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_drv.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 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list); 35 36 bool xsk_is_setup_for_bpf_map(struct xdp_sock *xs) 37 { 38 return READ_ONCE(xs->rx) && READ_ONCE(xs->umem) && 39 READ_ONCE(xs->umem->fq); 40 } 41 42 void xsk_set_rx_need_wakeup(struct xdp_umem *umem) 43 { 44 if (umem->need_wakeup & XDP_WAKEUP_RX) 45 return; 46 47 umem->fq->ring->flags |= XDP_RING_NEED_WAKEUP; 48 umem->need_wakeup |= XDP_WAKEUP_RX; 49 } 50 EXPORT_SYMBOL(xsk_set_rx_need_wakeup); 51 52 void xsk_set_tx_need_wakeup(struct xdp_umem *umem) 53 { 54 struct xdp_sock *xs; 55 56 if (umem->need_wakeup & XDP_WAKEUP_TX) 57 return; 58 59 rcu_read_lock(); 60 list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) { 61 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 62 } 63 rcu_read_unlock(); 64 65 umem->need_wakeup |= XDP_WAKEUP_TX; 66 } 67 EXPORT_SYMBOL(xsk_set_tx_need_wakeup); 68 69 void xsk_clear_rx_need_wakeup(struct xdp_umem *umem) 70 { 71 if (!(umem->need_wakeup & XDP_WAKEUP_RX)) 72 return; 73 74 umem->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP; 75 umem->need_wakeup &= ~XDP_WAKEUP_RX; 76 } 77 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup); 78 79 void xsk_clear_tx_need_wakeup(struct xdp_umem *umem) 80 { 81 struct xdp_sock *xs; 82 83 if (!(umem->need_wakeup & XDP_WAKEUP_TX)) 84 return; 85 86 rcu_read_lock(); 87 list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) { 88 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP; 89 } 90 rcu_read_unlock(); 91 92 umem->need_wakeup &= ~XDP_WAKEUP_TX; 93 } 94 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup); 95 96 bool xsk_umem_uses_need_wakeup(struct xdp_umem *umem) 97 { 98 return umem->flags & XDP_UMEM_USES_NEED_WAKEUP; 99 } 100 EXPORT_SYMBOL(xsk_umem_uses_need_wakeup); 101 102 void xp_release(struct xdp_buff_xsk *xskb) 103 { 104 xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb; 105 } 106 107 static u64 xp_get_handle(struct xdp_buff_xsk *xskb) 108 { 109 u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start; 110 111 offset += xskb->pool->headroom; 112 if (!xskb->pool->unaligned) 113 return xskb->orig_addr + offset; 114 return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT); 115 } 116 117 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 118 { 119 struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp); 120 u64 addr; 121 int err; 122 123 addr = xp_get_handle(xskb); 124 err = xskq_prod_reserve_desc(xs->rx, addr, len); 125 if (err) { 126 xs->rx_queue_full++; 127 return err; 128 } 129 130 xp_release(xskb); 131 return 0; 132 } 133 134 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len) 135 { 136 void *from_buf, *to_buf; 137 u32 metalen; 138 139 if (unlikely(xdp_data_meta_unsupported(from))) { 140 from_buf = from->data; 141 to_buf = to->data; 142 metalen = 0; 143 } else { 144 from_buf = from->data_meta; 145 metalen = from->data - from->data_meta; 146 to_buf = to->data - metalen; 147 } 148 149 memcpy(to_buf, from_buf, len + metalen); 150 } 151 152 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len, 153 bool explicit_free) 154 { 155 struct xdp_buff *xsk_xdp; 156 int err; 157 158 if (len > xsk_umem_get_rx_frame_size(xs->umem)) { 159 xs->rx_dropped++; 160 return -ENOSPC; 161 } 162 163 xsk_xdp = xsk_buff_alloc(xs->umem); 164 if (!xsk_xdp) { 165 xs->rx_dropped++; 166 return -ENOSPC; 167 } 168 169 xsk_copy_xdp(xsk_xdp, xdp, len); 170 err = __xsk_rcv_zc(xs, xsk_xdp, len); 171 if (err) { 172 xsk_buff_free(xsk_xdp); 173 return err; 174 } 175 if (explicit_free) 176 xdp_return_buff(xdp); 177 return 0; 178 } 179 180 static bool xsk_is_bound(struct xdp_sock *xs) 181 { 182 if (READ_ONCE(xs->state) == XSK_BOUND) { 183 /* Matches smp_wmb() in bind(). */ 184 smp_rmb(); 185 return true; 186 } 187 return false; 188 } 189 190 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, 191 bool explicit_free) 192 { 193 u32 len; 194 195 if (!xsk_is_bound(xs)) 196 return -EINVAL; 197 198 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) 199 return -EINVAL; 200 201 len = xdp->data_end - xdp->data; 202 203 return xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL ? 204 __xsk_rcv_zc(xs, xdp, len) : 205 __xsk_rcv(xs, xdp, len, explicit_free); 206 } 207 208 static void xsk_flush(struct xdp_sock *xs) 209 { 210 xskq_prod_submit(xs->rx); 211 __xskq_cons_release(xs->umem->fq); 212 sock_def_readable(&xs->sk); 213 } 214 215 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 216 { 217 int err; 218 219 spin_lock_bh(&xs->rx_lock); 220 err = xsk_rcv(xs, xdp, false); 221 xsk_flush(xs); 222 spin_unlock_bh(&xs->rx_lock); 223 return err; 224 } 225 226 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp) 227 { 228 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list); 229 int err; 230 231 err = xsk_rcv(xs, xdp, true); 232 if (err) 233 return err; 234 235 if (!xs->flush_node.prev) 236 list_add(&xs->flush_node, flush_list); 237 238 return 0; 239 } 240 241 void __xsk_map_flush(void) 242 { 243 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list); 244 struct xdp_sock *xs, *tmp; 245 246 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) { 247 xsk_flush(xs); 248 __list_del_clearprev(&xs->flush_node); 249 } 250 } 251 252 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries) 253 { 254 xskq_prod_submit_n(umem->cq, nb_entries); 255 } 256 EXPORT_SYMBOL(xsk_umem_complete_tx); 257 258 void xsk_umem_consume_tx_done(struct xdp_umem *umem) 259 { 260 struct xdp_sock *xs; 261 262 rcu_read_lock(); 263 list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) { 264 __xskq_cons_release(xs->tx); 265 xs->sk.sk_write_space(&xs->sk); 266 } 267 rcu_read_unlock(); 268 } 269 EXPORT_SYMBOL(xsk_umem_consume_tx_done); 270 271 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc) 272 { 273 struct xdp_sock *xs; 274 275 rcu_read_lock(); 276 list_for_each_entry_rcu(xs, &umem->xsk_tx_list, list) { 277 if (!xskq_cons_peek_desc(xs->tx, desc, umem)) { 278 xs->tx->queue_empty_descs++; 279 continue; 280 } 281 282 /* This is the backpressure mechanism for the Tx path. 283 * Reserve space in the completion queue and only proceed 284 * if there is space in it. This avoids having to implement 285 * any buffering in the Tx path. 286 */ 287 if (xskq_prod_reserve_addr(umem->cq, desc->addr)) 288 goto out; 289 290 xskq_cons_release(xs->tx); 291 rcu_read_unlock(); 292 return true; 293 } 294 295 out: 296 rcu_read_unlock(); 297 return false; 298 } 299 EXPORT_SYMBOL(xsk_umem_consume_tx); 300 301 static int xsk_wakeup(struct xdp_sock *xs, u8 flags) 302 { 303 struct net_device *dev = xs->dev; 304 int err; 305 306 rcu_read_lock(); 307 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags); 308 rcu_read_unlock(); 309 310 return err; 311 } 312 313 static int xsk_zc_xmit(struct xdp_sock *xs) 314 { 315 return xsk_wakeup(xs, XDP_WAKEUP_TX); 316 } 317 318 static void xsk_destruct_skb(struct sk_buff *skb) 319 { 320 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg; 321 struct xdp_sock *xs = xdp_sk(skb->sk); 322 unsigned long flags; 323 324 spin_lock_irqsave(&xs->tx_completion_lock, flags); 325 xskq_prod_submit_addr(xs->umem->cq, addr); 326 spin_unlock_irqrestore(&xs->tx_completion_lock, flags); 327 328 sock_wfree(skb); 329 } 330 331 static int xsk_generic_xmit(struct sock *sk) 332 { 333 struct xdp_sock *xs = xdp_sk(sk); 334 u32 max_batch = TX_BATCH_SIZE; 335 bool sent_frame = false; 336 struct xdp_desc desc; 337 struct sk_buff *skb; 338 int err = 0; 339 340 mutex_lock(&xs->mutex); 341 342 if (xs->queue_id >= xs->dev->real_num_tx_queues) 343 goto out; 344 345 while (xskq_cons_peek_desc(xs->tx, &desc, xs->umem)) { 346 char *buffer; 347 u64 addr; 348 u32 len; 349 350 if (max_batch-- == 0) { 351 err = -EAGAIN; 352 goto out; 353 } 354 355 len = desc.len; 356 skb = sock_alloc_send_skb(sk, len, 1, &err); 357 if (unlikely(!skb)) 358 goto out; 359 360 skb_put(skb, len); 361 addr = desc.addr; 362 buffer = xsk_buff_raw_get_data(xs->umem, addr); 363 err = skb_store_bits(skb, 0, buffer, len); 364 /* This is the backpressure mechanism for the Tx path. 365 * Reserve space in the completion queue and only proceed 366 * if there is space in it. This avoids having to implement 367 * any buffering in the Tx path. 368 */ 369 if (unlikely(err) || xskq_prod_reserve(xs->umem->cq)) { 370 kfree_skb(skb); 371 goto out; 372 } 373 374 skb->dev = xs->dev; 375 skb->priority = sk->sk_priority; 376 skb->mark = sk->sk_mark; 377 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr; 378 skb->destructor = xsk_destruct_skb; 379 380 /* Hinder dev_direct_xmit from freeing the packet and 381 * therefore completing it in the destructor 382 */ 383 refcount_inc(&skb->users); 384 err = dev_direct_xmit(skb, xs->queue_id); 385 if (err == NETDEV_TX_BUSY) { 386 /* Tell user-space to retry the send */ 387 skb->destructor = sock_wfree; 388 /* Free skb without triggering the perf drop trace */ 389 consume_skb(skb); 390 err = -EAGAIN; 391 goto out; 392 } 393 394 xskq_cons_release(xs->tx); 395 /* Ignore NET_XMIT_CN as packet might have been sent */ 396 if (err == NET_XMIT_DROP) { 397 /* SKB completed but not sent */ 398 kfree_skb(skb); 399 err = -EBUSY; 400 goto out; 401 } 402 403 consume_skb(skb); 404 sent_frame = true; 405 } 406 407 xs->tx->queue_empty_descs++; 408 409 out: 410 if (sent_frame) 411 sk->sk_write_space(sk); 412 413 mutex_unlock(&xs->mutex); 414 return err; 415 } 416 417 static int __xsk_sendmsg(struct sock *sk) 418 { 419 struct xdp_sock *xs = xdp_sk(sk); 420 421 if (unlikely(!(xs->dev->flags & IFF_UP))) 422 return -ENETDOWN; 423 if (unlikely(!xs->tx)) 424 return -ENOBUFS; 425 426 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk); 427 } 428 429 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) 430 { 431 bool need_wait = !(m->msg_flags & MSG_DONTWAIT); 432 struct sock *sk = sock->sk; 433 struct xdp_sock *xs = xdp_sk(sk); 434 435 if (unlikely(!xsk_is_bound(xs))) 436 return -ENXIO; 437 if (unlikely(need_wait)) 438 return -EOPNOTSUPP; 439 440 return __xsk_sendmsg(sk); 441 } 442 443 static __poll_t xsk_poll(struct file *file, struct socket *sock, 444 struct poll_table_struct *wait) 445 { 446 __poll_t mask = datagram_poll(file, sock, wait); 447 struct sock *sk = sock->sk; 448 struct xdp_sock *xs = xdp_sk(sk); 449 struct xdp_umem *umem; 450 451 if (unlikely(!xsk_is_bound(xs))) 452 return mask; 453 454 umem = xs->umem; 455 456 if (umem->need_wakeup) { 457 if (xs->zc) 458 xsk_wakeup(xs, umem->need_wakeup); 459 else 460 /* Poll needs to drive Tx also in copy mode */ 461 __xsk_sendmsg(sk); 462 } 463 464 if (xs->rx && !xskq_prod_is_empty(xs->rx)) 465 mask |= EPOLLIN | EPOLLRDNORM; 466 if (xs->tx && !xskq_cons_is_full(xs->tx)) 467 mask |= EPOLLOUT | EPOLLWRNORM; 468 469 return mask; 470 } 471 472 static int xsk_init_queue(u32 entries, struct xsk_queue **queue, 473 bool umem_queue) 474 { 475 struct xsk_queue *q; 476 477 if (entries == 0 || *queue || !is_power_of_2(entries)) 478 return -EINVAL; 479 480 q = xskq_create(entries, umem_queue); 481 if (!q) 482 return -ENOMEM; 483 484 /* Make sure queue is ready before it can be seen by others */ 485 smp_wmb(); 486 WRITE_ONCE(*queue, q); 487 return 0; 488 } 489 490 static void xsk_unbind_dev(struct xdp_sock *xs) 491 { 492 struct net_device *dev = xs->dev; 493 494 if (xs->state != XSK_BOUND) 495 return; 496 WRITE_ONCE(xs->state, XSK_UNBOUND); 497 498 /* Wait for driver to stop using the xdp socket. */ 499 xdp_del_sk_umem(xs->umem, xs); 500 xs->dev = NULL; 501 synchronize_net(); 502 dev_put(dev); 503 } 504 505 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs, 506 struct xdp_sock ***map_entry) 507 { 508 struct xsk_map *map = NULL; 509 struct xsk_map_node *node; 510 511 *map_entry = NULL; 512 513 spin_lock_bh(&xs->map_list_lock); 514 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node, 515 node); 516 if (node) { 517 WARN_ON(xsk_map_inc(node->map)); 518 map = node->map; 519 *map_entry = node->map_entry; 520 } 521 spin_unlock_bh(&xs->map_list_lock); 522 return map; 523 } 524 525 static void xsk_delete_from_maps(struct xdp_sock *xs) 526 { 527 /* This function removes the current XDP socket from all the 528 * maps it resides in. We need to take extra care here, due to 529 * the two locks involved. Each map has a lock synchronizing 530 * updates to the entries, and each socket has a lock that 531 * synchronizes access to the list of maps (map_list). For 532 * deadlock avoidance the locks need to be taken in the order 533 * "map lock"->"socket map list lock". We start off by 534 * accessing the socket map list, and take a reference to the 535 * map to guarantee existence between the 536 * xsk_get_map_list_entry() and xsk_map_try_sock_delete() 537 * calls. Then we ask the map to remove the socket, which 538 * tries to remove the socket from the map. Note that there 539 * might be updates to the map between 540 * xsk_get_map_list_entry() and xsk_map_try_sock_delete(). 541 */ 542 struct xdp_sock **map_entry = NULL; 543 struct xsk_map *map; 544 545 while ((map = xsk_get_map_list_entry(xs, &map_entry))) { 546 xsk_map_try_sock_delete(map, xs, map_entry); 547 xsk_map_put(map); 548 } 549 } 550 551 static int xsk_release(struct socket *sock) 552 { 553 struct sock *sk = sock->sk; 554 struct xdp_sock *xs = xdp_sk(sk); 555 struct net *net; 556 557 if (!sk) 558 return 0; 559 560 net = sock_net(sk); 561 562 mutex_lock(&net->xdp.lock); 563 sk_del_node_init_rcu(sk); 564 mutex_unlock(&net->xdp.lock); 565 566 local_bh_disable(); 567 sock_prot_inuse_add(net, sk->sk_prot, -1); 568 local_bh_enable(); 569 570 xsk_delete_from_maps(xs); 571 mutex_lock(&xs->mutex); 572 xsk_unbind_dev(xs); 573 mutex_unlock(&xs->mutex); 574 575 xskq_destroy(xs->rx); 576 xskq_destroy(xs->tx); 577 578 sock_orphan(sk); 579 sock->sk = NULL; 580 581 sk_refcnt_debug_release(sk); 582 sock_put(sk); 583 584 return 0; 585 } 586 587 static struct socket *xsk_lookup_xsk_from_fd(int fd) 588 { 589 struct socket *sock; 590 int err; 591 592 sock = sockfd_lookup(fd, &err); 593 if (!sock) 594 return ERR_PTR(-ENOTSOCK); 595 596 if (sock->sk->sk_family != PF_XDP) { 597 sockfd_put(sock); 598 return ERR_PTR(-ENOPROTOOPT); 599 } 600 601 return sock; 602 } 603 604 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len) 605 { 606 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr; 607 struct sock *sk = sock->sk; 608 struct xdp_sock *xs = xdp_sk(sk); 609 struct net_device *dev; 610 u32 flags, qid; 611 int err = 0; 612 613 if (addr_len < sizeof(struct sockaddr_xdp)) 614 return -EINVAL; 615 if (sxdp->sxdp_family != AF_XDP) 616 return -EINVAL; 617 618 flags = sxdp->sxdp_flags; 619 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY | 620 XDP_USE_NEED_WAKEUP)) 621 return -EINVAL; 622 623 rtnl_lock(); 624 mutex_lock(&xs->mutex); 625 if (xs->state != XSK_READY) { 626 err = -EBUSY; 627 goto out_release; 628 } 629 630 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex); 631 if (!dev) { 632 err = -ENODEV; 633 goto out_release; 634 } 635 636 if (!xs->rx && !xs->tx) { 637 err = -EINVAL; 638 goto out_unlock; 639 } 640 641 qid = sxdp->sxdp_queue_id; 642 643 if (flags & XDP_SHARED_UMEM) { 644 struct xdp_sock *umem_xs; 645 struct socket *sock; 646 647 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) || 648 (flags & XDP_USE_NEED_WAKEUP)) { 649 /* Cannot specify flags for shared sockets. */ 650 err = -EINVAL; 651 goto out_unlock; 652 } 653 654 if (xs->umem) { 655 /* We have already our own. */ 656 err = -EINVAL; 657 goto out_unlock; 658 } 659 660 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd); 661 if (IS_ERR(sock)) { 662 err = PTR_ERR(sock); 663 goto out_unlock; 664 } 665 666 umem_xs = xdp_sk(sock->sk); 667 if (!xsk_is_bound(umem_xs)) { 668 err = -EBADF; 669 sockfd_put(sock); 670 goto out_unlock; 671 } 672 if (umem_xs->dev != dev || umem_xs->queue_id != qid) { 673 err = -EINVAL; 674 sockfd_put(sock); 675 goto out_unlock; 676 } 677 678 xdp_get_umem(umem_xs->umem); 679 WRITE_ONCE(xs->umem, umem_xs->umem); 680 sockfd_put(sock); 681 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) { 682 err = -EINVAL; 683 goto out_unlock; 684 } else { 685 /* This xsk has its own umem. */ 686 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags); 687 if (err) 688 goto out_unlock; 689 } 690 691 xs->dev = dev; 692 xs->zc = xs->umem->zc; 693 xs->queue_id = qid; 694 xdp_add_sk_umem(xs->umem, xs); 695 696 out_unlock: 697 if (err) { 698 dev_put(dev); 699 } else { 700 /* Matches smp_rmb() in bind() for shared umem 701 * sockets, and xsk_is_bound(). 702 */ 703 smp_wmb(); 704 WRITE_ONCE(xs->state, XSK_BOUND); 705 } 706 out_release: 707 mutex_unlock(&xs->mutex); 708 rtnl_unlock(); 709 return err; 710 } 711 712 struct xdp_umem_reg_v1 { 713 __u64 addr; /* Start of packet data area */ 714 __u64 len; /* Length of packet data area */ 715 __u32 chunk_size; 716 __u32 headroom; 717 }; 718 719 static int xsk_setsockopt(struct socket *sock, int level, int optname, 720 sockptr_t optval, unsigned int optlen) 721 { 722 struct sock *sk = sock->sk; 723 struct xdp_sock *xs = xdp_sk(sk); 724 int err; 725 726 if (level != SOL_XDP) 727 return -ENOPROTOOPT; 728 729 switch (optname) { 730 case XDP_RX_RING: 731 case XDP_TX_RING: 732 { 733 struct xsk_queue **q; 734 int entries; 735 736 if (optlen < sizeof(entries)) 737 return -EINVAL; 738 if (copy_from_sockptr(&entries, optval, sizeof(entries))) 739 return -EFAULT; 740 741 mutex_lock(&xs->mutex); 742 if (xs->state != XSK_READY) { 743 mutex_unlock(&xs->mutex); 744 return -EBUSY; 745 } 746 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx; 747 err = xsk_init_queue(entries, q, false); 748 if (!err && optname == XDP_TX_RING) 749 /* Tx needs to be explicitly woken up the first time */ 750 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 751 mutex_unlock(&xs->mutex); 752 return err; 753 } 754 case XDP_UMEM_REG: 755 { 756 size_t mr_size = sizeof(struct xdp_umem_reg); 757 struct xdp_umem_reg mr = {}; 758 struct xdp_umem *umem; 759 760 if (optlen < sizeof(struct xdp_umem_reg_v1)) 761 return -EINVAL; 762 else if (optlen < sizeof(mr)) 763 mr_size = sizeof(struct xdp_umem_reg_v1); 764 765 if (copy_from_sockptr(&mr, optval, mr_size)) 766 return -EFAULT; 767 768 mutex_lock(&xs->mutex); 769 if (xs->state != XSK_READY || xs->umem) { 770 mutex_unlock(&xs->mutex); 771 return -EBUSY; 772 } 773 774 umem = xdp_umem_create(&mr); 775 if (IS_ERR(umem)) { 776 mutex_unlock(&xs->mutex); 777 return PTR_ERR(umem); 778 } 779 780 /* Make sure umem is ready before it can be seen by others */ 781 smp_wmb(); 782 WRITE_ONCE(xs->umem, umem); 783 mutex_unlock(&xs->mutex); 784 return 0; 785 } 786 case XDP_UMEM_FILL_RING: 787 case XDP_UMEM_COMPLETION_RING: 788 { 789 struct xsk_queue **q; 790 int entries; 791 792 if (copy_from_sockptr(&entries, optval, sizeof(entries))) 793 return -EFAULT; 794 795 mutex_lock(&xs->mutex); 796 if (xs->state != XSK_READY) { 797 mutex_unlock(&xs->mutex); 798 return -EBUSY; 799 } 800 if (!xs->umem) { 801 mutex_unlock(&xs->mutex); 802 return -EINVAL; 803 } 804 805 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq : 806 &xs->umem->cq; 807 err = xsk_init_queue(entries, q, true); 808 if (optname == XDP_UMEM_FILL_RING) 809 xp_set_fq(xs->umem->pool, *q); 810 mutex_unlock(&xs->mutex); 811 return err; 812 } 813 default: 814 break; 815 } 816 817 return -ENOPROTOOPT; 818 } 819 820 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring) 821 { 822 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer); 823 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer); 824 ring->desc = offsetof(struct xdp_rxtx_ring, desc); 825 } 826 827 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring) 828 { 829 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer); 830 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer); 831 ring->desc = offsetof(struct xdp_umem_ring, desc); 832 } 833 834 struct xdp_statistics_v1 { 835 __u64 rx_dropped; 836 __u64 rx_invalid_descs; 837 __u64 tx_invalid_descs; 838 }; 839 840 static int xsk_getsockopt(struct socket *sock, int level, int optname, 841 char __user *optval, int __user *optlen) 842 { 843 struct sock *sk = sock->sk; 844 struct xdp_sock *xs = xdp_sk(sk); 845 int len; 846 847 if (level != SOL_XDP) 848 return -ENOPROTOOPT; 849 850 if (get_user(len, optlen)) 851 return -EFAULT; 852 if (len < 0) 853 return -EINVAL; 854 855 switch (optname) { 856 case XDP_STATISTICS: 857 { 858 struct xdp_statistics stats = {}; 859 bool extra_stats = true; 860 size_t stats_size; 861 862 if (len < sizeof(struct xdp_statistics_v1)) { 863 return -EINVAL; 864 } else if (len < sizeof(stats)) { 865 extra_stats = false; 866 stats_size = sizeof(struct xdp_statistics_v1); 867 } else { 868 stats_size = sizeof(stats); 869 } 870 871 mutex_lock(&xs->mutex); 872 stats.rx_dropped = xs->rx_dropped; 873 if (extra_stats) { 874 stats.rx_ring_full = xs->rx_queue_full; 875 stats.rx_fill_ring_empty_descs = 876 xs->umem ? xskq_nb_queue_empty_descs(xs->umem->fq) : 0; 877 stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx); 878 } else { 879 stats.rx_dropped += xs->rx_queue_full; 880 } 881 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx); 882 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx); 883 mutex_unlock(&xs->mutex); 884 885 if (copy_to_user(optval, &stats, stats_size)) 886 return -EFAULT; 887 if (put_user(stats_size, optlen)) 888 return -EFAULT; 889 890 return 0; 891 } 892 case XDP_MMAP_OFFSETS: 893 { 894 struct xdp_mmap_offsets off; 895 struct xdp_mmap_offsets_v1 off_v1; 896 bool flags_supported = true; 897 void *to_copy; 898 899 if (len < sizeof(off_v1)) 900 return -EINVAL; 901 else if (len < sizeof(off)) 902 flags_supported = false; 903 904 if (flags_supported) { 905 /* xdp_ring_offset is identical to xdp_ring_offset_v1 906 * except for the flags field added to the end. 907 */ 908 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 909 &off.rx); 910 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 911 &off.tx); 912 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 913 &off.fr); 914 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 915 &off.cr); 916 off.rx.flags = offsetof(struct xdp_rxtx_ring, 917 ptrs.flags); 918 off.tx.flags = offsetof(struct xdp_rxtx_ring, 919 ptrs.flags); 920 off.fr.flags = offsetof(struct xdp_umem_ring, 921 ptrs.flags); 922 off.cr.flags = offsetof(struct xdp_umem_ring, 923 ptrs.flags); 924 925 len = sizeof(off); 926 to_copy = &off; 927 } else { 928 xsk_enter_rxtx_offsets(&off_v1.rx); 929 xsk_enter_rxtx_offsets(&off_v1.tx); 930 xsk_enter_umem_offsets(&off_v1.fr); 931 xsk_enter_umem_offsets(&off_v1.cr); 932 933 len = sizeof(off_v1); 934 to_copy = &off_v1; 935 } 936 937 if (copy_to_user(optval, to_copy, len)) 938 return -EFAULT; 939 if (put_user(len, optlen)) 940 return -EFAULT; 941 942 return 0; 943 } 944 case XDP_OPTIONS: 945 { 946 struct xdp_options opts = {}; 947 948 if (len < sizeof(opts)) 949 return -EINVAL; 950 951 mutex_lock(&xs->mutex); 952 if (xs->zc) 953 opts.flags |= XDP_OPTIONS_ZEROCOPY; 954 mutex_unlock(&xs->mutex); 955 956 len = sizeof(opts); 957 if (copy_to_user(optval, &opts, len)) 958 return -EFAULT; 959 if (put_user(len, optlen)) 960 return -EFAULT; 961 962 return 0; 963 } 964 default: 965 break; 966 } 967 968 return -EOPNOTSUPP; 969 } 970 971 static int xsk_mmap(struct file *file, struct socket *sock, 972 struct vm_area_struct *vma) 973 { 974 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT; 975 unsigned long size = vma->vm_end - vma->vm_start; 976 struct xdp_sock *xs = xdp_sk(sock->sk); 977 struct xsk_queue *q = NULL; 978 struct xdp_umem *umem; 979 unsigned long pfn; 980 struct page *qpg; 981 982 if (READ_ONCE(xs->state) != XSK_READY) 983 return -EBUSY; 984 985 if (offset == XDP_PGOFF_RX_RING) { 986 q = READ_ONCE(xs->rx); 987 } else if (offset == XDP_PGOFF_TX_RING) { 988 q = READ_ONCE(xs->tx); 989 } else { 990 umem = READ_ONCE(xs->umem); 991 if (!umem) 992 return -EINVAL; 993 994 /* Matches the smp_wmb() in XDP_UMEM_REG */ 995 smp_rmb(); 996 if (offset == XDP_UMEM_PGOFF_FILL_RING) 997 q = READ_ONCE(umem->fq); 998 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING) 999 q = READ_ONCE(umem->cq); 1000 } 1001 1002 if (!q) 1003 return -EINVAL; 1004 1005 /* Matches the smp_wmb() in xsk_init_queue */ 1006 smp_rmb(); 1007 qpg = virt_to_head_page(q->ring); 1008 if (size > page_size(qpg)) 1009 return -EINVAL; 1010 1011 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT; 1012 return remap_pfn_range(vma, vma->vm_start, pfn, 1013 size, vma->vm_page_prot); 1014 } 1015 1016 static int xsk_notifier(struct notifier_block *this, 1017 unsigned long msg, void *ptr) 1018 { 1019 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 1020 struct net *net = dev_net(dev); 1021 struct sock *sk; 1022 1023 switch (msg) { 1024 case NETDEV_UNREGISTER: 1025 mutex_lock(&net->xdp.lock); 1026 sk_for_each(sk, &net->xdp.list) { 1027 struct xdp_sock *xs = xdp_sk(sk); 1028 1029 mutex_lock(&xs->mutex); 1030 if (xs->dev == dev) { 1031 sk->sk_err = ENETDOWN; 1032 if (!sock_flag(sk, SOCK_DEAD)) 1033 sk->sk_error_report(sk); 1034 1035 xsk_unbind_dev(xs); 1036 1037 /* Clear device references in umem. */ 1038 xdp_umem_clear_dev(xs->umem); 1039 } 1040 mutex_unlock(&xs->mutex); 1041 } 1042 mutex_unlock(&net->xdp.lock); 1043 break; 1044 } 1045 return NOTIFY_DONE; 1046 } 1047 1048 static struct proto xsk_proto = { 1049 .name = "XDP", 1050 .owner = THIS_MODULE, 1051 .obj_size = sizeof(struct xdp_sock), 1052 }; 1053 1054 static const struct proto_ops xsk_proto_ops = { 1055 .family = PF_XDP, 1056 .owner = THIS_MODULE, 1057 .release = xsk_release, 1058 .bind = xsk_bind, 1059 .connect = sock_no_connect, 1060 .socketpair = sock_no_socketpair, 1061 .accept = sock_no_accept, 1062 .getname = sock_no_getname, 1063 .poll = xsk_poll, 1064 .ioctl = sock_no_ioctl, 1065 .listen = sock_no_listen, 1066 .shutdown = sock_no_shutdown, 1067 .setsockopt = xsk_setsockopt, 1068 .getsockopt = xsk_getsockopt, 1069 .sendmsg = xsk_sendmsg, 1070 .recvmsg = sock_no_recvmsg, 1071 .mmap = xsk_mmap, 1072 .sendpage = sock_no_sendpage, 1073 }; 1074 1075 static void xsk_destruct(struct sock *sk) 1076 { 1077 struct xdp_sock *xs = xdp_sk(sk); 1078 1079 if (!sock_flag(sk, SOCK_DEAD)) 1080 return; 1081 1082 xdp_put_umem(xs->umem); 1083 1084 sk_refcnt_debug_dec(sk); 1085 } 1086 1087 static int xsk_create(struct net *net, struct socket *sock, int protocol, 1088 int kern) 1089 { 1090 struct sock *sk; 1091 struct xdp_sock *xs; 1092 1093 if (!ns_capable(net->user_ns, CAP_NET_RAW)) 1094 return -EPERM; 1095 if (sock->type != SOCK_RAW) 1096 return -ESOCKTNOSUPPORT; 1097 1098 if (protocol) 1099 return -EPROTONOSUPPORT; 1100 1101 sock->state = SS_UNCONNECTED; 1102 1103 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern); 1104 if (!sk) 1105 return -ENOBUFS; 1106 1107 sock->ops = &xsk_proto_ops; 1108 1109 sock_init_data(sock, sk); 1110 1111 sk->sk_family = PF_XDP; 1112 1113 sk->sk_destruct = xsk_destruct; 1114 sk_refcnt_debug_inc(sk); 1115 1116 sock_set_flag(sk, SOCK_RCU_FREE); 1117 1118 xs = xdp_sk(sk); 1119 xs->state = XSK_READY; 1120 mutex_init(&xs->mutex); 1121 spin_lock_init(&xs->rx_lock); 1122 spin_lock_init(&xs->tx_completion_lock); 1123 1124 INIT_LIST_HEAD(&xs->map_list); 1125 spin_lock_init(&xs->map_list_lock); 1126 1127 mutex_lock(&net->xdp.lock); 1128 sk_add_node_rcu(sk, &net->xdp.list); 1129 mutex_unlock(&net->xdp.lock); 1130 1131 local_bh_disable(); 1132 sock_prot_inuse_add(net, &xsk_proto, 1); 1133 local_bh_enable(); 1134 1135 return 0; 1136 } 1137 1138 static const struct net_proto_family xsk_family_ops = { 1139 .family = PF_XDP, 1140 .create = xsk_create, 1141 .owner = THIS_MODULE, 1142 }; 1143 1144 static struct notifier_block xsk_netdev_notifier = { 1145 .notifier_call = xsk_notifier, 1146 }; 1147 1148 static int __net_init xsk_net_init(struct net *net) 1149 { 1150 mutex_init(&net->xdp.lock); 1151 INIT_HLIST_HEAD(&net->xdp.list); 1152 return 0; 1153 } 1154 1155 static void __net_exit xsk_net_exit(struct net *net) 1156 { 1157 WARN_ON_ONCE(!hlist_empty(&net->xdp.list)); 1158 } 1159 1160 static struct pernet_operations xsk_net_ops = { 1161 .init = xsk_net_init, 1162 .exit = xsk_net_exit, 1163 }; 1164 1165 static int __init xsk_init(void) 1166 { 1167 int err, cpu; 1168 1169 err = proto_register(&xsk_proto, 0 /* no slab */); 1170 if (err) 1171 goto out; 1172 1173 err = sock_register(&xsk_family_ops); 1174 if (err) 1175 goto out_proto; 1176 1177 err = register_pernet_subsys(&xsk_net_ops); 1178 if (err) 1179 goto out_sk; 1180 1181 err = register_netdevice_notifier(&xsk_netdev_notifier); 1182 if (err) 1183 goto out_pernet; 1184 1185 for_each_possible_cpu(cpu) 1186 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu)); 1187 return 0; 1188 1189 out_pernet: 1190 unregister_pernet_subsys(&xsk_net_ops); 1191 out_sk: 1192 sock_unregister(PF_XDP); 1193 out_proto: 1194 proto_unregister(&xsk_proto); 1195 out: 1196 return err; 1197 } 1198 1199 fs_initcall(xsk_init); 1200