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 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 bool xsk_umem_has_addrs(struct xdp_umem *umem, u32 cnt) 43 { 44 return xskq_cons_has_entries(umem->fq, cnt); 45 } 46 EXPORT_SYMBOL(xsk_umem_has_addrs); 47 48 bool xsk_umem_peek_addr(struct xdp_umem *umem, u64 *addr) 49 { 50 return xskq_cons_peek_addr(umem->fq, addr, umem); 51 } 52 EXPORT_SYMBOL(xsk_umem_peek_addr); 53 54 void xsk_umem_release_addr(struct xdp_umem *umem) 55 { 56 xskq_cons_release(umem->fq); 57 } 58 EXPORT_SYMBOL(xsk_umem_release_addr); 59 60 void xsk_set_rx_need_wakeup(struct xdp_umem *umem) 61 { 62 if (umem->need_wakeup & XDP_WAKEUP_RX) 63 return; 64 65 umem->fq->ring->flags |= XDP_RING_NEED_WAKEUP; 66 umem->need_wakeup |= XDP_WAKEUP_RX; 67 } 68 EXPORT_SYMBOL(xsk_set_rx_need_wakeup); 69 70 void xsk_set_tx_need_wakeup(struct xdp_umem *umem) 71 { 72 struct xdp_sock *xs; 73 74 if (umem->need_wakeup & XDP_WAKEUP_TX) 75 return; 76 77 rcu_read_lock(); 78 list_for_each_entry_rcu(xs, &umem->xsk_list, list) { 79 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 80 } 81 rcu_read_unlock(); 82 83 umem->need_wakeup |= XDP_WAKEUP_TX; 84 } 85 EXPORT_SYMBOL(xsk_set_tx_need_wakeup); 86 87 void xsk_clear_rx_need_wakeup(struct xdp_umem *umem) 88 { 89 if (!(umem->need_wakeup & XDP_WAKEUP_RX)) 90 return; 91 92 umem->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP; 93 umem->need_wakeup &= ~XDP_WAKEUP_RX; 94 } 95 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup); 96 97 void xsk_clear_tx_need_wakeup(struct xdp_umem *umem) 98 { 99 struct xdp_sock *xs; 100 101 if (!(umem->need_wakeup & XDP_WAKEUP_TX)) 102 return; 103 104 rcu_read_lock(); 105 list_for_each_entry_rcu(xs, &umem->xsk_list, list) { 106 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP; 107 } 108 rcu_read_unlock(); 109 110 umem->need_wakeup &= ~XDP_WAKEUP_TX; 111 } 112 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup); 113 114 bool xsk_umem_uses_need_wakeup(struct xdp_umem *umem) 115 { 116 return umem->flags & XDP_UMEM_USES_NEED_WAKEUP; 117 } 118 EXPORT_SYMBOL(xsk_umem_uses_need_wakeup); 119 120 /* If a buffer crosses a page boundary, we need to do 2 memcpy's, one for 121 * each page. This is only required in copy mode. 122 */ 123 static void __xsk_rcv_memcpy(struct xdp_umem *umem, u64 addr, void *from_buf, 124 u32 len, u32 metalen) 125 { 126 void *to_buf = xdp_umem_get_data(umem, addr); 127 128 addr = xsk_umem_add_offset_to_addr(addr); 129 if (xskq_cons_crosses_non_contig_pg(umem, addr, len + metalen)) { 130 void *next_pg_addr = umem->pages[(addr >> PAGE_SHIFT) + 1].addr; 131 u64 page_start = addr & ~(PAGE_SIZE - 1); 132 u64 first_len = PAGE_SIZE - (addr - page_start); 133 134 memcpy(to_buf, from_buf, first_len + metalen); 135 memcpy(next_pg_addr, from_buf + first_len, len - first_len); 136 137 return; 138 } 139 140 memcpy(to_buf, from_buf, len + metalen); 141 } 142 143 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 144 { 145 u64 offset = xs->umem->headroom; 146 u64 addr, memcpy_addr; 147 void *from_buf; 148 u32 metalen; 149 int err; 150 151 if (!xskq_cons_peek_addr(xs->umem->fq, &addr, xs->umem) || 152 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) { 153 xs->rx_dropped++; 154 return -ENOSPC; 155 } 156 157 if (unlikely(xdp_data_meta_unsupported(xdp))) { 158 from_buf = xdp->data; 159 metalen = 0; 160 } else { 161 from_buf = xdp->data_meta; 162 metalen = xdp->data - xdp->data_meta; 163 } 164 165 memcpy_addr = xsk_umem_adjust_offset(xs->umem, addr, offset); 166 __xsk_rcv_memcpy(xs->umem, memcpy_addr, from_buf, len, metalen); 167 168 offset += metalen; 169 addr = xsk_umem_adjust_offset(xs->umem, addr, offset); 170 err = xskq_prod_reserve_desc(xs->rx, addr, len); 171 if (!err) { 172 xskq_cons_release(xs->umem->fq); 173 xdp_return_buff(xdp); 174 return 0; 175 } 176 177 xs->rx_dropped++; 178 return err; 179 } 180 181 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 182 { 183 int err = xskq_prod_reserve_desc(xs->rx, xdp->handle, len); 184 185 if (err) 186 xs->rx_dropped++; 187 188 return err; 189 } 190 191 static bool xsk_is_bound(struct xdp_sock *xs) 192 { 193 if (READ_ONCE(xs->state) == XSK_BOUND) { 194 /* Matches smp_wmb() in bind(). */ 195 smp_rmb(); 196 return true; 197 } 198 return false; 199 } 200 201 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 202 { 203 u32 len; 204 205 if (!xsk_is_bound(xs)) 206 return -EINVAL; 207 208 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) 209 return -EINVAL; 210 211 len = xdp->data_end - xdp->data; 212 213 return (xdp->rxq->mem.type == MEM_TYPE_ZERO_COPY) ? 214 __xsk_rcv_zc(xs, xdp, len) : __xsk_rcv(xs, xdp, len); 215 } 216 217 static void xsk_flush(struct xdp_sock *xs) 218 { 219 xskq_prod_submit(xs->rx); 220 __xskq_cons_release(xs->umem->fq); 221 sock_def_readable(&xs->sk); 222 } 223 224 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 225 { 226 u32 metalen = xdp->data - xdp->data_meta; 227 u32 len = xdp->data_end - xdp->data; 228 u64 offset = xs->umem->headroom; 229 void *buffer; 230 u64 addr; 231 int err; 232 233 spin_lock_bh(&xs->rx_lock); 234 235 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) { 236 err = -EINVAL; 237 goto out_unlock; 238 } 239 240 if (!xskq_cons_peek_addr(xs->umem->fq, &addr, xs->umem) || 241 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) { 242 err = -ENOSPC; 243 goto out_drop; 244 } 245 246 addr = xsk_umem_adjust_offset(xs->umem, addr, offset); 247 buffer = xdp_umem_get_data(xs->umem, addr); 248 memcpy(buffer, xdp->data_meta, len + metalen); 249 250 addr = xsk_umem_adjust_offset(xs->umem, addr, metalen); 251 err = xskq_prod_reserve_desc(xs->rx, addr, len); 252 if (err) 253 goto out_drop; 254 255 xskq_cons_release(xs->umem->fq); 256 xskq_prod_submit(xs->rx); 257 258 spin_unlock_bh(&xs->rx_lock); 259 260 xs->sk.sk_data_ready(&xs->sk); 261 return 0; 262 263 out_drop: 264 xs->rx_dropped++; 265 out_unlock: 266 spin_unlock_bh(&xs->rx_lock); 267 return err; 268 } 269 270 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp) 271 { 272 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list); 273 int err; 274 275 err = xsk_rcv(xs, xdp); 276 if (err) 277 return err; 278 279 if (!xs->flush_node.prev) 280 list_add(&xs->flush_node, flush_list); 281 282 return 0; 283 } 284 285 void __xsk_map_flush(void) 286 { 287 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list); 288 struct xdp_sock *xs, *tmp; 289 290 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) { 291 xsk_flush(xs); 292 __list_del_clearprev(&xs->flush_node); 293 } 294 } 295 296 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries) 297 { 298 xskq_prod_submit_n(umem->cq, nb_entries); 299 } 300 EXPORT_SYMBOL(xsk_umem_complete_tx); 301 302 void xsk_umem_consume_tx_done(struct xdp_umem *umem) 303 { 304 struct xdp_sock *xs; 305 306 rcu_read_lock(); 307 list_for_each_entry_rcu(xs, &umem->xsk_list, list) { 308 __xskq_cons_release(xs->tx); 309 xs->sk.sk_write_space(&xs->sk); 310 } 311 rcu_read_unlock(); 312 } 313 EXPORT_SYMBOL(xsk_umem_consume_tx_done); 314 315 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc) 316 { 317 struct xdp_sock *xs; 318 319 rcu_read_lock(); 320 list_for_each_entry_rcu(xs, &umem->xsk_list, list) { 321 if (!xskq_cons_peek_desc(xs->tx, desc, umem)) 322 continue; 323 324 /* This is the backpreassure mechanism for the Tx path. 325 * Reserve space in the completion queue and only proceed 326 * if there is space in it. This avoids having to implement 327 * any buffering in the Tx path. 328 */ 329 if (xskq_prod_reserve_addr(umem->cq, desc->addr)) 330 goto out; 331 332 xskq_cons_release(xs->tx); 333 rcu_read_unlock(); 334 return true; 335 } 336 337 out: 338 rcu_read_unlock(); 339 return false; 340 } 341 EXPORT_SYMBOL(xsk_umem_consume_tx); 342 343 static int xsk_wakeup(struct xdp_sock *xs, u8 flags) 344 { 345 struct net_device *dev = xs->dev; 346 int err; 347 348 rcu_read_lock(); 349 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags); 350 rcu_read_unlock(); 351 352 return err; 353 } 354 355 static int xsk_zc_xmit(struct xdp_sock *xs) 356 { 357 return xsk_wakeup(xs, XDP_WAKEUP_TX); 358 } 359 360 static void xsk_destruct_skb(struct sk_buff *skb) 361 { 362 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg; 363 struct xdp_sock *xs = xdp_sk(skb->sk); 364 unsigned long flags; 365 366 spin_lock_irqsave(&xs->tx_completion_lock, flags); 367 xskq_prod_submit_addr(xs->umem->cq, addr); 368 spin_unlock_irqrestore(&xs->tx_completion_lock, flags); 369 370 sock_wfree(skb); 371 } 372 373 static int xsk_generic_xmit(struct sock *sk) 374 { 375 struct xdp_sock *xs = xdp_sk(sk); 376 u32 max_batch = TX_BATCH_SIZE; 377 bool sent_frame = false; 378 struct xdp_desc desc; 379 struct sk_buff *skb; 380 int err = 0; 381 382 mutex_lock(&xs->mutex); 383 384 if (xs->queue_id >= xs->dev->real_num_tx_queues) 385 goto out; 386 387 while (xskq_cons_peek_desc(xs->tx, &desc, xs->umem)) { 388 char *buffer; 389 u64 addr; 390 u32 len; 391 392 if (max_batch-- == 0) { 393 err = -EAGAIN; 394 goto out; 395 } 396 397 len = desc.len; 398 skb = sock_alloc_send_skb(sk, len, 1, &err); 399 if (unlikely(!skb)) { 400 err = -EAGAIN; 401 goto out; 402 } 403 404 skb_put(skb, len); 405 addr = desc.addr; 406 buffer = xdp_umem_get_data(xs->umem, addr); 407 err = skb_store_bits(skb, 0, buffer, len); 408 /* This is the backpreassure mechanism for the Tx path. 409 * Reserve space in the completion queue and only proceed 410 * if there is space in it. This avoids having to implement 411 * any buffering in the Tx path. 412 */ 413 if (unlikely(err) || xskq_prod_reserve(xs->umem->cq)) { 414 kfree_skb(skb); 415 goto out; 416 } 417 418 skb->dev = xs->dev; 419 skb->priority = sk->sk_priority; 420 skb->mark = sk->sk_mark; 421 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr; 422 skb->destructor = xsk_destruct_skb; 423 424 err = dev_direct_xmit(skb, xs->queue_id); 425 xskq_cons_release(xs->tx); 426 /* Ignore NET_XMIT_CN as packet might have been sent */ 427 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) { 428 /* SKB completed but not sent */ 429 err = -EBUSY; 430 goto out; 431 } 432 433 sent_frame = true; 434 } 435 436 out: 437 if (sent_frame) 438 sk->sk_write_space(sk); 439 440 mutex_unlock(&xs->mutex); 441 return err; 442 } 443 444 static int __xsk_sendmsg(struct sock *sk) 445 { 446 struct xdp_sock *xs = xdp_sk(sk); 447 448 if (unlikely(!(xs->dev->flags & IFF_UP))) 449 return -ENETDOWN; 450 if (unlikely(!xs->tx)) 451 return -ENOBUFS; 452 453 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk); 454 } 455 456 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) 457 { 458 bool need_wait = !(m->msg_flags & MSG_DONTWAIT); 459 struct sock *sk = sock->sk; 460 struct xdp_sock *xs = xdp_sk(sk); 461 462 if (unlikely(!xsk_is_bound(xs))) 463 return -ENXIO; 464 if (unlikely(need_wait)) 465 return -EOPNOTSUPP; 466 467 return __xsk_sendmsg(sk); 468 } 469 470 static __poll_t xsk_poll(struct file *file, struct socket *sock, 471 struct poll_table_struct *wait) 472 { 473 __poll_t mask = datagram_poll(file, sock, wait); 474 struct sock *sk = sock->sk; 475 struct xdp_sock *xs = xdp_sk(sk); 476 struct xdp_umem *umem; 477 478 if (unlikely(!xsk_is_bound(xs))) 479 return mask; 480 481 umem = xs->umem; 482 483 if (umem->need_wakeup) { 484 if (xs->zc) 485 xsk_wakeup(xs, umem->need_wakeup); 486 else 487 /* Poll needs to drive Tx also in copy mode */ 488 __xsk_sendmsg(sk); 489 } 490 491 if (xs->rx && !xskq_prod_is_empty(xs->rx)) 492 mask |= EPOLLIN | EPOLLRDNORM; 493 if (xs->tx && !xskq_cons_is_full(xs->tx)) 494 mask |= EPOLLOUT | EPOLLWRNORM; 495 496 return mask; 497 } 498 499 static int xsk_init_queue(u32 entries, struct xsk_queue **queue, 500 bool umem_queue) 501 { 502 struct xsk_queue *q; 503 504 if (entries == 0 || *queue || !is_power_of_2(entries)) 505 return -EINVAL; 506 507 q = xskq_create(entries, umem_queue); 508 if (!q) 509 return -ENOMEM; 510 511 /* Make sure queue is ready before it can be seen by others */ 512 smp_wmb(); 513 WRITE_ONCE(*queue, q); 514 return 0; 515 } 516 517 static void xsk_unbind_dev(struct xdp_sock *xs) 518 { 519 struct net_device *dev = xs->dev; 520 521 if (xs->state != XSK_BOUND) 522 return; 523 WRITE_ONCE(xs->state, XSK_UNBOUND); 524 525 /* Wait for driver to stop using the xdp socket. */ 526 xdp_del_sk_umem(xs->umem, xs); 527 xs->dev = NULL; 528 synchronize_net(); 529 dev_put(dev); 530 } 531 532 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs, 533 struct xdp_sock ***map_entry) 534 { 535 struct xsk_map *map = NULL; 536 struct xsk_map_node *node; 537 538 *map_entry = NULL; 539 540 spin_lock_bh(&xs->map_list_lock); 541 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node, 542 node); 543 if (node) { 544 WARN_ON(xsk_map_inc(node->map)); 545 map = node->map; 546 *map_entry = node->map_entry; 547 } 548 spin_unlock_bh(&xs->map_list_lock); 549 return map; 550 } 551 552 static void xsk_delete_from_maps(struct xdp_sock *xs) 553 { 554 /* This function removes the current XDP socket from all the 555 * maps it resides in. We need to take extra care here, due to 556 * the two locks involved. Each map has a lock synchronizing 557 * updates to the entries, and each socket has a lock that 558 * synchronizes access to the list of maps (map_list). For 559 * deadlock avoidance the locks need to be taken in the order 560 * "map lock"->"socket map list lock". We start off by 561 * accessing the socket map list, and take a reference to the 562 * map to guarantee existence between the 563 * xsk_get_map_list_entry() and xsk_map_try_sock_delete() 564 * calls. Then we ask the map to remove the socket, which 565 * tries to remove the socket from the map. Note that there 566 * might be updates to the map between 567 * xsk_get_map_list_entry() and xsk_map_try_sock_delete(). 568 */ 569 struct xdp_sock **map_entry = NULL; 570 struct xsk_map *map; 571 572 while ((map = xsk_get_map_list_entry(xs, &map_entry))) { 573 xsk_map_try_sock_delete(map, xs, map_entry); 574 xsk_map_put(map); 575 } 576 } 577 578 static int xsk_release(struct socket *sock) 579 { 580 struct sock *sk = sock->sk; 581 struct xdp_sock *xs = xdp_sk(sk); 582 struct net *net; 583 584 if (!sk) 585 return 0; 586 587 net = sock_net(sk); 588 589 mutex_lock(&net->xdp.lock); 590 sk_del_node_init_rcu(sk); 591 mutex_unlock(&net->xdp.lock); 592 593 local_bh_disable(); 594 sock_prot_inuse_add(net, sk->sk_prot, -1); 595 local_bh_enable(); 596 597 xsk_delete_from_maps(xs); 598 mutex_lock(&xs->mutex); 599 xsk_unbind_dev(xs); 600 mutex_unlock(&xs->mutex); 601 602 xskq_destroy(xs->rx); 603 xskq_destroy(xs->tx); 604 605 sock_orphan(sk); 606 sock->sk = NULL; 607 608 sk_refcnt_debug_release(sk); 609 sock_put(sk); 610 611 return 0; 612 } 613 614 static struct socket *xsk_lookup_xsk_from_fd(int fd) 615 { 616 struct socket *sock; 617 int err; 618 619 sock = sockfd_lookup(fd, &err); 620 if (!sock) 621 return ERR_PTR(-ENOTSOCK); 622 623 if (sock->sk->sk_family != PF_XDP) { 624 sockfd_put(sock); 625 return ERR_PTR(-ENOPROTOOPT); 626 } 627 628 return sock; 629 } 630 631 /* Check if umem pages are contiguous. 632 * If zero-copy mode, use the DMA address to do the page contiguity check 633 * For all other modes we use addr (kernel virtual address) 634 * Store the result in the low bits of addr. 635 */ 636 static void xsk_check_page_contiguity(struct xdp_umem *umem, u32 flags) 637 { 638 struct xdp_umem_page *pgs = umem->pages; 639 int i, is_contig; 640 641 for (i = 0; i < umem->npgs - 1; i++) { 642 is_contig = (flags & XDP_ZEROCOPY) ? 643 (pgs[i].dma + PAGE_SIZE == pgs[i + 1].dma) : 644 (pgs[i].addr + PAGE_SIZE == pgs[i + 1].addr); 645 pgs[i].addr += is_contig << XSK_NEXT_PG_CONTIG_SHIFT; 646 } 647 } 648 649 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len) 650 { 651 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr; 652 struct sock *sk = sock->sk; 653 struct xdp_sock *xs = xdp_sk(sk); 654 struct net_device *dev; 655 u32 flags, qid; 656 int err = 0; 657 658 if (addr_len < sizeof(struct sockaddr_xdp)) 659 return -EINVAL; 660 if (sxdp->sxdp_family != AF_XDP) 661 return -EINVAL; 662 663 flags = sxdp->sxdp_flags; 664 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY | 665 XDP_USE_NEED_WAKEUP)) 666 return -EINVAL; 667 668 rtnl_lock(); 669 mutex_lock(&xs->mutex); 670 if (xs->state != XSK_READY) { 671 err = -EBUSY; 672 goto out_release; 673 } 674 675 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex); 676 if (!dev) { 677 err = -ENODEV; 678 goto out_release; 679 } 680 681 if (!xs->rx && !xs->tx) { 682 err = -EINVAL; 683 goto out_unlock; 684 } 685 686 qid = sxdp->sxdp_queue_id; 687 688 if (flags & XDP_SHARED_UMEM) { 689 struct xdp_sock *umem_xs; 690 struct socket *sock; 691 692 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) || 693 (flags & XDP_USE_NEED_WAKEUP)) { 694 /* Cannot specify flags for shared sockets. */ 695 err = -EINVAL; 696 goto out_unlock; 697 } 698 699 if (xs->umem) { 700 /* We have already our own. */ 701 err = -EINVAL; 702 goto out_unlock; 703 } 704 705 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd); 706 if (IS_ERR(sock)) { 707 err = PTR_ERR(sock); 708 goto out_unlock; 709 } 710 711 umem_xs = xdp_sk(sock->sk); 712 if (!xsk_is_bound(umem_xs)) { 713 err = -EBADF; 714 sockfd_put(sock); 715 goto out_unlock; 716 } 717 if (umem_xs->dev != dev || umem_xs->queue_id != qid) { 718 err = -EINVAL; 719 sockfd_put(sock); 720 goto out_unlock; 721 } 722 723 xdp_get_umem(umem_xs->umem); 724 WRITE_ONCE(xs->umem, umem_xs->umem); 725 sockfd_put(sock); 726 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) { 727 err = -EINVAL; 728 goto out_unlock; 729 } else { 730 /* This xsk has its own umem. */ 731 xskq_set_umem(xs->umem->fq, xs->umem->size, 732 xs->umem->chunk_mask); 733 xskq_set_umem(xs->umem->cq, xs->umem->size, 734 xs->umem->chunk_mask); 735 736 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags); 737 if (err) 738 goto out_unlock; 739 740 xsk_check_page_contiguity(xs->umem, flags); 741 } 742 743 xs->dev = dev; 744 xs->zc = xs->umem->zc; 745 xs->queue_id = qid; 746 xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask); 747 xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask); 748 xdp_add_sk_umem(xs->umem, xs); 749 750 out_unlock: 751 if (err) { 752 dev_put(dev); 753 } else { 754 /* Matches smp_rmb() in bind() for shared umem 755 * sockets, and xsk_is_bound(). 756 */ 757 smp_wmb(); 758 WRITE_ONCE(xs->state, XSK_BOUND); 759 } 760 out_release: 761 mutex_unlock(&xs->mutex); 762 rtnl_unlock(); 763 return err; 764 } 765 766 struct xdp_umem_reg_v1 { 767 __u64 addr; /* Start of packet data area */ 768 __u64 len; /* Length of packet data area */ 769 __u32 chunk_size; 770 __u32 headroom; 771 }; 772 773 static int xsk_setsockopt(struct socket *sock, int level, int optname, 774 char __user *optval, unsigned int optlen) 775 { 776 struct sock *sk = sock->sk; 777 struct xdp_sock *xs = xdp_sk(sk); 778 int err; 779 780 if (level != SOL_XDP) 781 return -ENOPROTOOPT; 782 783 switch (optname) { 784 case XDP_RX_RING: 785 case XDP_TX_RING: 786 { 787 struct xsk_queue **q; 788 int entries; 789 790 if (optlen < sizeof(entries)) 791 return -EINVAL; 792 if (copy_from_user(&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 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx; 801 err = xsk_init_queue(entries, q, false); 802 if (!err && optname == XDP_TX_RING) 803 /* Tx needs to be explicitly woken up the first time */ 804 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 805 mutex_unlock(&xs->mutex); 806 return err; 807 } 808 case XDP_UMEM_REG: 809 { 810 size_t mr_size = sizeof(struct xdp_umem_reg); 811 struct xdp_umem_reg mr = {}; 812 struct xdp_umem *umem; 813 814 if (optlen < sizeof(struct xdp_umem_reg_v1)) 815 return -EINVAL; 816 else if (optlen < sizeof(mr)) 817 mr_size = sizeof(struct xdp_umem_reg_v1); 818 819 if (copy_from_user(&mr, optval, mr_size)) 820 return -EFAULT; 821 822 mutex_lock(&xs->mutex); 823 if (xs->state != XSK_READY || xs->umem) { 824 mutex_unlock(&xs->mutex); 825 return -EBUSY; 826 } 827 828 umem = xdp_umem_create(&mr); 829 if (IS_ERR(umem)) { 830 mutex_unlock(&xs->mutex); 831 return PTR_ERR(umem); 832 } 833 834 /* Make sure umem is ready before it can be seen by others */ 835 smp_wmb(); 836 WRITE_ONCE(xs->umem, umem); 837 mutex_unlock(&xs->mutex); 838 return 0; 839 } 840 case XDP_UMEM_FILL_RING: 841 case XDP_UMEM_COMPLETION_RING: 842 { 843 struct xsk_queue **q; 844 int entries; 845 846 if (copy_from_user(&entries, optval, sizeof(entries))) 847 return -EFAULT; 848 849 mutex_lock(&xs->mutex); 850 if (xs->state != XSK_READY) { 851 mutex_unlock(&xs->mutex); 852 return -EBUSY; 853 } 854 if (!xs->umem) { 855 mutex_unlock(&xs->mutex); 856 return -EINVAL; 857 } 858 859 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq : 860 &xs->umem->cq; 861 err = xsk_init_queue(entries, q, true); 862 mutex_unlock(&xs->mutex); 863 return err; 864 } 865 default: 866 break; 867 } 868 869 return -ENOPROTOOPT; 870 } 871 872 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring) 873 { 874 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer); 875 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer); 876 ring->desc = offsetof(struct xdp_rxtx_ring, desc); 877 } 878 879 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring) 880 { 881 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer); 882 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer); 883 ring->desc = offsetof(struct xdp_umem_ring, desc); 884 } 885 886 static int xsk_getsockopt(struct socket *sock, int level, int optname, 887 char __user *optval, int __user *optlen) 888 { 889 struct sock *sk = sock->sk; 890 struct xdp_sock *xs = xdp_sk(sk); 891 int len; 892 893 if (level != SOL_XDP) 894 return -ENOPROTOOPT; 895 896 if (get_user(len, optlen)) 897 return -EFAULT; 898 if (len < 0) 899 return -EINVAL; 900 901 switch (optname) { 902 case XDP_STATISTICS: 903 { 904 struct xdp_statistics stats; 905 906 if (len < sizeof(stats)) 907 return -EINVAL; 908 909 mutex_lock(&xs->mutex); 910 stats.rx_dropped = xs->rx_dropped; 911 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx); 912 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx); 913 mutex_unlock(&xs->mutex); 914 915 if (copy_to_user(optval, &stats, sizeof(stats))) 916 return -EFAULT; 917 if (put_user(sizeof(stats), optlen)) 918 return -EFAULT; 919 920 return 0; 921 } 922 case XDP_MMAP_OFFSETS: 923 { 924 struct xdp_mmap_offsets off; 925 struct xdp_mmap_offsets_v1 off_v1; 926 bool flags_supported = true; 927 void *to_copy; 928 929 if (len < sizeof(off_v1)) 930 return -EINVAL; 931 else if (len < sizeof(off)) 932 flags_supported = false; 933 934 if (flags_supported) { 935 /* xdp_ring_offset is identical to xdp_ring_offset_v1 936 * except for the flags field added to the end. 937 */ 938 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 939 &off.rx); 940 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 941 &off.tx); 942 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 943 &off.fr); 944 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 945 &off.cr); 946 off.rx.flags = offsetof(struct xdp_rxtx_ring, 947 ptrs.flags); 948 off.tx.flags = offsetof(struct xdp_rxtx_ring, 949 ptrs.flags); 950 off.fr.flags = offsetof(struct xdp_umem_ring, 951 ptrs.flags); 952 off.cr.flags = offsetof(struct xdp_umem_ring, 953 ptrs.flags); 954 955 len = sizeof(off); 956 to_copy = &off; 957 } else { 958 xsk_enter_rxtx_offsets(&off_v1.rx); 959 xsk_enter_rxtx_offsets(&off_v1.tx); 960 xsk_enter_umem_offsets(&off_v1.fr); 961 xsk_enter_umem_offsets(&off_v1.cr); 962 963 len = sizeof(off_v1); 964 to_copy = &off_v1; 965 } 966 967 if (copy_to_user(optval, to_copy, len)) 968 return -EFAULT; 969 if (put_user(len, optlen)) 970 return -EFAULT; 971 972 return 0; 973 } 974 case XDP_OPTIONS: 975 { 976 struct xdp_options opts = {}; 977 978 if (len < sizeof(opts)) 979 return -EINVAL; 980 981 mutex_lock(&xs->mutex); 982 if (xs->zc) 983 opts.flags |= XDP_OPTIONS_ZEROCOPY; 984 mutex_unlock(&xs->mutex); 985 986 len = sizeof(opts); 987 if (copy_to_user(optval, &opts, len)) 988 return -EFAULT; 989 if (put_user(len, optlen)) 990 return -EFAULT; 991 992 return 0; 993 } 994 default: 995 break; 996 } 997 998 return -EOPNOTSUPP; 999 } 1000 1001 static int xsk_mmap(struct file *file, struct socket *sock, 1002 struct vm_area_struct *vma) 1003 { 1004 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT; 1005 unsigned long size = vma->vm_end - vma->vm_start; 1006 struct xdp_sock *xs = xdp_sk(sock->sk); 1007 struct xsk_queue *q = NULL; 1008 struct xdp_umem *umem; 1009 unsigned long pfn; 1010 struct page *qpg; 1011 1012 if (READ_ONCE(xs->state) != XSK_READY) 1013 return -EBUSY; 1014 1015 if (offset == XDP_PGOFF_RX_RING) { 1016 q = READ_ONCE(xs->rx); 1017 } else if (offset == XDP_PGOFF_TX_RING) { 1018 q = READ_ONCE(xs->tx); 1019 } else { 1020 umem = READ_ONCE(xs->umem); 1021 if (!umem) 1022 return -EINVAL; 1023 1024 /* Matches the smp_wmb() in XDP_UMEM_REG */ 1025 smp_rmb(); 1026 if (offset == XDP_UMEM_PGOFF_FILL_RING) 1027 q = READ_ONCE(umem->fq); 1028 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING) 1029 q = READ_ONCE(umem->cq); 1030 } 1031 1032 if (!q) 1033 return -EINVAL; 1034 1035 /* Matches the smp_wmb() in xsk_init_queue */ 1036 smp_rmb(); 1037 qpg = virt_to_head_page(q->ring); 1038 if (size > page_size(qpg)) 1039 return -EINVAL; 1040 1041 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT; 1042 return remap_pfn_range(vma, vma->vm_start, pfn, 1043 size, vma->vm_page_prot); 1044 } 1045 1046 static int xsk_notifier(struct notifier_block *this, 1047 unsigned long msg, void *ptr) 1048 { 1049 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 1050 struct net *net = dev_net(dev); 1051 struct sock *sk; 1052 1053 switch (msg) { 1054 case NETDEV_UNREGISTER: 1055 mutex_lock(&net->xdp.lock); 1056 sk_for_each(sk, &net->xdp.list) { 1057 struct xdp_sock *xs = xdp_sk(sk); 1058 1059 mutex_lock(&xs->mutex); 1060 if (xs->dev == dev) { 1061 sk->sk_err = ENETDOWN; 1062 if (!sock_flag(sk, SOCK_DEAD)) 1063 sk->sk_error_report(sk); 1064 1065 xsk_unbind_dev(xs); 1066 1067 /* Clear device references in umem. */ 1068 xdp_umem_clear_dev(xs->umem); 1069 } 1070 mutex_unlock(&xs->mutex); 1071 } 1072 mutex_unlock(&net->xdp.lock); 1073 break; 1074 } 1075 return NOTIFY_DONE; 1076 } 1077 1078 static struct proto xsk_proto = { 1079 .name = "XDP", 1080 .owner = THIS_MODULE, 1081 .obj_size = sizeof(struct xdp_sock), 1082 }; 1083 1084 static const struct proto_ops xsk_proto_ops = { 1085 .family = PF_XDP, 1086 .owner = THIS_MODULE, 1087 .release = xsk_release, 1088 .bind = xsk_bind, 1089 .connect = sock_no_connect, 1090 .socketpair = sock_no_socketpair, 1091 .accept = sock_no_accept, 1092 .getname = sock_no_getname, 1093 .poll = xsk_poll, 1094 .ioctl = sock_no_ioctl, 1095 .listen = sock_no_listen, 1096 .shutdown = sock_no_shutdown, 1097 .setsockopt = xsk_setsockopt, 1098 .getsockopt = xsk_getsockopt, 1099 .sendmsg = xsk_sendmsg, 1100 .recvmsg = sock_no_recvmsg, 1101 .mmap = xsk_mmap, 1102 .sendpage = sock_no_sendpage, 1103 }; 1104 1105 static void xsk_destruct(struct sock *sk) 1106 { 1107 struct xdp_sock *xs = xdp_sk(sk); 1108 1109 if (!sock_flag(sk, SOCK_DEAD)) 1110 return; 1111 1112 xdp_put_umem(xs->umem); 1113 1114 sk_refcnt_debug_dec(sk); 1115 } 1116 1117 static int xsk_create(struct net *net, struct socket *sock, int protocol, 1118 int kern) 1119 { 1120 struct sock *sk; 1121 struct xdp_sock *xs; 1122 1123 if (!ns_capable(net->user_ns, CAP_NET_RAW)) 1124 return -EPERM; 1125 if (sock->type != SOCK_RAW) 1126 return -ESOCKTNOSUPPORT; 1127 1128 if (protocol) 1129 return -EPROTONOSUPPORT; 1130 1131 sock->state = SS_UNCONNECTED; 1132 1133 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern); 1134 if (!sk) 1135 return -ENOBUFS; 1136 1137 sock->ops = &xsk_proto_ops; 1138 1139 sock_init_data(sock, sk); 1140 1141 sk->sk_family = PF_XDP; 1142 1143 sk->sk_destruct = xsk_destruct; 1144 sk_refcnt_debug_inc(sk); 1145 1146 sock_set_flag(sk, SOCK_RCU_FREE); 1147 1148 xs = xdp_sk(sk); 1149 xs->state = XSK_READY; 1150 mutex_init(&xs->mutex); 1151 spin_lock_init(&xs->rx_lock); 1152 spin_lock_init(&xs->tx_completion_lock); 1153 1154 INIT_LIST_HEAD(&xs->map_list); 1155 spin_lock_init(&xs->map_list_lock); 1156 1157 mutex_lock(&net->xdp.lock); 1158 sk_add_node_rcu(sk, &net->xdp.list); 1159 mutex_unlock(&net->xdp.lock); 1160 1161 local_bh_disable(); 1162 sock_prot_inuse_add(net, &xsk_proto, 1); 1163 local_bh_enable(); 1164 1165 return 0; 1166 } 1167 1168 static const struct net_proto_family xsk_family_ops = { 1169 .family = PF_XDP, 1170 .create = xsk_create, 1171 .owner = THIS_MODULE, 1172 }; 1173 1174 static struct notifier_block xsk_netdev_notifier = { 1175 .notifier_call = xsk_notifier, 1176 }; 1177 1178 static int __net_init xsk_net_init(struct net *net) 1179 { 1180 mutex_init(&net->xdp.lock); 1181 INIT_HLIST_HEAD(&net->xdp.list); 1182 return 0; 1183 } 1184 1185 static void __net_exit xsk_net_exit(struct net *net) 1186 { 1187 WARN_ON_ONCE(!hlist_empty(&net->xdp.list)); 1188 } 1189 1190 static struct pernet_operations xsk_net_ops = { 1191 .init = xsk_net_init, 1192 .exit = xsk_net_exit, 1193 }; 1194 1195 static int __init xsk_init(void) 1196 { 1197 int err, cpu; 1198 1199 err = proto_register(&xsk_proto, 0 /* no slab */); 1200 if (err) 1201 goto out; 1202 1203 err = sock_register(&xsk_family_ops); 1204 if (err) 1205 goto out_proto; 1206 1207 err = register_pernet_subsys(&xsk_net_ops); 1208 if (err) 1209 goto out_sk; 1210 1211 err = register_netdevice_notifier(&xsk_netdev_notifier); 1212 if (err) 1213 goto out_pernet; 1214 1215 for_each_possible_cpu(cpu) 1216 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu)); 1217 return 0; 1218 1219 out_pernet: 1220 unregister_pernet_subsys(&xsk_net_ops); 1221 out_sk: 1222 sock_unregister(PF_XDP); 1223 out_proto: 1224 proto_unregister(&xsk_proto); 1225 out: 1226 return err; 1227 } 1228 1229 fs_initcall(xsk_init); 1230