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 #include "xsk.h" 31 32 #define TX_BATCH_SIZE 16 33 34 bool xsk_is_setup_for_bpf_map(struct xdp_sock *xs) 35 { 36 return READ_ONCE(xs->rx) && READ_ONCE(xs->umem) && 37 READ_ONCE(xs->umem->fq); 38 } 39 40 u64 *xsk_umem_peek_addr(struct xdp_umem *umem, u64 *addr) 41 { 42 return xskq_peek_addr(umem->fq, addr); 43 } 44 EXPORT_SYMBOL(xsk_umem_peek_addr); 45 46 void xsk_umem_discard_addr(struct xdp_umem *umem) 47 { 48 xskq_discard_addr(umem->fq); 49 } 50 EXPORT_SYMBOL(xsk_umem_discard_addr); 51 52 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 53 { 54 void *to_buf, *from_buf; 55 u32 metalen; 56 u64 addr; 57 int err; 58 59 if (!xskq_peek_addr(xs->umem->fq, &addr) || 60 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) { 61 xs->rx_dropped++; 62 return -ENOSPC; 63 } 64 65 addr += xs->umem->headroom; 66 67 if (unlikely(xdp_data_meta_unsupported(xdp))) { 68 from_buf = xdp->data; 69 metalen = 0; 70 } else { 71 from_buf = xdp->data_meta; 72 metalen = xdp->data - xdp->data_meta; 73 } 74 75 to_buf = xdp_umem_get_data(xs->umem, addr); 76 memcpy(to_buf, from_buf, len + metalen); 77 addr += metalen; 78 err = xskq_produce_batch_desc(xs->rx, addr, len); 79 if (!err) { 80 xskq_discard_addr(xs->umem->fq); 81 xdp_return_buff(xdp); 82 return 0; 83 } 84 85 xs->rx_dropped++; 86 return err; 87 } 88 89 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 90 { 91 int err = xskq_produce_batch_desc(xs->rx, (u64)xdp->handle, len); 92 93 if (err) 94 xs->rx_dropped++; 95 96 return err; 97 } 98 99 int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 100 { 101 u32 len; 102 103 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) 104 return -EINVAL; 105 106 len = xdp->data_end - xdp->data; 107 108 return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ? 109 __xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len); 110 } 111 112 void xsk_flush(struct xdp_sock *xs) 113 { 114 xskq_produce_flush_desc(xs->rx); 115 xs->sk.sk_data_ready(&xs->sk); 116 } 117 118 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 119 { 120 u32 metalen = xdp->data - xdp->data_meta; 121 u32 len = xdp->data_end - xdp->data; 122 void *buffer; 123 u64 addr; 124 int err; 125 126 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) 127 return -EINVAL; 128 129 if (!xskq_peek_addr(xs->umem->fq, &addr) || 130 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) { 131 xs->rx_dropped++; 132 return -ENOSPC; 133 } 134 135 addr += xs->umem->headroom; 136 137 buffer = xdp_umem_get_data(xs->umem, addr); 138 memcpy(buffer, xdp->data_meta, len + metalen); 139 addr += metalen; 140 err = xskq_produce_batch_desc(xs->rx, addr, len); 141 if (!err) { 142 xskq_discard_addr(xs->umem->fq); 143 xsk_flush(xs); 144 return 0; 145 } 146 147 xs->rx_dropped++; 148 return err; 149 } 150 151 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries) 152 { 153 xskq_produce_flush_addr_n(umem->cq, nb_entries); 154 } 155 EXPORT_SYMBOL(xsk_umem_complete_tx); 156 157 void xsk_umem_consume_tx_done(struct xdp_umem *umem) 158 { 159 struct xdp_sock *xs; 160 161 rcu_read_lock(); 162 list_for_each_entry_rcu(xs, &umem->xsk_list, list) { 163 xs->sk.sk_write_space(&xs->sk); 164 } 165 rcu_read_unlock(); 166 } 167 EXPORT_SYMBOL(xsk_umem_consume_tx_done); 168 169 bool xsk_umem_consume_tx(struct xdp_umem *umem, dma_addr_t *dma, u32 *len) 170 { 171 struct xdp_desc desc; 172 struct xdp_sock *xs; 173 174 rcu_read_lock(); 175 list_for_each_entry_rcu(xs, &umem->xsk_list, list) { 176 if (!xskq_peek_desc(xs->tx, &desc)) 177 continue; 178 179 if (xskq_produce_addr_lazy(umem->cq, desc.addr)) 180 goto out; 181 182 *dma = xdp_umem_get_dma(umem, desc.addr); 183 *len = desc.len; 184 185 xskq_discard_desc(xs->tx); 186 rcu_read_unlock(); 187 return true; 188 } 189 190 out: 191 rcu_read_unlock(); 192 return false; 193 } 194 EXPORT_SYMBOL(xsk_umem_consume_tx); 195 196 static int xsk_zc_xmit(struct sock *sk) 197 { 198 struct xdp_sock *xs = xdp_sk(sk); 199 struct net_device *dev = xs->dev; 200 201 return dev->netdev_ops->ndo_xsk_async_xmit(dev, xs->queue_id); 202 } 203 204 static void xsk_destruct_skb(struct sk_buff *skb) 205 { 206 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg; 207 struct xdp_sock *xs = xdp_sk(skb->sk); 208 unsigned long flags; 209 210 spin_lock_irqsave(&xs->tx_completion_lock, flags); 211 WARN_ON_ONCE(xskq_produce_addr(xs->umem->cq, addr)); 212 spin_unlock_irqrestore(&xs->tx_completion_lock, flags); 213 214 sock_wfree(skb); 215 } 216 217 static int xsk_generic_xmit(struct sock *sk, struct msghdr *m, 218 size_t total_len) 219 { 220 u32 max_batch = TX_BATCH_SIZE; 221 struct xdp_sock *xs = xdp_sk(sk); 222 bool sent_frame = false; 223 struct xdp_desc desc; 224 struct sk_buff *skb; 225 int err = 0; 226 227 mutex_lock(&xs->mutex); 228 229 while (xskq_peek_desc(xs->tx, &desc)) { 230 char *buffer; 231 u64 addr; 232 u32 len; 233 234 if (max_batch-- == 0) { 235 err = -EAGAIN; 236 goto out; 237 } 238 239 if (xskq_reserve_addr(xs->umem->cq)) 240 goto out; 241 242 if (xs->queue_id >= xs->dev->real_num_tx_queues) 243 goto out; 244 245 len = desc.len; 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 xskq_discard_desc(xs->tx); 269 /* Ignore NET_XMIT_CN as packet might have been sent */ 270 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) { 271 /* SKB completed but not sent */ 272 err = -EBUSY; 273 goto out; 274 } 275 276 sent_frame = true; 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 (unlikely(!xs->tx)) 298 return -ENOBUFS; 299 if (need_wait) 300 return -EOPNOTSUPP; 301 302 return (xs->zc) ? xsk_zc_xmit(sk) : xsk_generic_xmit(sk, m, total_len); 303 } 304 305 static unsigned int xsk_poll(struct file *file, struct socket *sock, 306 struct poll_table_struct *wait) 307 { 308 unsigned int mask = datagram_poll(file, sock, wait); 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 mutex_lock(&net->xdp.lock); 350 sk_del_node_init_rcu(sk); 351 mutex_unlock(&net->xdp.lock); 352 353 local_bh_disable(); 354 sock_prot_inuse_add(net, sk->sk_prot, -1); 355 local_bh_enable(); 356 357 if (xs->dev) { 358 struct net_device *dev = xs->dev; 359 360 /* Wait for driver to stop using the xdp socket. */ 361 xdp_del_sk_umem(xs->umem, xs); 362 xs->dev = NULL; 363 synchronize_net(); 364 dev_put(dev); 365 } 366 367 xskq_destroy(xs->rx); 368 xskq_destroy(xs->tx); 369 370 sock_orphan(sk); 371 sock->sk = NULL; 372 373 sk_refcnt_debug_release(sk); 374 sock_put(sk); 375 376 return 0; 377 } 378 379 static struct socket *xsk_lookup_xsk_from_fd(int fd) 380 { 381 struct socket *sock; 382 int err; 383 384 sock = sockfd_lookup(fd, &err); 385 if (!sock) 386 return ERR_PTR(-ENOTSOCK); 387 388 if (sock->sk->sk_family != PF_XDP) { 389 sockfd_put(sock); 390 return ERR_PTR(-ENOPROTOOPT); 391 } 392 393 return sock; 394 } 395 396 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len) 397 { 398 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr; 399 struct sock *sk = sock->sk; 400 struct xdp_sock *xs = xdp_sk(sk); 401 struct net_device *dev; 402 u32 flags, qid; 403 int err = 0; 404 405 if (addr_len < sizeof(struct sockaddr_xdp)) 406 return -EINVAL; 407 if (sxdp->sxdp_family != AF_XDP) 408 return -EINVAL; 409 410 mutex_lock(&xs->mutex); 411 if (xs->dev) { 412 err = -EBUSY; 413 goto out_release; 414 } 415 416 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex); 417 if (!dev) { 418 err = -ENODEV; 419 goto out_release; 420 } 421 422 if (!xs->rx && !xs->tx) { 423 err = -EINVAL; 424 goto out_unlock; 425 } 426 427 qid = sxdp->sxdp_queue_id; 428 flags = sxdp->sxdp_flags; 429 430 if (flags & XDP_SHARED_UMEM) { 431 struct xdp_sock *umem_xs; 432 struct socket *sock; 433 434 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY)) { 435 /* Cannot specify flags for shared sockets. */ 436 err = -EINVAL; 437 goto out_unlock; 438 } 439 440 if (xs->umem) { 441 /* We have already our own. */ 442 err = -EINVAL; 443 goto out_unlock; 444 } 445 446 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd); 447 if (IS_ERR(sock)) { 448 err = PTR_ERR(sock); 449 goto out_unlock; 450 } 451 452 umem_xs = xdp_sk(sock->sk); 453 if (!umem_xs->umem) { 454 /* No umem to inherit. */ 455 err = -EBADF; 456 sockfd_put(sock); 457 goto out_unlock; 458 } else if (umem_xs->dev != dev || umem_xs->queue_id != qid) { 459 err = -EINVAL; 460 sockfd_put(sock); 461 goto out_unlock; 462 } 463 464 xdp_get_umem(umem_xs->umem); 465 xs->umem = umem_xs->umem; 466 sockfd_put(sock); 467 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) { 468 err = -EINVAL; 469 goto out_unlock; 470 } else { 471 /* This xsk has its own umem. */ 472 xskq_set_umem(xs->umem->fq, xs->umem->size, 473 xs->umem->chunk_mask); 474 xskq_set_umem(xs->umem->cq, xs->umem->size, 475 xs->umem->chunk_mask); 476 477 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags); 478 if (err) 479 goto out_unlock; 480 } 481 482 xs->dev = dev; 483 xs->zc = xs->umem->zc; 484 xs->queue_id = qid; 485 xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask); 486 xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask); 487 xdp_add_sk_umem(xs->umem, xs); 488 489 out_unlock: 490 if (err) 491 dev_put(dev); 492 out_release: 493 mutex_unlock(&xs->mutex); 494 return err; 495 } 496 497 static int xsk_setsockopt(struct socket *sock, int level, int optname, 498 char __user *optval, unsigned int optlen) 499 { 500 struct sock *sk = sock->sk; 501 struct xdp_sock *xs = xdp_sk(sk); 502 int err; 503 504 if (level != SOL_XDP) 505 return -ENOPROTOOPT; 506 507 switch (optname) { 508 case XDP_RX_RING: 509 case XDP_TX_RING: 510 { 511 struct xsk_queue **q; 512 int entries; 513 514 if (optlen < sizeof(entries)) 515 return -EINVAL; 516 if (copy_from_user(&entries, optval, sizeof(entries))) 517 return -EFAULT; 518 519 mutex_lock(&xs->mutex); 520 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx; 521 err = xsk_init_queue(entries, q, false); 522 mutex_unlock(&xs->mutex); 523 return err; 524 } 525 case XDP_UMEM_REG: 526 { 527 struct xdp_umem_reg mr; 528 struct xdp_umem *umem; 529 530 if (copy_from_user(&mr, optval, sizeof(mr))) 531 return -EFAULT; 532 533 mutex_lock(&xs->mutex); 534 if (xs->umem) { 535 mutex_unlock(&xs->mutex); 536 return -EBUSY; 537 } 538 539 umem = xdp_umem_create(&mr); 540 if (IS_ERR(umem)) { 541 mutex_unlock(&xs->mutex); 542 return PTR_ERR(umem); 543 } 544 545 /* Make sure umem is ready before it can be seen by others */ 546 smp_wmb(); 547 xs->umem = umem; 548 mutex_unlock(&xs->mutex); 549 return 0; 550 } 551 case XDP_UMEM_FILL_RING: 552 case XDP_UMEM_COMPLETION_RING: 553 { 554 struct xsk_queue **q; 555 int entries; 556 557 if (copy_from_user(&entries, optval, sizeof(entries))) 558 return -EFAULT; 559 560 mutex_lock(&xs->mutex); 561 if (!xs->umem) { 562 mutex_unlock(&xs->mutex); 563 return -EINVAL; 564 } 565 566 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq : 567 &xs->umem->cq; 568 err = xsk_init_queue(entries, q, true); 569 mutex_unlock(&xs->mutex); 570 return err; 571 } 572 default: 573 break; 574 } 575 576 return -ENOPROTOOPT; 577 } 578 579 static int xsk_getsockopt(struct socket *sock, int level, int optname, 580 char __user *optval, int __user *optlen) 581 { 582 struct sock *sk = sock->sk; 583 struct xdp_sock *xs = xdp_sk(sk); 584 int len; 585 586 if (level != SOL_XDP) 587 return -ENOPROTOOPT; 588 589 if (get_user(len, optlen)) 590 return -EFAULT; 591 if (len < 0) 592 return -EINVAL; 593 594 switch (optname) { 595 case XDP_STATISTICS: 596 { 597 struct xdp_statistics stats; 598 599 if (len < sizeof(stats)) 600 return -EINVAL; 601 602 mutex_lock(&xs->mutex); 603 stats.rx_dropped = xs->rx_dropped; 604 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx); 605 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx); 606 mutex_unlock(&xs->mutex); 607 608 if (copy_to_user(optval, &stats, sizeof(stats))) 609 return -EFAULT; 610 if (put_user(sizeof(stats), optlen)) 611 return -EFAULT; 612 613 return 0; 614 } 615 case XDP_MMAP_OFFSETS: 616 { 617 struct xdp_mmap_offsets off; 618 619 if (len < sizeof(off)) 620 return -EINVAL; 621 622 off.rx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer); 623 off.rx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer); 624 off.rx.desc = offsetof(struct xdp_rxtx_ring, desc); 625 off.tx.producer = offsetof(struct xdp_rxtx_ring, ptrs.producer); 626 off.tx.consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer); 627 off.tx.desc = offsetof(struct xdp_rxtx_ring, desc); 628 629 off.fr.producer = offsetof(struct xdp_umem_ring, ptrs.producer); 630 off.fr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer); 631 off.fr.desc = offsetof(struct xdp_umem_ring, desc); 632 off.cr.producer = offsetof(struct xdp_umem_ring, ptrs.producer); 633 off.cr.consumer = offsetof(struct xdp_umem_ring, ptrs.consumer); 634 off.cr.desc = offsetof(struct xdp_umem_ring, desc); 635 636 len = sizeof(off); 637 if (copy_to_user(optval, &off, len)) 638 return -EFAULT; 639 if (put_user(len, optlen)) 640 return -EFAULT; 641 642 return 0; 643 } 644 default: 645 break; 646 } 647 648 return -EOPNOTSUPP; 649 } 650 651 static int xsk_mmap(struct file *file, struct socket *sock, 652 struct vm_area_struct *vma) 653 { 654 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT; 655 unsigned long size = vma->vm_end - vma->vm_start; 656 struct xdp_sock *xs = xdp_sk(sock->sk); 657 struct xsk_queue *q = NULL; 658 struct xdp_umem *umem; 659 unsigned long pfn; 660 struct page *qpg; 661 662 if (offset == XDP_PGOFF_RX_RING) { 663 q = READ_ONCE(xs->rx); 664 } else if (offset == XDP_PGOFF_TX_RING) { 665 q = READ_ONCE(xs->tx); 666 } else { 667 umem = READ_ONCE(xs->umem); 668 if (!umem) 669 return -EINVAL; 670 671 /* Matches the smp_wmb() in XDP_UMEM_REG */ 672 smp_rmb(); 673 if (offset == XDP_UMEM_PGOFF_FILL_RING) 674 q = READ_ONCE(umem->fq); 675 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING) 676 q = READ_ONCE(umem->cq); 677 } 678 679 if (!q) 680 return -EINVAL; 681 682 /* Matches the smp_wmb() in xsk_init_queue */ 683 smp_rmb(); 684 qpg = virt_to_head_page(q->ring); 685 if (size > (PAGE_SIZE << compound_order(qpg))) 686 return -EINVAL; 687 688 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT; 689 return remap_pfn_range(vma, vma->vm_start, pfn, 690 size, vma->vm_page_prot); 691 } 692 693 static struct proto xsk_proto = { 694 .name = "XDP", 695 .owner = THIS_MODULE, 696 .obj_size = sizeof(struct xdp_sock), 697 }; 698 699 static const struct proto_ops xsk_proto_ops = { 700 .family = PF_XDP, 701 .owner = THIS_MODULE, 702 .release = xsk_release, 703 .bind = xsk_bind, 704 .connect = sock_no_connect, 705 .socketpair = sock_no_socketpair, 706 .accept = sock_no_accept, 707 .getname = sock_no_getname, 708 .poll = xsk_poll, 709 .ioctl = sock_no_ioctl, 710 .listen = sock_no_listen, 711 .shutdown = sock_no_shutdown, 712 .setsockopt = xsk_setsockopt, 713 .getsockopt = xsk_getsockopt, 714 .sendmsg = xsk_sendmsg, 715 .recvmsg = sock_no_recvmsg, 716 .mmap = xsk_mmap, 717 .sendpage = sock_no_sendpage, 718 }; 719 720 static void xsk_destruct(struct sock *sk) 721 { 722 struct xdp_sock *xs = xdp_sk(sk); 723 724 if (!sock_flag(sk, SOCK_DEAD)) 725 return; 726 727 xdp_put_umem(xs->umem); 728 729 sk_refcnt_debug_dec(sk); 730 } 731 732 static int xsk_create(struct net *net, struct socket *sock, int protocol, 733 int kern) 734 { 735 struct sock *sk; 736 struct xdp_sock *xs; 737 738 if (!ns_capable(net->user_ns, CAP_NET_RAW)) 739 return -EPERM; 740 if (sock->type != SOCK_RAW) 741 return -ESOCKTNOSUPPORT; 742 743 if (protocol) 744 return -EPROTONOSUPPORT; 745 746 sock->state = SS_UNCONNECTED; 747 748 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern); 749 if (!sk) 750 return -ENOBUFS; 751 752 sock->ops = &xsk_proto_ops; 753 754 sock_init_data(sock, sk); 755 756 sk->sk_family = PF_XDP; 757 758 sk->sk_destruct = xsk_destruct; 759 sk_refcnt_debug_inc(sk); 760 761 sock_set_flag(sk, SOCK_RCU_FREE); 762 763 xs = xdp_sk(sk); 764 mutex_init(&xs->mutex); 765 spin_lock_init(&xs->tx_completion_lock); 766 767 mutex_lock(&net->xdp.lock); 768 sk_add_node_rcu(sk, &net->xdp.list); 769 mutex_unlock(&net->xdp.lock); 770 771 local_bh_disable(); 772 sock_prot_inuse_add(net, &xsk_proto, 1); 773 local_bh_enable(); 774 775 return 0; 776 } 777 778 static const struct net_proto_family xsk_family_ops = { 779 .family = PF_XDP, 780 .create = xsk_create, 781 .owner = THIS_MODULE, 782 }; 783 784 static int __net_init xsk_net_init(struct net *net) 785 { 786 mutex_init(&net->xdp.lock); 787 INIT_HLIST_HEAD(&net->xdp.list); 788 return 0; 789 } 790 791 static void __net_exit xsk_net_exit(struct net *net) 792 { 793 WARN_ON_ONCE(!hlist_empty(&net->xdp.list)); 794 } 795 796 static struct pernet_operations xsk_net_ops = { 797 .init = xsk_net_init, 798 .exit = xsk_net_exit, 799 }; 800 801 static int __init xsk_init(void) 802 { 803 int err; 804 805 err = proto_register(&xsk_proto, 0 /* no slab */); 806 if (err) 807 goto out; 808 809 err = sock_register(&xsk_family_ops); 810 if (err) 811 goto out_proto; 812 813 err = register_pernet_subsys(&xsk_net_ops); 814 if (err) 815 goto out_sk; 816 return 0; 817 818 out_sk: 819 sock_unregister(PF_XDP); 820 out_proto: 821 proto_unregister(&xsk_proto); 822 out: 823 return err; 824 } 825 826 fs_initcall(xsk_init); 827