutex_unlock(&xs->mutex);
 541                 return 0;
 542         }
 543         case XDP_UMEM_FILL_RING:
 544         case XDP_UMEM_COMPLETION_RING:
 545         {
 546                 struct xsk_queue **q;
 547                 int entries;
 548
 549                 if (copy_from_user(&entries, optval, sizeof(entries)))
 550                         return -EFAULT;
 551
 552                 mutex_lock(&xs->mutex);
 553                 if (!xs->umem) {
 554                         mutex_unlock(&xs->mutex);
 555                         return -EINVAL;
 556                 }
 557
 558                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq :
 559                         &xs->umem->cq;
 560                 err = xsk_init_queue(entries, q, true);
 561                 mutex_unlock(&xs->mutex);
 562                 return err;
 563         }
 564         default:
 565                 break;
 566         }
 567
 568         return -ENOPROTOOPT;
 569 }
 570
 571 static int xsk_getsockopt(struct socket *sock, int level, int optname,
 572                           char __user *optval, int __user *optlen)
 573 {
 574         struct sock *sk = sock->sk;
 575         struct xdp_sock *xs = xdp_sk(sk);
 576         int len;
 577
 578         if (level != SOL_XDP)
 579                 return -ENOPROTOOPT;
 580
 581         if (get_user(len, optlen))
 582                 return -EFAULT;
 583         if (len < 0)
 584                 return -EINVAL;
 585
 586         switch (optname) {
 587         case XDP_STATISTICS:
 588         {
 589                 struct xdp_statistics stats;
 590
 591                 if (len < sizeof(stats))
 592                         return -EINVAL;
 593
 594                 mutex_lock(&xs->mutex);
 595                 stats.rx_dropped = xs->rx_dropped;
 596                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
 597                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
 598                 mutex_unlock(&xs->mutex);
 599
 600                 if (copy_to_user(optval, &stats, sizeof(stats)))
 601                         return -EFAULT;
 602                 if (put_user(sizeof(stats), optlen))
 603                         return -EFAULT;
 604
 605                 return 0;
 606         }
 607         case XDP_MMAP_OFFSETS:
 608         {
 609                 struct xdp_mmap_offsets off;
 610
 611                 if (len < sizeof(off))
 612                         return -EINVAL;
 613
 614                 off.rx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 615                 off.rx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 616                 off.rx.desc     = offsetof(struct xdp_rxtx_ring, desc);
 617                 off.tx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 618                 off.tx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 619                 off.tx.desc     = offsetof(struct xdp_rxtx_ring, desc);
 620
 621                 off.fr.producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 622                 off.fr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 623                 off.fr.desc     = offsetof(struct xdp_umem_ring, desc);
 624                 off.cr.producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 625                 off.cr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 626                 off.cr.desc     = offsetof(struct xdp_umem_ring, desc);
 627
 628                 len = sizeof(off);
 629                 if (copy_to_user(optval, &off, len))
 630                         return -EFAULT;
 631                 if (put_user(len, optlen))
 632                         return -EFAULT;
 633
 634                 return 0;
 635         }
 636         default:
 637                 break;
 638         }
 639
 640         return -EOPNOTSUPP;
 641 }
 642
 643 static int xsk_mmap(struct file *file, struct socket *sock,
 644                     struct vm_area_struct *vma)
 645 {
 646         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
 647         unsigned long size = vma->vm_end - vma->vm_start;
 648         struct xdp_sock *xs = xdp_sk(sock->sk);
 649         struct xsk_queue *q = NULL;
 650         struct xdp_umem *umem;
 651         unsigned long pfn;
 652         struct page *qpg;
 653
 654         if (offset == XDP_PGOFF_RX_RING) {
 655                 q = READ_ONCE(xs->rx);
 656         } else if (offset == XDP_PGOFF_TX_RING) {
 657                 q = READ_ONCE(xs->tx);
 658         } else {
 659                 umem = READ_ONCE(xs->umem);
 660                 if (!umem)
 661                         return -EINVAL;
 662
 663                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
 664                         q = READ_ONCE(umem->fq);
 665                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
 666                         q = READ_ONCE(umem->cq);
 667         }
 668
 669         if (!q)
 670                 return -EINVAL;
 671
 672         qpg = virt_to_head_page(q->ring);
 673         if (size > (PAGE_SIZE << compound_order(qpg)))
 674                 return -EINVAL;
 675
 676         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
 677         return remap_pfn_range(vma, vma->vm_start, pfn,
 678                                size, vma->vm_page_prot);
 679 }
 680
 681 static struct proto xsk_proto = {
 682         .name =         "XDP",
 683         .owner =        THIS_MODULE,
 684         .obj_size =     sizeof(struct xdp_sock),
 685 };
 686
 687 static const struct proto_ops xsk_proto_ops = {
 688         .family         = PF_XDP,
 689         .owner          = THIS_MODULE,
 690         .release        = xsk_release,
 691         .bind           = xsk_bind,
 692         .connect        = sock_no_connect,
 693         .socketpair     = sock_no_socketpair,
 694         .accept         = sock_no_accept,
 695         .getname        = sock_no_getname,
 696         .poll_mask      = xsk_poll_mask,
 697         .ioctl          = sock_no_ioctl,
 698         .listen         = sock_no_listen,
 699         .shutdown       = sock_no_shutdown,
 700         .setsockopt     = xsk_setsockopt,
 701         .getsockopt     = xsk_getsockopt,
 702         .sendmsg        = xsk_sendmsg,
 703         .recvmsg        = sock_no_recvmsg,
 704         .mmap           = xsk_mmap,
 705         .sendpage       = sock_no_sendpage,
 706 };
 707
 708 static void xsk_destruct(struct sock *sk)
 709 {
 710         struct xdp_sock *xs = xdp_sk(sk);
 711
 712         if (!sock_flag(sk, SOCK_DEAD))
 713                 return;
 714
 715         xskq_destroy(xs->rx);
 716         xskq_destroy(xs->tx);
 717         xdp_del_sk_umem(xs->umem, xs);
 718         xdp_put_umem(xs->umem);
 719
 720         sk_refcnt_debug_dec(sk);
 721 }
 722
 723 static int xsk_create(struct net *net, struct socket *sock, int protocol,
 724                       int kern)
 725 {
 726         struct sock *sk;
 727         struct xdp_sock *xs;
 728
 729         if (!ns_capable(net->user_ns, CAP_NET_RAW))
 730                 return -EPERM;
 731         if (sock->type != SOCK_RAW)
 732                 return -ESOCKTNOSUPPORT;
 733
 734         if (protocol)
 735                 return -EPROTONOSUPPORT;
 736
 737         sock->state = SS_UNCONNECTED;
 738
 739         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
 740         if (!sk)
 741                 return -ENOBUFS;
 742
 743         sock->ops = &xsk_proto_ops;
 744
 745         sock_init_data(sock, sk);
 746
 747         sk->sk_family = PF_XDP;
 748
 749         sk->sk_destruct = xsk_destruct;
 750         sk_refcnt_debug_inc(sk);
 751
 752         xs = xdp_sk(sk);
 753         mutex_init(&xs->mutex);
 754
 755         local_bh_disable();
 756         sock_prot_inuse_add(net, &xsk_proto, 1);
 757         local_bh_enable();
 758
 759         return 0;
 760 }
 761
 762 static const struct net_proto_family xsk_family_ops = {
 763         .family = PF_XDP,
 764         .create = xsk_create,
 765         .owner  = THIS_MODULE,
 766 };
 767
 768 static int __init xsk_init(void)
 769 {
 770         int err;
 771
 772         err = proto_register(&xsk_proto, 0 /* no slab */);
 773         if (err)
 774                 goto out;
 775
 776         err = sock_register(&xsk_family_ops);
 777         if (err)
 778                 goto out_proto;
 779
 780         return 0;
 781
 782 out_proto:
 783         proto_unregister(&xsk_proto);
 784 out:
 785         return err;
 786 }
 787
 788 fs_initcall(xsk_init);