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 sock_def_readable(&xs->sk); 221 } 222 223 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 224 { 225 u32 metalen = xdp->data - xdp->data_meta; 226 u32 len = xdp->data_end - xdp->data; 227 u64 offset = xs->umem->headroom; 228 void *buffer; 229 u64 addr; 230 int err; 231 232 spin_lock_bh(&xs->rx_lock); 233 234 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) { 235 err = -EINVAL; 236 goto out_unlock; 237 } 238 239 if (!xskq_cons_peek_addr(xs->umem->fq, &addr, xs->umem) || 240 len > xs->umem->chunk_size_nohr - XDP_PACKET_HEADROOM) { 241 err = -ENOSPC; 242 goto out_drop; 243 } 244 245 addr = xsk_umem_adjust_offset(xs->umem, addr, offset); 246 buffer = xdp_umem_get_data(xs->umem, addr); 247 memcpy(buffer, xdp->data_meta, len + metalen); 248 249 addr = xsk_umem_adjust_offset(xs->umem, addr, metalen); 250 err = xskq_prod_reserve_desc(xs->rx, addr, len); 251 if (err) 252 goto out_drop; 253 254 xskq_cons_release(xs->umem->fq); 255 xskq_prod_submit(xs->rx); 256 257 spin_unlock_bh(&xs->rx_lock); 258 259 xs->sk.sk_data_ready(&xs->sk); 260 return 0; 261 262 out_drop: 263 xs->rx_dropped++; 264 out_unlock: 265 spin_unlock_bh(&xs->rx_lock); 266 return err; 267 } 268 269 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp) 270 { 271 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list); 272 int err; 273 274 err = xsk_rcv(xs, xdp); 275 if (err) 276 return err; 277 278 if (!xs->flush_node.prev) 279 list_add(&xs->flush_node, flush_list); 280 281 return 0; 282 } 283 284 void __xsk_map_flush(void) 285 { 286 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list); 287 struct xdp_sock *xs, *tmp; 288 289 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) { 290 xsk_flush(xs); 291 __list_del_clearprev(&xs->flush_node); 292 } 293 } 294 295 void xsk_umem_complete_tx(struct xdp_umem *umem, u32 nb_entries) 296 { 297 xskq_prod_submit_n(umem->cq, nb_entries); 298 } 299 EXPORT_SYMBOL(xsk_umem_complete_tx); 300 301 void xsk_umem_consume_tx_done(struct xdp_umem *umem) 302 { 303 struct xdp_sock *xs; 304 305 rcu_read_lock(); 306 list_for_each_entry_rcu(xs, &umem->xsk_list, list) { 307 xs->sk.sk_write_space(&xs->sk); 308 } 309 rcu_read_unlock(); 310 } 311 EXPORT_SYMBOL(xsk_umem_consume_tx_done); 312 313 bool xsk_umem_consume_tx(struct xdp_umem *umem, struct xdp_desc *desc) 314 { 315 struct xdp_sock *xs; 316 317 rcu_read_lock(); 318 list_for_each_entry_rcu(xs, &umem->xsk_list, list) { 319 if (!xskq_cons_peek_desc(xs->tx, desc, umem)) 320 continue; 321 322 /* This is the backpreassure mechanism for the Tx path. 323 * Reserve space in the completion queue and only proceed 324 * if there is space in it. This avoids having to implement 325 * any buffering in the Tx path. 326 */ 327 if (xskq_prod_reserve_addr(umem->cq, desc->addr)) 328 goto out; 329 330 xskq_cons_release(xs->tx); 331 rcu_read_unlock(); 332 return true; 333 } 334 335 out: 336 rcu_read_unlock(); 337 return false; 338 } 339 EXPORT_SYMBOL(xsk_umem_consume_tx); 340 341 static int xsk_wakeup(struct xdp_sock *xs, u8 flags) 342 { 343 struct net_device *dev = xs->dev; 344 int err; 345 346 rcu_read_lock(); 347 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags); 348 rcu_read_unlock(); 349 350 return err; 351 } 352 353 static int xsk_zc_xmit(struct xdp_sock *xs) 354 { 355 return xsk_wakeup(xs, XDP_WAKEUP_TX); 356 } 357 358 static void xsk_destruct_skb(struct sk_buff *skb) 359 { 360 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg; 361 struct xdp_sock *xs = xdp_sk(skb->sk); 362 unsigned long flags; 363 364 spin_lock_irqsave(&xs->tx_completion_lock, flags); 365 xskq_prod_submit_addr(xs->umem->cq, addr); 366 spin_unlock_irqrestore(&xs->tx_completion_lock, flags); 367 368 sock_wfree(skb); 369 } 370 371 static int xsk_generic_xmit(struct sock *sk) 372 { 373 struct xdp_sock *xs = xdp_sk(sk); 374 u32 max_batch = TX_BATCH_SIZE; 375 bool sent_frame = false; 376 struct xdp_desc desc; 377 struct sk_buff *skb; 378 int err = 0; 379 380 mutex_lock(&xs->mutex); 381 382 if (xs->queue_id >= xs->dev->real_num_tx_queues) 383 goto out; 384 385 while (xskq_cons_peek_desc(xs->tx, &desc, xs->umem)) { 386 char *buffer; 387 u64 addr; 388 u32 len; 389 390 if (max_batch-- == 0) { 391 err = -EAGAIN; 392 goto out; 393 } 394 395 len = desc.len; 396 skb = sock_alloc_send_skb(sk, len, 1, &err); 397 if (unlikely(!skb)) { 398 err = -EAGAIN; 399 goto out; 400 } 401 402 skb_put(skb, len); 403 addr = desc.addr; 404 buffer = xdp_umem_get_data(xs->umem, addr); 405 err = skb_store_bits(skb, 0, buffer, len); 406 /* This is the backpreassure mechanism for the Tx path. 407 * Reserve space in the completion queue and only proceed 408 * if there is space in it. This avoids having to implement 409 * any buffering in the Tx path. 410 */ 411 if (unlikely(err) || xskq_prod_reserve(xs->umem->cq)) { 412 kfree_skb(skb); 413 goto out; 414 } 415 416 skb->dev = xs->dev; 417 skb->priority = sk->sk_priority; 418 skb->mark = sk->sk_mark; 419 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr; 420 skb->destructor = xsk_destruct_skb; 421 422 err = dev_direct_xmit(skb, xs->queue_id); 423 xskq_cons_release(xs->tx); 424 /* Ignore NET_XMIT_CN as packet might have been sent */ 425 if (err == NET_XMIT_DROP || err == NETDEV_TX_BUSY) { 426 /* SKB completed but not sent */ 427 err = -EBUSY; 428 goto out; 429 } 430 431 sent_frame = true; 432 } 433 434 out: 435 if (sent_frame) 436 sk->sk_write_space(sk); 437 438 mutex_unlock(&xs->mutex); 439 return err; 440 } 441 442 static int __xsk_sendmsg(struct sock *sk) 443 { 444 struct xdp_sock *xs = xdp_sk(sk); 445 446 if (unlikely(!(xs->dev->flags & IFF_UP))) 447 return -ENETDOWN; 448 if (unlikely(!xs->tx)) 449 return -ENOBUFS; 450 451 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk); 452 } 453 454 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) 455 { 456 bool need_wait = !(m->msg_flags & MSG_DONTWAIT); 457 struct sock *sk = sock->sk; 458 struct xdp_sock *xs = xdp_sk(sk); 459 460 if (unlikely(!xsk_is_bound(xs))) 461 return -ENXIO; 462 if (unlikely(need_wait)) 463 return -EOPNOTSUPP; 464 465 return __xsk_sendmsg(sk); 466 } 467 468 static __poll_t xsk_poll(struct file *file, struct socket *sock, 469 struct poll_table_struct *wait) 470 { 471 __poll_t mask = datagram_poll(file, sock, wait); 472 struct sock *sk = sock->sk; 473 struct xdp_sock *xs = xdp_sk(sk); 474 struct xdp_umem *umem; 475 476 if (unlikely(!xsk_is_bound(xs))) 477 return mask; 478 479 umem = xs->umem; 480 481 if (umem->need_wakeup) { 482 if (xs->zc) 483 xsk_wakeup(xs, umem->need_wakeup); 484 else 485 /* Poll needs to drive Tx also in copy mode */ 486 __xsk_sendmsg(sk); 487 } 488 489 if (xs->rx && !xskq_prod_is_empty(xs->rx)) 490 mask |= EPOLLIN | EPOLLRDNORM; 491 if (xs->tx && !xskq_cons_is_full(xs->tx)) 492 mask |= EPOLLOUT | EPOLLWRNORM; 493 494 return mask; 495 } 496 497 static int xsk_init_queue(u32 entries, struct xsk_queue **queue, 498 bool umem_queue) 499 { 500 struct xsk_queue *q; 501 502 if (entries == 0 || *queue || !is_power_of_2(entries)) 503 return -EINVAL; 504 505 q = xskq_create(entries, umem_queue); 506 if (!q) 507 return -ENOMEM; 508 509 /* Make sure queue is ready before it can be seen by others */ 510 smp_wmb(); 511 WRITE_ONCE(*queue, q); 512 return 0; 513 } 514 515 static void xsk_unbind_dev(struct xdp_sock *xs) 516 { 517 struct net_device *dev = xs->dev; 518 519 if (xs->state != XSK_BOUND) 520 return; 521 WRITE_ONCE(xs->state, XSK_UNBOUND); 522 523 /* Wait for driver to stop using the xdp socket. */ 524 xdp_del_sk_umem(xs->umem, xs); 525 xs->dev = NULL; 526 synchronize_net(); 527 dev_put(dev); 528 } 529 530 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs, 531 struct xdp_sock ***map_entry) 532 { 533 struct xsk_map *map = NULL; 534 struct xsk_map_node *node; 535 536 *map_entry = NULL; 537 538 spin_lock_bh(&xs->map_list_lock); 539 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node, 540 node); 541 if (node) { 542 WARN_ON(xsk_map_inc(node->map)); 543 map = node->map; 544 *map_entry = node->map_entry; 545 } 546 spin_unlock_bh(&xs->map_list_lock); 547 return map; 548 } 549 550 static void xsk_delete_from_maps(struct xdp_sock *xs) 551 { 552 /* This function removes the current XDP socket from all the 553 * maps it resides in. We need to take extra care here, due to 554 * the two locks involved. Each map has a lock synchronizing 555 * updates to the entries, and each socket has a lock that 556 * synchronizes access to the list of maps (map_list). For 557 * deadlock avoidance the locks need to be taken in the order 558 * "map lock"->"socket map list lock". We start off by 559 * accessing the socket map list, and take a reference to the 560 * map to guarantee existence between the 561 * xsk_get_map_list_entry() and xsk_map_try_sock_delete() 562 * calls. Then we ask the map to remove the socket, which 563 * tries to remove the socket from the map. Note that there 564 * might be updates to the map between 565 * xsk_get_map_list_entry() and xsk_map_try_sock_delete(). 566 */ 567 struct xdp_sock **map_entry = NULL; 568 struct xsk_map *map; 569 570 while ((map = xsk_get_map_list_entry(xs, &map_entry))) { 571 xsk_map_try_sock_delete(map, xs, map_entry); 572 xsk_map_put(map); 573 } 574 } 575 576 static int xsk_release(struct socket *sock) 577 { 578 struct sock *sk = sock->sk; 579 struct xdp_sock *xs = xdp_sk(sk); 580 struct net *net; 581 582 if (!sk) 583 return 0; 584 585 net = sock_net(sk); 586 587 mutex_lock(&net->xdp.lock); 588 sk_del_node_init_rcu(sk); 589 mutex_unlock(&net->xdp.lock); 590 591 local_bh_disable(); 592 sock_prot_inuse_add(net, sk->sk_prot, -1); 593 local_bh_enable(); 594 595 xsk_delete_from_maps(xs); 596 mutex_lock(&xs->mutex); 597 xsk_unbind_dev(xs); 598 mutex_unlock(&xs->mutex); 599 600 xskq_destroy(xs->rx); 601 xskq_destroy(xs->tx); 602 603 sock_orphan(sk); 604 sock->sk = NULL; 605 606 sk_refcnt_debug_release(sk); 607 sock_put(sk); 608 609 return 0; 610 } 611 612 static struct socket *xsk_lookup_xsk_from_fd(int fd) 613 { 614 struct socket *sock; 615 int err; 616 617 sock = sockfd_lookup(fd, &err); 618 if (!sock) 619 return ERR_PTR(-ENOTSOCK); 620 621 if (sock->sk->sk_family != PF_XDP) { 622 sockfd_put(sock); 623 return ERR_PTR(-ENOPROTOOPT); 624 } 625 626 return sock; 627 } 628 629 /* Check if umem pages are contiguous. 630 * If zero-copy mode, use the DMA address to do the page contiguity check 631 * For all other modes we use addr (kernel virtual address) 632 * Store the result in the low bits of addr. 633 */ 634 static void xsk_check_page_contiguity(struct xdp_umem *umem, u32 flags) 635 { 636 struct xdp_umem_page *pgs = umem->pages; 637 int i, is_contig; 638 639 for (i = 0; i < umem->npgs - 1; i++) { 640 is_contig = (flags & XDP_ZEROCOPY) ? 641 (pgs[i].dma + PAGE_SIZE == pgs[i + 1].dma) : 642 (pgs[i].addr + PAGE_SIZE == pgs[i + 1].addr); 643 pgs[i].addr += is_contig << XSK_NEXT_PG_CONTIG_SHIFT; 644 } 645 } 646 647 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len) 648 { 649 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr; 650 struct sock *sk = sock->sk; 651 struct xdp_sock *xs = xdp_sk(sk); 652 struct net_device *dev; 653 u32 flags, qid; 654 int err = 0; 655 656 if (addr_len < sizeof(struct sockaddr_xdp)) 657 return -EINVAL; 658 if (sxdp->sxdp_family != AF_XDP) 659 return -EINVAL; 660 661 flags = sxdp->sxdp_flags; 662 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY | 663 XDP_USE_NEED_WAKEUP)) 664 return -EINVAL; 665 666 rtnl_lock(); 667 mutex_lock(&xs->mutex); 668 if (xs->state != XSK_READY) { 669 err = -EBUSY; 670 goto out_release; 671 } 672 673 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex); 674 if (!dev) { 675 err = -ENODEV; 676 goto out_release; 677 } 678 679 if (!xs->rx && !xs->tx) { 680 err = -EINVAL; 681 goto out_unlock; 682 } 683 684 qid = sxdp->sxdp_queue_id; 685 686 if (flags & XDP_SHARED_UMEM) { 687 struct xdp_sock *umem_xs; 688 struct socket *sock; 689 690 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) || 691 (flags & XDP_USE_NEED_WAKEUP)) { 692 /* Cannot specify flags for shared sockets. */ 693 err = -EINVAL; 694 goto out_unlock; 695 } 696 697 if (xs->umem) { 698 /* We have already our own. */ 699 err = -EINVAL; 700 goto out_unlock; 701 } 702 703 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd); 704 if (IS_ERR(sock)) { 705 err = PTR_ERR(sock); 706 goto out_unlock; 707 } 708 709 umem_xs = xdp_sk(sock->sk); 710 if (!xsk_is_bound(umem_xs)) { 711 err = -EBADF; 712 sockfd_put(sock); 713 goto out_unlock; 714 } 715 if (umem_xs->dev != dev || umem_xs->queue_id != qid) { 716 err = -EINVAL; 717 sockfd_put(sock); 718 goto out_unlock; 719 } 720 721 xdp_get_umem(umem_xs->umem); 722 WRITE_ONCE(xs->umem, umem_xs->umem); 723 sockfd_put(sock); 724 } else if (!xs->umem || !xdp_umem_validate_queues(xs->umem)) { 725 err = -EINVAL; 726 goto out_unlock; 727 } else { 728 /* This xsk has its own umem. */ 729 xskq_set_umem(xs->umem->fq, xs->umem->size, 730 xs->umem->chunk_mask); 731 xskq_set_umem(xs->umem->cq, xs->umem->size, 732 xs->umem->chunk_mask); 733 734 err = xdp_umem_assign_dev(xs->umem, dev, qid, flags); 735 if (err) 736 goto out_unlock; 737 738 xsk_check_page_contiguity(xs->umem, flags); 739 } 740 741 xs->dev = dev; 742 xs->zc = xs->umem->zc; 743 xs->queue_id = qid; 744 xskq_set_umem(xs->rx, xs->umem->size, xs->umem->chunk_mask); 745 xskq_set_umem(xs->tx, xs->umem->size, xs->umem->chunk_mask); 746 xdp_add_sk_umem(xs->umem, xs); 747 748 out_unlock: 749 if (err) { 750 dev_put(dev); 751 } else { 752 /* Matches smp_rmb() in bind() for shared umem 753 * sockets, and xsk_is_bound(). 754 */ 755 smp_wmb(); 756 WRITE_ONCE(xs->state, XSK_BOUND); 757 } 758 out_release: 759 mutex_unlock(&xs->mutex); 760 rtnl_unlock(); 761 return err; 762 } 763 764 struct xdp_umem_reg_v1 { 765 __u64 addr; /* Start of packet data area */ 766 __u64 len; /* Length of packet data area */ 767 __u32 chunk_size; 768 __u32 headroom; 769 }; 770 771 static int xsk_setsockopt(struct socket *sock, int level, int optname, 772 char __user *optval, unsigned int optlen) 773 { 774 struct sock *sk = sock->sk; 775 struct xdp_sock *xs = xdp_sk(sk); 776 int err; 777 778 if (level != SOL_XDP) 779 return -ENOPROTOOPT; 780 781 switch (optname) { 782 case XDP_RX_RING: 783 case XDP_TX_RING: 784 { 785 struct xsk_queue **q; 786 int entries; 787 788 if (optlen < sizeof(entries)) 789 return -EINVAL; 790 if (copy_from_user(&entries, optval, sizeof(entries))) 791 return -EFAULT; 792 793 mutex_lock(&xs->mutex); 794 if (xs->state != XSK_READY) { 795 mutex_unlock(&xs->mutex); 796 return -EBUSY; 797 } 798 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx; 799 err = xsk_init_queue(entries, q, false); 800 if (!err && optname == XDP_TX_RING) 801 /* Tx needs to be explicitly woken up the first time */ 802 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 803 mutex_unlock(&xs->mutex); 804 return err; 805 } 806 case XDP_UMEM_REG: 807 { 808 size_t mr_size = sizeof(struct xdp_umem_reg); 809 struct xdp_umem_reg mr = {}; 810 struct xdp_umem *umem; 811 812 if (optlen < sizeof(struct xdp_umem_reg_v1)) 813 return -EINVAL; 814 else if (optlen < sizeof(mr)) 815 mr_size = sizeof(struct xdp_umem_reg_v1); 816 817 if (copy_from_user(&mr, optval, mr_size)) 818 return -EFAULT; 819 820 mutex_lock(&xs->mutex); 821 if (xs->state != XSK_READY || xs->umem) { 822 mutex_unlock(&xs->mutex); 823 return -EBUSY; 824 } 825 826 umem = xdp_umem_create(&mr); 827 if (IS_ERR(umem)) { 828 mutex_unlock(&xs->mutex); 829 return PTR_ERR(umem); 830 } 831 832 /* Make sure umem is ready before it can be seen by others */ 833 smp_wmb(); 834 WRITE_ONCE(xs->umem, umem); 835 mutex_unlock(&xs->mutex); 836 return 0; 837 } 838 case XDP_UMEM_FILL_RING: 839 case XDP_UMEM_COMPLETION_RING: 840 { 841 struct xsk_queue **q; 842 int entries; 843 844 if (copy_from_user(&entries, optval, sizeof(entries))) 845 return -EFAULT; 846 847 mutex_lock(&xs->mutex); 848 if (xs->state != XSK_READY) { 849 mutex_unlock(&xs->mutex); 850 return -EBUSY; 851 } 852 if (!xs->umem) { 853 mutex_unlock(&xs->mutex); 854 return -EINVAL; 855 } 856 857 q = (optname == XDP_UMEM_FILL_RING) ? &xs->umem->fq : 858 &xs->umem->cq; 859 err = xsk_init_queue(entries, q, true); 860 mutex_unlock(&xs->mutex); 861 return err; 862 } 863 default: 864 break; 865 } 866 867 return -ENOPROTOOPT; 868 } 869 870 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring) 871 { 872 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer); 873 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer); 874 ring->desc = offsetof(struct xdp_rxtx_ring, desc); 875 } 876 877 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring) 878 { 879 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer); 880 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer); 881 ring->desc = offsetof(struct xdp_umem_ring, desc); 882 } 883 884 static int xsk_getsockopt(struct socket *sock, int level, int optname, 885 char __user *optval, int __user *optlen) 886 { 887 struct sock *sk = sock->sk; 888 struct xdp_sock *xs = xdp_sk(sk); 889 int len; 890 891 if (level != SOL_XDP) 892 return -ENOPROTOOPT; 893 894 if (get_user(len, optlen)) 895 return -EFAULT; 896 if (len < 0) 897 return -EINVAL; 898 899 switch (optname) { 900 case XDP_STATISTICS: 901 { 902 struct xdp_statistics stats; 903 904 if (len < sizeof(stats)) 905 return -EINVAL; 906 907 mutex_lock(&xs->mutex); 908 stats.rx_dropped = xs->rx_dropped; 909 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx); 910 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx); 911 mutex_unlock(&xs->mutex); 912 913 if (copy_to_user(optval, &stats, sizeof(stats))) 914 return -EFAULT; 915 if (put_user(sizeof(stats), optlen)) 916 return -EFAULT; 917 918 return 0; 919 } 920 case XDP_MMAP_OFFSETS: 921 { 922 struct xdp_mmap_offsets off; 923 struct xdp_mmap_offsets_v1 off_v1; 924 bool flags_supported = true; 925 void *to_copy; 926 927 if (len < sizeof(off_v1)) 928 return -EINVAL; 929 else if (len < sizeof(off)) 930 flags_supported = false; 931 932 if (flags_supported) { 933 /* xdp_ring_offset is identical to xdp_ring_offset_v1 934 * except for the flags field added to the end. 935 */ 936 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 937 &off.rx); 938 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 939 &off.tx); 940 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 941 &off.fr); 942 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 943 &off.cr); 944 off.rx.flags = offsetof(struct xdp_rxtx_ring, 945 ptrs.flags); 946 off.tx.flags = offsetof(struct xdp_rxtx_ring, 947 ptrs.flags); 948 off.fr.flags = offsetof(struct xdp_umem_ring, 949 ptrs.flags); 950 off.cr.flags = offsetof(struct xdp_umem_ring, 951 ptrs.flags); 952 953 len = sizeof(off); 954 to_copy = &off; 955 } else { 956 xsk_enter_rxtx_offsets(&off_v1.rx); 957 xsk_enter_rxtx_offsets(&off_v1.tx); 958 xsk_enter_umem_offsets(&off_v1.fr); 959 xsk_enter_umem_offsets(&off_v1.cr); 960 961 len = sizeof(off_v1); 962 to_copy = &off_v1; 963 } 964 965 if (copy_to_user(optval, to_copy, len)) 966 return -EFAULT; 967 if (put_user(len, optlen)) 968 return -EFAULT; 969 970 return 0; 971 } 972 case XDP_OPTIONS: 973 { 974 struct xdp_options opts = {}; 975 976 if (len < sizeof(opts)) 977 return -EINVAL; 978 979 mutex_lock(&xs->mutex); 980 if (xs->zc) 981 opts.flags |= XDP_OPTIONS_ZEROCOPY; 982 mutex_unlock(&xs->mutex); 983 984 len = sizeof(opts); 985 if (copy_to_user(optval, &opts, len)) 986 return -EFAULT; 987 if (put_user(len, optlen)) 988 return -EFAULT; 989 990 return 0; 991 } 992 default: 993 break; 994 } 995 996 return -EOPNOTSUPP; 997 } 998 999 static int xsk_mmap(struct file *file, struct socket *sock, 1000 struct vm_area_struct *vma) 1001 { 1002 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT; 1003 unsigned long size = vma->vm_end - vma->vm_start; 1004 struct xdp_sock *xs = xdp_sk(sock->sk); 1005 struct xsk_queue *q = NULL; 1006 struct xdp_umem *umem; 1007 unsigned long pfn; 1008 struct page *qpg; 1009 1010 if (READ_ONCE(xs->state) != XSK_READY) 1011 return -EBUSY; 1012 1013 if (offset == XDP_PGOFF_RX_RING) { 1014 q = READ_ONCE(xs->rx); 1015 } else if (offset == XDP_PGOFF_TX_RING) { 1016 q = READ_ONCE(xs->tx); 1017 } else { 1018 umem = READ_ONCE(xs->umem); 1019 if (!umem) 1020 return -EINVAL; 1021 1022 /* Matches the smp_wmb() in XDP_UMEM_REG */ 1023 smp_rmb(); 1024 if (offset == XDP_UMEM_PGOFF_FILL_RING) 1025 q = READ_ONCE(umem->fq); 1026 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING) 1027 q = READ_ONCE(umem->cq); 1028 } 1029 1030 if (!q) 1031 return -EINVAL; 1032 1033 /* Matches the smp_wmb() in xsk_init_queue */ 1034 smp_rmb(); 1035 qpg = virt_to_head_page(q->ring); 1036 if (size > page_size(qpg)) 1037 return -EINVAL; 1038 1039 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT; 1040 return remap_pfn_range(vma, vma->vm_start, pfn, 1041 size, vma->vm_page_prot); 1042 } 1043 1044 static int xsk_notifier(struct notifier_block *this, 1045 unsigned long msg, void *ptr) 1046 { 1047 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 1048 struct net *net = dev_net(dev); 1049 struct sock *sk; 1050 1051 switch (msg) { 1052 case NETDEV_UNREGISTER: 1053 mutex_lock(&net->xdp.lock); 1054 sk_for_each(sk, &net->xdp.list) { 1055 struct xdp_sock *xs = xdp_sk(sk); 1056 1057 mutex_lock(&xs->mutex); 1058 if (xs->dev == dev) { 1059 sk->sk_err = ENETDOWN; 1060 if (!sock_flag(sk, SOCK_DEAD)) 1061 sk->sk_error_report(sk); 1062 1063 xsk_unbind_dev(xs); 1064 1065 /* Clear device references in umem. */ 1066 xdp_umem_clear_dev(xs->umem); 1067 } 1068 mutex_unlock(&xs->mutex); 1069 } 1070 mutex_unlock(&net->xdp.lock); 1071 break; 1072 } 1073 return NOTIFY_DONE; 1074 } 1075 1076 static struct proto xsk_proto = { 1077 .name = "XDP", 1078 .owner = THIS_MODULE, 1079 .obj_size = sizeof(struct xdp_sock), 1080 }; 1081 1082 static const struct proto_ops xsk_proto_ops = { 1083 .family = PF_XDP, 1084 .owner = THIS_MODULE, 1085 .release = xsk_release, 1086 .bind = xsk_bind, 1087 .connect = sock_no_connect, 1088 .socketpair = sock_no_socketpair, 1089 .accept = sock_no_accept, 1090 .getname = sock_no_getname, 1091 .poll = xsk_poll, 1092 .ioctl = sock_no_ioctl, 1093 .listen = sock_no_listen, 1094 .shutdown = sock_no_shutdown, 1095 .setsockopt = xsk_setsockopt, 1096 .getsockopt = xsk_getsockopt, 1097 .sendmsg = xsk_sendmsg, 1098 .recvmsg = sock_no_recvmsg, 1099 .mmap = xsk_mmap, 1100 .sendpage = sock_no_sendpage, 1101 }; 1102 1103 static void xsk_destruct(struct sock *sk) 1104 { 1105 struct xdp_sock *xs = xdp_sk(sk); 1106 1107 if (!sock_flag(sk, SOCK_DEAD)) 1108 return; 1109 1110 xdp_put_umem(xs->umem); 1111 1112 sk_refcnt_debug_dec(sk); 1113 } 1114 1115 static int xsk_create(struct net *net, struct socket *sock, int protocol, 1116 int kern) 1117 { 1118 struct sock *sk; 1119 struct xdp_sock *xs; 1120 1121 if (!ns_capable(net->user_ns, CAP_NET_RAW)) 1122 return -EPERM; 1123 if (sock->type != SOCK_RAW) 1124 return -ESOCKTNOSUPPORT; 1125 1126 if (protocol) 1127 return -EPROTONOSUPPORT; 1128 1129 sock->state = SS_UNCONNECTED; 1130 1131 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern); 1132 if (!sk) 1133 return -ENOBUFS; 1134 1135 sock->ops = &xsk_proto_ops; 1136 1137 sock_init_data(sock, sk); 1138 1139 sk->sk_family = PF_XDP; 1140 1141 sk->sk_destruct = xsk_destruct; 1142 sk_refcnt_debug_inc(sk); 1143 1144 sock_set_flag(sk, SOCK_RCU_FREE); 1145 1146 xs = xdp_sk(sk); 1147 xs->state = XSK_READY; 1148 mutex_init(&xs->mutex); 1149 spin_lock_init(&xs->rx_lock); 1150 spin_lock_init(&xs->tx_completion_lock); 1151 1152 INIT_LIST_HEAD(&xs->map_list); 1153 spin_lock_init(&xs->map_list_lock); 1154 1155 mutex_lock(&net->xdp.lock); 1156 sk_add_node_rcu(sk, &net->xdp.list); 1157 mutex_unlock(&net->xdp.lock); 1158 1159 local_bh_disable(); 1160 sock_prot_inuse_add(net, &xsk_proto, 1); 1161 local_bh_enable(); 1162 1163 return 0; 1164 } 1165 1166 static const struct net_proto_family xsk_family_ops = { 1167 .family = PF_XDP, 1168 .create = xsk_create, 1169 .owner = THIS_MODULE, 1170 }; 1171 1172 static struct notifier_block xsk_netdev_notifier = { 1173 .notifier_call = xsk_notifier, 1174 }; 1175 1176 static int __net_init xsk_net_init(struct net *net) 1177 { 1178 mutex_init(&net->xdp.lock); 1179 INIT_HLIST_HEAD(&net->xdp.list); 1180 return 0; 1181 } 1182 1183 static void __net_exit xsk_net_exit(struct net *net) 1184 { 1185 WARN_ON_ONCE(!hlist_empty(&net->xdp.list)); 1186 } 1187 1188 static struct pernet_operations xsk_net_ops = { 1189 .init = xsk_net_init, 1190 .exit = xsk_net_exit, 1191 }; 1192 1193 static int __init xsk_init(void) 1194 { 1195 int err, cpu; 1196 1197 err = proto_register(&xsk_proto, 0 /* no slab */); 1198 if (err) 1199 goto out; 1200 1201 err = sock_register(&xsk_family_ops); 1202 if (err) 1203 goto out_proto; 1204 1205 err = register_pernet_subsys(&xsk_net_ops); 1206 if (err) 1207 goto out_sk; 1208 1209 err = register_netdevice_notifier(&xsk_netdev_notifier); 1210 if (err) 1211 goto out_pernet; 1212 1213 for_each_possible_cpu(cpu) 1214 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu)); 1215 return 0; 1216 1217 out_pernet: 1218 unregister_pernet_subsys(&xsk_net_ops); 1219 out_sk: 1220 sock_unregister(PF_XDP); 1221 out_proto: 1222 proto_unregister(&xsk_proto); 1223 out: 1224 return err; 1225 } 1226 1227 fs_initcall(xsk_init); 1228