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