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/busy_poll.h> 27 #include <net/xdp.h> 28 29 #include "xsk_queue.h" 30 #include "xdp_umem.h" 31 #include "xsk.h" 32 33 #define TX_BATCH_SIZE 16 34 35 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list); 36 37 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool) 38 { 39 if (pool->cached_need_wakeup & XDP_WAKEUP_RX) 40 return; 41 42 pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP; 43 pool->cached_need_wakeup |= XDP_WAKEUP_RX; 44 } 45 EXPORT_SYMBOL(xsk_set_rx_need_wakeup); 46 47 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool) 48 { 49 struct xdp_sock *xs; 50 51 if (pool->cached_need_wakeup & XDP_WAKEUP_TX) 52 return; 53 54 rcu_read_lock(); 55 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) { 56 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 57 } 58 rcu_read_unlock(); 59 60 pool->cached_need_wakeup |= XDP_WAKEUP_TX; 61 } 62 EXPORT_SYMBOL(xsk_set_tx_need_wakeup); 63 64 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool) 65 { 66 if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX)) 67 return; 68 69 pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP; 70 pool->cached_need_wakeup &= ~XDP_WAKEUP_RX; 71 } 72 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup); 73 74 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool) 75 { 76 struct xdp_sock *xs; 77 78 if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX)) 79 return; 80 81 rcu_read_lock(); 82 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) { 83 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP; 84 } 85 rcu_read_unlock(); 86 87 pool->cached_need_wakeup &= ~XDP_WAKEUP_TX; 88 } 89 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup); 90 91 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool) 92 { 93 return pool->uses_need_wakeup; 94 } 95 EXPORT_SYMBOL(xsk_uses_need_wakeup); 96 97 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev, 98 u16 queue_id) 99 { 100 if (queue_id < dev->real_num_rx_queues) 101 return dev->_rx[queue_id].pool; 102 if (queue_id < dev->real_num_tx_queues) 103 return dev->_tx[queue_id].pool; 104 105 return NULL; 106 } 107 EXPORT_SYMBOL(xsk_get_pool_from_qid); 108 109 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id) 110 { 111 if (queue_id < dev->num_rx_queues) 112 dev->_rx[queue_id].pool = NULL; 113 if (queue_id < dev->num_tx_queues) 114 dev->_tx[queue_id].pool = NULL; 115 } 116 117 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do 118 * not know if the device has more tx queues than rx, or the opposite. 119 * This might also change during run time. 120 */ 121 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool, 122 u16 queue_id) 123 { 124 if (queue_id >= max_t(unsigned int, 125 dev->real_num_rx_queues, 126 dev->real_num_tx_queues)) 127 return -EINVAL; 128 129 if (queue_id < dev->real_num_rx_queues) 130 dev->_rx[queue_id].pool = pool; 131 if (queue_id < dev->real_num_tx_queues) 132 dev->_tx[queue_id].pool = pool; 133 134 return 0; 135 } 136 137 void xp_release(struct xdp_buff_xsk *xskb) 138 { 139 xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb; 140 } 141 142 static u64 xp_get_handle(struct xdp_buff_xsk *xskb) 143 { 144 u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start; 145 146 offset += xskb->pool->headroom; 147 if (!xskb->pool->unaligned) 148 return xskb->orig_addr + offset; 149 return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT); 150 } 151 152 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 153 { 154 struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp); 155 u64 addr; 156 int err; 157 158 addr = xp_get_handle(xskb); 159 err = xskq_prod_reserve_desc(xs->rx, addr, len); 160 if (err) { 161 xs->rx_queue_full++; 162 return err; 163 } 164 165 xp_release(xskb); 166 return 0; 167 } 168 169 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len) 170 { 171 void *from_buf, *to_buf; 172 u32 metalen; 173 174 if (unlikely(xdp_data_meta_unsupported(from))) { 175 from_buf = from->data; 176 to_buf = to->data; 177 metalen = 0; 178 } else { 179 from_buf = from->data_meta; 180 metalen = from->data - from->data_meta; 181 to_buf = to->data - metalen; 182 } 183 184 memcpy(to_buf, from_buf, len + metalen); 185 } 186 187 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 188 { 189 struct xdp_buff *xsk_xdp; 190 int err; 191 u32 len; 192 193 len = xdp->data_end - xdp->data; 194 if (len > xsk_pool_get_rx_frame_size(xs->pool)) { 195 xs->rx_dropped++; 196 return -ENOSPC; 197 } 198 199 xsk_xdp = xsk_buff_alloc(xs->pool); 200 if (!xsk_xdp) { 201 xs->rx_dropped++; 202 return -ENOSPC; 203 } 204 205 xsk_copy_xdp(xsk_xdp, xdp, len); 206 err = __xsk_rcv_zc(xs, xsk_xdp, len); 207 if (err) { 208 xsk_buff_free(xsk_xdp); 209 return err; 210 } 211 return 0; 212 } 213 214 static bool xsk_tx_writeable(struct xdp_sock *xs) 215 { 216 if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2) 217 return false; 218 219 return true; 220 } 221 222 static bool xsk_is_bound(struct xdp_sock *xs) 223 { 224 if (READ_ONCE(xs->state) == XSK_BOUND) { 225 /* Matches smp_wmb() in bind(). */ 226 smp_rmb(); 227 return true; 228 } 229 return false; 230 } 231 232 static int xsk_rcv_check(struct xdp_sock *xs, struct xdp_buff *xdp) 233 { 234 if (!xsk_is_bound(xs)) 235 return -EINVAL; 236 237 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) 238 return -EINVAL; 239 240 sk_mark_napi_id_once_xdp(&xs->sk, xdp); 241 return 0; 242 } 243 244 static void xsk_flush(struct xdp_sock *xs) 245 { 246 xskq_prod_submit(xs->rx); 247 __xskq_cons_release(xs->pool->fq); 248 sock_def_readable(&xs->sk); 249 } 250 251 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 252 { 253 int err; 254 255 spin_lock_bh(&xs->rx_lock); 256 err = xsk_rcv_check(xs, xdp); 257 if (!err) { 258 err = __xsk_rcv(xs, xdp); 259 xsk_flush(xs); 260 } 261 spin_unlock_bh(&xs->rx_lock); 262 return err; 263 } 264 265 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 266 { 267 int err; 268 u32 len; 269 270 err = xsk_rcv_check(xs, xdp); 271 if (err) 272 return err; 273 274 if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) { 275 len = xdp->data_end - xdp->data; 276 return __xsk_rcv_zc(xs, xdp, len); 277 } 278 279 err = __xsk_rcv(xs, xdp); 280 if (!err) 281 xdp_return_buff(xdp); 282 return err; 283 } 284 285 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp) 286 { 287 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list); 288 int err; 289 290 err = xsk_rcv(xs, xdp); 291 if (err) 292 return err; 293 294 if (!xs->flush_node.prev) 295 list_add(&xs->flush_node, flush_list); 296 297 return 0; 298 } 299 300 void __xsk_map_flush(void) 301 { 302 struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list); 303 struct xdp_sock *xs, *tmp; 304 305 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) { 306 xsk_flush(xs); 307 __list_del_clearprev(&xs->flush_node); 308 } 309 } 310 311 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries) 312 { 313 xskq_prod_submit_n(pool->cq, nb_entries); 314 } 315 EXPORT_SYMBOL(xsk_tx_completed); 316 317 void xsk_tx_release(struct xsk_buff_pool *pool) 318 { 319 struct xdp_sock *xs; 320 321 rcu_read_lock(); 322 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) { 323 __xskq_cons_release(xs->tx); 324 if (xsk_tx_writeable(xs)) 325 xs->sk.sk_write_space(&xs->sk); 326 } 327 rcu_read_unlock(); 328 } 329 EXPORT_SYMBOL(xsk_tx_release); 330 331 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc) 332 { 333 struct xdp_sock *xs; 334 335 rcu_read_lock(); 336 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) { 337 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) { 338 xs->tx->queue_empty_descs++; 339 continue; 340 } 341 342 /* This is the backpressure mechanism for the Tx path. 343 * Reserve space in the completion queue and only proceed 344 * if there is space in it. This avoids having to implement 345 * any buffering in the Tx path. 346 */ 347 if (xskq_prod_reserve_addr(pool->cq, desc->addr)) 348 goto out; 349 350 xskq_cons_release(xs->tx); 351 rcu_read_unlock(); 352 return true; 353 } 354 355 out: 356 rcu_read_unlock(); 357 return false; 358 } 359 EXPORT_SYMBOL(xsk_tx_peek_desc); 360 361 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, struct xdp_desc *descs, 362 u32 max_entries) 363 { 364 u32 nb_pkts = 0; 365 366 while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts])) 367 nb_pkts++; 368 369 xsk_tx_release(pool); 370 return nb_pkts; 371 } 372 373 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, struct xdp_desc *descs, 374 u32 max_entries) 375 { 376 struct xdp_sock *xs; 377 u32 nb_pkts; 378 379 rcu_read_lock(); 380 if (!list_is_singular(&pool->xsk_tx_list)) { 381 /* Fallback to the non-batched version */ 382 rcu_read_unlock(); 383 return xsk_tx_peek_release_fallback(pool, descs, max_entries); 384 } 385 386 xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list); 387 if (!xs) { 388 nb_pkts = 0; 389 goto out; 390 } 391 392 nb_pkts = xskq_cons_peek_desc_batch(xs->tx, descs, pool, max_entries); 393 if (!nb_pkts) { 394 xs->tx->queue_empty_descs++; 395 goto out; 396 } 397 398 /* This is the backpressure mechanism for the Tx path. Try to 399 * reserve space in the completion queue for all packets, but 400 * if there are fewer slots available, just process that many 401 * packets. This avoids having to implement any buffering in 402 * the Tx path. 403 */ 404 nb_pkts = xskq_prod_reserve_addr_batch(pool->cq, descs, nb_pkts); 405 if (!nb_pkts) 406 goto out; 407 408 xskq_cons_release_n(xs->tx, nb_pkts); 409 __xskq_cons_release(xs->tx); 410 xs->sk.sk_write_space(&xs->sk); 411 412 out: 413 rcu_read_unlock(); 414 return nb_pkts; 415 } 416 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch); 417 418 static int xsk_wakeup(struct xdp_sock *xs, u8 flags) 419 { 420 struct net_device *dev = xs->dev; 421 int err; 422 423 rcu_read_lock(); 424 err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags); 425 rcu_read_unlock(); 426 427 return err; 428 } 429 430 static int xsk_zc_xmit(struct xdp_sock *xs) 431 { 432 return xsk_wakeup(xs, XDP_WAKEUP_TX); 433 } 434 435 static void xsk_destruct_skb(struct sk_buff *skb) 436 { 437 u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg; 438 struct xdp_sock *xs = xdp_sk(skb->sk); 439 unsigned long flags; 440 441 spin_lock_irqsave(&xs->pool->cq_lock, flags); 442 xskq_prod_submit_addr(xs->pool->cq, addr); 443 spin_unlock_irqrestore(&xs->pool->cq_lock, flags); 444 445 sock_wfree(skb); 446 } 447 448 static int xsk_generic_xmit(struct sock *sk) 449 { 450 struct xdp_sock *xs = xdp_sk(sk); 451 u32 max_batch = TX_BATCH_SIZE; 452 bool sent_frame = false; 453 struct xdp_desc desc; 454 struct sk_buff *skb; 455 unsigned long flags; 456 int err = 0; 457 458 mutex_lock(&xs->mutex); 459 460 if (xs->queue_id >= xs->dev->real_num_tx_queues) 461 goto out; 462 463 while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) { 464 char *buffer; 465 u64 addr; 466 u32 len; 467 468 if (max_batch-- == 0) { 469 err = -EAGAIN; 470 goto out; 471 } 472 473 len = desc.len; 474 skb = sock_alloc_send_skb(sk, len, 1, &err); 475 if (unlikely(!skb)) 476 goto out; 477 478 skb_put(skb, len); 479 addr = desc.addr; 480 buffer = xsk_buff_raw_get_data(xs->pool, addr); 481 err = skb_store_bits(skb, 0, buffer, len); 482 /* This is the backpressure mechanism for the Tx path. 483 * Reserve space in the completion queue and only proceed 484 * if there is space in it. This avoids having to implement 485 * any buffering in the Tx path. 486 */ 487 spin_lock_irqsave(&xs->pool->cq_lock, flags); 488 if (unlikely(err) || xskq_prod_reserve(xs->pool->cq)) { 489 spin_unlock_irqrestore(&xs->pool->cq_lock, flags); 490 kfree_skb(skb); 491 goto out; 492 } 493 spin_unlock_irqrestore(&xs->pool->cq_lock, flags); 494 495 skb->dev = xs->dev; 496 skb->priority = sk->sk_priority; 497 skb->mark = sk->sk_mark; 498 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr; 499 skb->destructor = xsk_destruct_skb; 500 501 err = __dev_direct_xmit(skb, xs->queue_id); 502 if (err == NETDEV_TX_BUSY) { 503 /* Tell user-space to retry the send */ 504 skb->destructor = sock_wfree; 505 spin_lock_irqsave(&xs->pool->cq_lock, flags); 506 xskq_prod_cancel(xs->pool->cq); 507 spin_unlock_irqrestore(&xs->pool->cq_lock, flags); 508 /* Free skb without triggering the perf drop trace */ 509 consume_skb(skb); 510 err = -EAGAIN; 511 goto out; 512 } 513 514 xskq_cons_release(xs->tx); 515 /* Ignore NET_XMIT_CN as packet might have been sent */ 516 if (err == NET_XMIT_DROP) { 517 /* SKB completed but not sent */ 518 err = -EBUSY; 519 goto out; 520 } 521 522 sent_frame = true; 523 } 524 525 xs->tx->queue_empty_descs++; 526 527 out: 528 if (sent_frame) 529 if (xsk_tx_writeable(xs)) 530 sk->sk_write_space(sk); 531 532 mutex_unlock(&xs->mutex); 533 return err; 534 } 535 536 static int __xsk_sendmsg(struct sock *sk) 537 { 538 struct xdp_sock *xs = xdp_sk(sk); 539 540 if (unlikely(!(xs->dev->flags & IFF_UP))) 541 return -ENETDOWN; 542 if (unlikely(!xs->tx)) 543 return -ENOBUFS; 544 545 return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk); 546 } 547 548 static bool xsk_no_wakeup(struct sock *sk) 549 { 550 #ifdef CONFIG_NET_RX_BUSY_POLL 551 /* Prefer busy-polling, skip the wakeup. */ 552 return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) && 553 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID; 554 #else 555 return false; 556 #endif 557 } 558 559 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) 560 { 561 bool need_wait = !(m->msg_flags & MSG_DONTWAIT); 562 struct sock *sk = sock->sk; 563 struct xdp_sock *xs = xdp_sk(sk); 564 struct xsk_buff_pool *pool; 565 566 if (unlikely(!xsk_is_bound(xs))) 567 return -ENXIO; 568 if (unlikely(need_wait)) 569 return -EOPNOTSUPP; 570 571 if (sk_can_busy_loop(sk)) 572 sk_busy_loop(sk, 1); /* only support non-blocking sockets */ 573 574 if (xsk_no_wakeup(sk)) 575 return 0; 576 577 pool = xs->pool; 578 if (pool->cached_need_wakeup & XDP_WAKEUP_TX) 579 return __xsk_sendmsg(sk); 580 return 0; 581 } 582 583 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags) 584 { 585 bool need_wait = !(flags & MSG_DONTWAIT); 586 struct sock *sk = sock->sk; 587 struct xdp_sock *xs = xdp_sk(sk); 588 589 if (unlikely(!xsk_is_bound(xs))) 590 return -ENXIO; 591 if (unlikely(!(xs->dev->flags & IFF_UP))) 592 return -ENETDOWN; 593 if (unlikely(!xs->rx)) 594 return -ENOBUFS; 595 if (unlikely(need_wait)) 596 return -EOPNOTSUPP; 597 598 if (sk_can_busy_loop(sk)) 599 sk_busy_loop(sk, 1); /* only support non-blocking sockets */ 600 601 if (xsk_no_wakeup(sk)) 602 return 0; 603 604 if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc) 605 return xsk_wakeup(xs, XDP_WAKEUP_RX); 606 return 0; 607 } 608 609 static __poll_t xsk_poll(struct file *file, struct socket *sock, 610 struct poll_table_struct *wait) 611 { 612 __poll_t mask = 0; 613 struct sock *sk = sock->sk; 614 struct xdp_sock *xs = xdp_sk(sk); 615 struct xsk_buff_pool *pool; 616 617 sock_poll_wait(file, sock, wait); 618 619 if (unlikely(!xsk_is_bound(xs))) 620 return mask; 621 622 pool = xs->pool; 623 624 if (pool->cached_need_wakeup) { 625 if (xs->zc) 626 xsk_wakeup(xs, pool->cached_need_wakeup); 627 else 628 /* Poll needs to drive Tx also in copy mode */ 629 __xsk_sendmsg(sk); 630 } 631 632 if (xs->rx && !xskq_prod_is_empty(xs->rx)) 633 mask |= EPOLLIN | EPOLLRDNORM; 634 if (xs->tx && xsk_tx_writeable(xs)) 635 mask |= EPOLLOUT | EPOLLWRNORM; 636 637 return mask; 638 } 639 640 static int xsk_init_queue(u32 entries, struct xsk_queue **queue, 641 bool umem_queue) 642 { 643 struct xsk_queue *q; 644 645 if (entries == 0 || *queue || !is_power_of_2(entries)) 646 return -EINVAL; 647 648 q = xskq_create(entries, umem_queue); 649 if (!q) 650 return -ENOMEM; 651 652 /* Make sure queue is ready before it can be seen by others */ 653 smp_wmb(); 654 WRITE_ONCE(*queue, q); 655 return 0; 656 } 657 658 static void xsk_unbind_dev(struct xdp_sock *xs) 659 { 660 struct net_device *dev = xs->dev; 661 662 if (xs->state != XSK_BOUND) 663 return; 664 WRITE_ONCE(xs->state, XSK_UNBOUND); 665 666 /* Wait for driver to stop using the xdp socket. */ 667 xp_del_xsk(xs->pool, xs); 668 xs->dev = NULL; 669 synchronize_net(); 670 dev_put(dev); 671 } 672 673 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs, 674 struct xdp_sock ***map_entry) 675 { 676 struct xsk_map *map = NULL; 677 struct xsk_map_node *node; 678 679 *map_entry = NULL; 680 681 spin_lock_bh(&xs->map_list_lock); 682 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node, 683 node); 684 if (node) { 685 bpf_map_inc(&node->map->map); 686 map = node->map; 687 *map_entry = node->map_entry; 688 } 689 spin_unlock_bh(&xs->map_list_lock); 690 return map; 691 } 692 693 static void xsk_delete_from_maps(struct xdp_sock *xs) 694 { 695 /* This function removes the current XDP socket from all the 696 * maps it resides in. We need to take extra care here, due to 697 * the two locks involved. Each map has a lock synchronizing 698 * updates to the entries, and each socket has a lock that 699 * synchronizes access to the list of maps (map_list). For 700 * deadlock avoidance the locks need to be taken in the order 701 * "map lock"->"socket map list lock". We start off by 702 * accessing the socket map list, and take a reference to the 703 * map to guarantee existence between the 704 * xsk_get_map_list_entry() and xsk_map_try_sock_delete() 705 * calls. Then we ask the map to remove the socket, which 706 * tries to remove the socket from the map. Note that there 707 * might be updates to the map between 708 * xsk_get_map_list_entry() and xsk_map_try_sock_delete(). 709 */ 710 struct xdp_sock **map_entry = NULL; 711 struct xsk_map *map; 712 713 while ((map = xsk_get_map_list_entry(xs, &map_entry))) { 714 xsk_map_try_sock_delete(map, xs, map_entry); 715 bpf_map_put(&map->map); 716 } 717 } 718 719 static int xsk_release(struct socket *sock) 720 { 721 struct sock *sk = sock->sk; 722 struct xdp_sock *xs = xdp_sk(sk); 723 struct net *net; 724 725 if (!sk) 726 return 0; 727 728 net = sock_net(sk); 729 730 mutex_lock(&net->xdp.lock); 731 sk_del_node_init_rcu(sk); 732 mutex_unlock(&net->xdp.lock); 733 734 local_bh_disable(); 735 sock_prot_inuse_add(net, sk->sk_prot, -1); 736 local_bh_enable(); 737 738 xsk_delete_from_maps(xs); 739 mutex_lock(&xs->mutex); 740 xsk_unbind_dev(xs); 741 mutex_unlock(&xs->mutex); 742 743 xskq_destroy(xs->rx); 744 xskq_destroy(xs->tx); 745 xskq_destroy(xs->fq_tmp); 746 xskq_destroy(xs->cq_tmp); 747 748 sock_orphan(sk); 749 sock->sk = NULL; 750 751 sk_refcnt_debug_release(sk); 752 sock_put(sk); 753 754 return 0; 755 } 756 757 static struct socket *xsk_lookup_xsk_from_fd(int fd) 758 { 759 struct socket *sock; 760 int err; 761 762 sock = sockfd_lookup(fd, &err); 763 if (!sock) 764 return ERR_PTR(-ENOTSOCK); 765 766 if (sock->sk->sk_family != PF_XDP) { 767 sockfd_put(sock); 768 return ERR_PTR(-ENOPROTOOPT); 769 } 770 771 return sock; 772 } 773 774 static bool xsk_validate_queues(struct xdp_sock *xs) 775 { 776 return xs->fq_tmp && xs->cq_tmp; 777 } 778 779 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len) 780 { 781 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr; 782 struct sock *sk = sock->sk; 783 struct xdp_sock *xs = xdp_sk(sk); 784 struct net_device *dev; 785 u32 flags, qid; 786 int err = 0; 787 788 if (addr_len < sizeof(struct sockaddr_xdp)) 789 return -EINVAL; 790 if (sxdp->sxdp_family != AF_XDP) 791 return -EINVAL; 792 793 flags = sxdp->sxdp_flags; 794 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY | 795 XDP_USE_NEED_WAKEUP)) 796 return -EINVAL; 797 798 rtnl_lock(); 799 mutex_lock(&xs->mutex); 800 if (xs->state != XSK_READY) { 801 err = -EBUSY; 802 goto out_release; 803 } 804 805 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex); 806 if (!dev) { 807 err = -ENODEV; 808 goto out_release; 809 } 810 811 if (!xs->rx && !xs->tx) { 812 err = -EINVAL; 813 goto out_unlock; 814 } 815 816 qid = sxdp->sxdp_queue_id; 817 818 if (flags & XDP_SHARED_UMEM) { 819 struct xdp_sock *umem_xs; 820 struct socket *sock; 821 822 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) || 823 (flags & XDP_USE_NEED_WAKEUP)) { 824 /* Cannot specify flags for shared sockets. */ 825 err = -EINVAL; 826 goto out_unlock; 827 } 828 829 if (xs->umem) { 830 /* We have already our own. */ 831 err = -EINVAL; 832 goto out_unlock; 833 } 834 835 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd); 836 if (IS_ERR(sock)) { 837 err = PTR_ERR(sock); 838 goto out_unlock; 839 } 840 841 umem_xs = xdp_sk(sock->sk); 842 if (!xsk_is_bound(umem_xs)) { 843 err = -EBADF; 844 sockfd_put(sock); 845 goto out_unlock; 846 } 847 848 if (umem_xs->queue_id != qid || umem_xs->dev != dev) { 849 /* Share the umem with another socket on another qid 850 * and/or device. 851 */ 852 xs->pool = xp_create_and_assign_umem(xs, 853 umem_xs->umem); 854 if (!xs->pool) { 855 err = -ENOMEM; 856 sockfd_put(sock); 857 goto out_unlock; 858 } 859 860 err = xp_assign_dev_shared(xs->pool, umem_xs->umem, 861 dev, qid); 862 if (err) { 863 xp_destroy(xs->pool); 864 xs->pool = NULL; 865 sockfd_put(sock); 866 goto out_unlock; 867 } 868 } else { 869 /* Share the buffer pool with the other socket. */ 870 if (xs->fq_tmp || xs->cq_tmp) { 871 /* Do not allow setting your own fq or cq. */ 872 err = -EINVAL; 873 sockfd_put(sock); 874 goto out_unlock; 875 } 876 877 xp_get_pool(umem_xs->pool); 878 xs->pool = umem_xs->pool; 879 } 880 881 xdp_get_umem(umem_xs->umem); 882 WRITE_ONCE(xs->umem, umem_xs->umem); 883 sockfd_put(sock); 884 } else if (!xs->umem || !xsk_validate_queues(xs)) { 885 err = -EINVAL; 886 goto out_unlock; 887 } else { 888 /* This xsk has its own umem. */ 889 xs->pool = xp_create_and_assign_umem(xs, xs->umem); 890 if (!xs->pool) { 891 err = -ENOMEM; 892 goto out_unlock; 893 } 894 895 err = xp_assign_dev(xs->pool, dev, qid, flags); 896 if (err) { 897 xp_destroy(xs->pool); 898 xs->pool = NULL; 899 goto out_unlock; 900 } 901 } 902 903 /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */ 904 xs->fq_tmp = NULL; 905 xs->cq_tmp = NULL; 906 907 xs->dev = dev; 908 xs->zc = xs->umem->zc; 909 xs->queue_id = qid; 910 xp_add_xsk(xs->pool, xs); 911 912 out_unlock: 913 if (err) { 914 dev_put(dev); 915 } else { 916 /* Matches smp_rmb() in bind() for shared umem 917 * sockets, and xsk_is_bound(). 918 */ 919 smp_wmb(); 920 WRITE_ONCE(xs->state, XSK_BOUND); 921 } 922 out_release: 923 mutex_unlock(&xs->mutex); 924 rtnl_unlock(); 925 return err; 926 } 927 928 struct xdp_umem_reg_v1 { 929 __u64 addr; /* Start of packet data area */ 930 __u64 len; /* Length of packet data area */ 931 __u32 chunk_size; 932 __u32 headroom; 933 }; 934 935 static int xsk_setsockopt(struct socket *sock, int level, int optname, 936 sockptr_t optval, unsigned int optlen) 937 { 938 struct sock *sk = sock->sk; 939 struct xdp_sock *xs = xdp_sk(sk); 940 int err; 941 942 if (level != SOL_XDP) 943 return -ENOPROTOOPT; 944 945 switch (optname) { 946 case XDP_RX_RING: 947 case XDP_TX_RING: 948 { 949 struct xsk_queue **q; 950 int entries; 951 952 if (optlen < sizeof(entries)) 953 return -EINVAL; 954 if (copy_from_sockptr(&entries, optval, sizeof(entries))) 955 return -EFAULT; 956 957 mutex_lock(&xs->mutex); 958 if (xs->state != XSK_READY) { 959 mutex_unlock(&xs->mutex); 960 return -EBUSY; 961 } 962 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx; 963 err = xsk_init_queue(entries, q, false); 964 if (!err && optname == XDP_TX_RING) 965 /* Tx needs to be explicitly woken up the first time */ 966 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 967 mutex_unlock(&xs->mutex); 968 return err; 969 } 970 case XDP_UMEM_REG: 971 { 972 size_t mr_size = sizeof(struct xdp_umem_reg); 973 struct xdp_umem_reg mr = {}; 974 struct xdp_umem *umem; 975 976 if (optlen < sizeof(struct xdp_umem_reg_v1)) 977 return -EINVAL; 978 else if (optlen < sizeof(mr)) 979 mr_size = sizeof(struct xdp_umem_reg_v1); 980 981 if (copy_from_sockptr(&mr, optval, mr_size)) 982 return -EFAULT; 983 984 mutex_lock(&xs->mutex); 985 if (xs->state != XSK_READY || xs->umem) { 986 mutex_unlock(&xs->mutex); 987 return -EBUSY; 988 } 989 990 umem = xdp_umem_create(&mr); 991 if (IS_ERR(umem)) { 992 mutex_unlock(&xs->mutex); 993 return PTR_ERR(umem); 994 } 995 996 /* Make sure umem is ready before it can be seen by others */ 997 smp_wmb(); 998 WRITE_ONCE(xs->umem, umem); 999 mutex_unlock(&xs->mutex); 1000 return 0; 1001 } 1002 case XDP_UMEM_FILL_RING: 1003 case XDP_UMEM_COMPLETION_RING: 1004 { 1005 struct xsk_queue **q; 1006 int entries; 1007 1008 if (copy_from_sockptr(&entries, optval, sizeof(entries))) 1009 return -EFAULT; 1010 1011 mutex_lock(&xs->mutex); 1012 if (xs->state != XSK_READY) { 1013 mutex_unlock(&xs->mutex); 1014 return -EBUSY; 1015 } 1016 1017 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp : 1018 &xs->cq_tmp; 1019 err = xsk_init_queue(entries, q, true); 1020 mutex_unlock(&xs->mutex); 1021 return err; 1022 } 1023 default: 1024 break; 1025 } 1026 1027 return -ENOPROTOOPT; 1028 } 1029 1030 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring) 1031 { 1032 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer); 1033 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer); 1034 ring->desc = offsetof(struct xdp_rxtx_ring, desc); 1035 } 1036 1037 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring) 1038 { 1039 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer); 1040 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer); 1041 ring->desc = offsetof(struct xdp_umem_ring, desc); 1042 } 1043 1044 struct xdp_statistics_v1 { 1045 __u64 rx_dropped; 1046 __u64 rx_invalid_descs; 1047 __u64 tx_invalid_descs; 1048 }; 1049 1050 static int xsk_getsockopt(struct socket *sock, int level, int optname, 1051 char __user *optval, int __user *optlen) 1052 { 1053 struct sock *sk = sock->sk; 1054 struct xdp_sock *xs = xdp_sk(sk); 1055 int len; 1056 1057 if (level != SOL_XDP) 1058 return -ENOPROTOOPT; 1059 1060 if (get_user(len, optlen)) 1061 return -EFAULT; 1062 if (len < 0) 1063 return -EINVAL; 1064 1065 switch (optname) { 1066 case XDP_STATISTICS: 1067 { 1068 struct xdp_statistics stats = {}; 1069 bool extra_stats = true; 1070 size_t stats_size; 1071 1072 if (len < sizeof(struct xdp_statistics_v1)) { 1073 return -EINVAL; 1074 } else if (len < sizeof(stats)) { 1075 extra_stats = false; 1076 stats_size = sizeof(struct xdp_statistics_v1); 1077 } else { 1078 stats_size = sizeof(stats); 1079 } 1080 1081 mutex_lock(&xs->mutex); 1082 stats.rx_dropped = xs->rx_dropped; 1083 if (extra_stats) { 1084 stats.rx_ring_full = xs->rx_queue_full; 1085 stats.rx_fill_ring_empty_descs = 1086 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0; 1087 stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx); 1088 } else { 1089 stats.rx_dropped += xs->rx_queue_full; 1090 } 1091 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx); 1092 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx); 1093 mutex_unlock(&xs->mutex); 1094 1095 if (copy_to_user(optval, &stats, stats_size)) 1096 return -EFAULT; 1097 if (put_user(stats_size, optlen)) 1098 return -EFAULT; 1099 1100 return 0; 1101 } 1102 case XDP_MMAP_OFFSETS: 1103 { 1104 struct xdp_mmap_offsets off; 1105 struct xdp_mmap_offsets_v1 off_v1; 1106 bool flags_supported = true; 1107 void *to_copy; 1108 1109 if (len < sizeof(off_v1)) 1110 return -EINVAL; 1111 else if (len < sizeof(off)) 1112 flags_supported = false; 1113 1114 if (flags_supported) { 1115 /* xdp_ring_offset is identical to xdp_ring_offset_v1 1116 * except for the flags field added to the end. 1117 */ 1118 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 1119 &off.rx); 1120 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 1121 &off.tx); 1122 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 1123 &off.fr); 1124 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 1125 &off.cr); 1126 off.rx.flags = offsetof(struct xdp_rxtx_ring, 1127 ptrs.flags); 1128 off.tx.flags = offsetof(struct xdp_rxtx_ring, 1129 ptrs.flags); 1130 off.fr.flags = offsetof(struct xdp_umem_ring, 1131 ptrs.flags); 1132 off.cr.flags = offsetof(struct xdp_umem_ring, 1133 ptrs.flags); 1134 1135 len = sizeof(off); 1136 to_copy = &off; 1137 } else { 1138 xsk_enter_rxtx_offsets(&off_v1.rx); 1139 xsk_enter_rxtx_offsets(&off_v1.tx); 1140 xsk_enter_umem_offsets(&off_v1.fr); 1141 xsk_enter_umem_offsets(&off_v1.cr); 1142 1143 len = sizeof(off_v1); 1144 to_copy = &off_v1; 1145 } 1146 1147 if (copy_to_user(optval, to_copy, len)) 1148 return -EFAULT; 1149 if (put_user(len, optlen)) 1150 return -EFAULT; 1151 1152 return 0; 1153 } 1154 case XDP_OPTIONS: 1155 { 1156 struct xdp_options opts = {}; 1157 1158 if (len < sizeof(opts)) 1159 return -EINVAL; 1160 1161 mutex_lock(&xs->mutex); 1162 if (xs->zc) 1163 opts.flags |= XDP_OPTIONS_ZEROCOPY; 1164 mutex_unlock(&xs->mutex); 1165 1166 len = sizeof(opts); 1167 if (copy_to_user(optval, &opts, len)) 1168 return -EFAULT; 1169 if (put_user(len, optlen)) 1170 return -EFAULT; 1171 1172 return 0; 1173 } 1174 default: 1175 break; 1176 } 1177 1178 return -EOPNOTSUPP; 1179 } 1180 1181 static int xsk_mmap(struct file *file, struct socket *sock, 1182 struct vm_area_struct *vma) 1183 { 1184 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT; 1185 unsigned long size = vma->vm_end - vma->vm_start; 1186 struct xdp_sock *xs = xdp_sk(sock->sk); 1187 struct xsk_queue *q = NULL; 1188 unsigned long pfn; 1189 struct page *qpg; 1190 1191 if (READ_ONCE(xs->state) != XSK_READY) 1192 return -EBUSY; 1193 1194 if (offset == XDP_PGOFF_RX_RING) { 1195 q = READ_ONCE(xs->rx); 1196 } else if (offset == XDP_PGOFF_TX_RING) { 1197 q = READ_ONCE(xs->tx); 1198 } else { 1199 /* Matches the smp_wmb() in XDP_UMEM_REG */ 1200 smp_rmb(); 1201 if (offset == XDP_UMEM_PGOFF_FILL_RING) 1202 q = READ_ONCE(xs->fq_tmp); 1203 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING) 1204 q = READ_ONCE(xs->cq_tmp); 1205 } 1206 1207 if (!q) 1208 return -EINVAL; 1209 1210 /* Matches the smp_wmb() in xsk_init_queue */ 1211 smp_rmb(); 1212 qpg = virt_to_head_page(q->ring); 1213 if (size > page_size(qpg)) 1214 return -EINVAL; 1215 1216 pfn = virt_to_phys(q->ring) >> PAGE_SHIFT; 1217 return remap_pfn_range(vma, vma->vm_start, pfn, 1218 size, vma->vm_page_prot); 1219 } 1220 1221 static int xsk_notifier(struct notifier_block *this, 1222 unsigned long msg, void *ptr) 1223 { 1224 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 1225 struct net *net = dev_net(dev); 1226 struct sock *sk; 1227 1228 switch (msg) { 1229 case NETDEV_UNREGISTER: 1230 mutex_lock(&net->xdp.lock); 1231 sk_for_each(sk, &net->xdp.list) { 1232 struct xdp_sock *xs = xdp_sk(sk); 1233 1234 mutex_lock(&xs->mutex); 1235 if (xs->dev == dev) { 1236 sk->sk_err = ENETDOWN; 1237 if (!sock_flag(sk, SOCK_DEAD)) 1238 sk->sk_error_report(sk); 1239 1240 xsk_unbind_dev(xs); 1241 1242 /* Clear device references. */ 1243 xp_clear_dev(xs->pool); 1244 } 1245 mutex_unlock(&xs->mutex); 1246 } 1247 mutex_unlock(&net->xdp.lock); 1248 break; 1249 } 1250 return NOTIFY_DONE; 1251 } 1252 1253 static struct proto xsk_proto = { 1254 .name = "XDP", 1255 .owner = THIS_MODULE, 1256 .obj_size = sizeof(struct xdp_sock), 1257 }; 1258 1259 static const struct proto_ops xsk_proto_ops = { 1260 .family = PF_XDP, 1261 .owner = THIS_MODULE, 1262 .release = xsk_release, 1263 .bind = xsk_bind, 1264 .connect = sock_no_connect, 1265 .socketpair = sock_no_socketpair, 1266 .accept = sock_no_accept, 1267 .getname = sock_no_getname, 1268 .poll = xsk_poll, 1269 .ioctl = sock_no_ioctl, 1270 .listen = sock_no_listen, 1271 .shutdown = sock_no_shutdown, 1272 .setsockopt = xsk_setsockopt, 1273 .getsockopt = xsk_getsockopt, 1274 .sendmsg = xsk_sendmsg, 1275 .recvmsg = xsk_recvmsg, 1276 .mmap = xsk_mmap, 1277 .sendpage = sock_no_sendpage, 1278 }; 1279 1280 static void xsk_destruct(struct sock *sk) 1281 { 1282 struct xdp_sock *xs = xdp_sk(sk); 1283 1284 if (!sock_flag(sk, SOCK_DEAD)) 1285 return; 1286 1287 if (!xp_put_pool(xs->pool)) 1288 xdp_put_umem(xs->umem, !xs->pool); 1289 1290 sk_refcnt_debug_dec(sk); 1291 } 1292 1293 static int xsk_create(struct net *net, struct socket *sock, int protocol, 1294 int kern) 1295 { 1296 struct xdp_sock *xs; 1297 struct sock *sk; 1298 1299 if (!ns_capable(net->user_ns, CAP_NET_RAW)) 1300 return -EPERM; 1301 if (sock->type != SOCK_RAW) 1302 return -ESOCKTNOSUPPORT; 1303 1304 if (protocol) 1305 return -EPROTONOSUPPORT; 1306 1307 sock->state = SS_UNCONNECTED; 1308 1309 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern); 1310 if (!sk) 1311 return -ENOBUFS; 1312 1313 sock->ops = &xsk_proto_ops; 1314 1315 sock_init_data(sock, sk); 1316 1317 sk->sk_family = PF_XDP; 1318 1319 sk->sk_destruct = xsk_destruct; 1320 sk_refcnt_debug_inc(sk); 1321 1322 sock_set_flag(sk, SOCK_RCU_FREE); 1323 1324 xs = xdp_sk(sk); 1325 xs->state = XSK_READY; 1326 mutex_init(&xs->mutex); 1327 spin_lock_init(&xs->rx_lock); 1328 1329 INIT_LIST_HEAD(&xs->map_list); 1330 spin_lock_init(&xs->map_list_lock); 1331 1332 mutex_lock(&net->xdp.lock); 1333 sk_add_node_rcu(sk, &net->xdp.list); 1334 mutex_unlock(&net->xdp.lock); 1335 1336 local_bh_disable(); 1337 sock_prot_inuse_add(net, &xsk_proto, 1); 1338 local_bh_enable(); 1339 1340 return 0; 1341 } 1342 1343 static const struct net_proto_family xsk_family_ops = { 1344 .family = PF_XDP, 1345 .create = xsk_create, 1346 .owner = THIS_MODULE, 1347 }; 1348 1349 static struct notifier_block xsk_netdev_notifier = { 1350 .notifier_call = xsk_notifier, 1351 }; 1352 1353 static int __net_init xsk_net_init(struct net *net) 1354 { 1355 mutex_init(&net->xdp.lock); 1356 INIT_HLIST_HEAD(&net->xdp.list); 1357 return 0; 1358 } 1359 1360 static void __net_exit xsk_net_exit(struct net *net) 1361 { 1362 WARN_ON_ONCE(!hlist_empty(&net->xdp.list)); 1363 } 1364 1365 static struct pernet_operations xsk_net_ops = { 1366 .init = xsk_net_init, 1367 .exit = xsk_net_exit, 1368 }; 1369 1370 static int __init xsk_init(void) 1371 { 1372 int err, cpu; 1373 1374 err = proto_register(&xsk_proto, 0 /* no slab */); 1375 if (err) 1376 goto out; 1377 1378 err = sock_register(&xsk_family_ops); 1379 if (err) 1380 goto out_proto; 1381 1382 err = register_pernet_subsys(&xsk_net_ops); 1383 if (err) 1384 goto out_sk; 1385 1386 err = register_netdevice_notifier(&xsk_netdev_notifier); 1387 if (err) 1388 goto out_pernet; 1389 1390 for_each_possible_cpu(cpu) 1391 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu)); 1392 return 0; 1393 1394 out_pernet: 1395 unregister_pernet_subsys(&xsk_net_ops); 1396 out_sk: 1397 sock_unregister(PF_XDP); 1398 out_proto: 1399 proto_unregister(&xsk_proto); 1400 out: 1401 return err; 1402 } 1403 1404 fs_initcall(xsk_init); 1405