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 <linux/vmalloc.h> 26 #include <net/xdp_sock_drv.h> 27 #include <net/busy_poll.h> 28 #include <net/netdev_lock.h> 29 #include <net/netdev_rx_queue.h> 30 #include <net/xdp.h> 31 32 #include "xsk_queue.h" 33 #include "xdp_umem.h" 34 #include "xsk.h" 35 36 #define TX_BATCH_SIZE 32 37 #define MAX_PER_SOCKET_BUDGET (TX_BATCH_SIZE) 38 39 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool) 40 { 41 if (pool->cached_need_wakeup & XDP_WAKEUP_RX) 42 return; 43 44 pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP; 45 pool->cached_need_wakeup |= XDP_WAKEUP_RX; 46 } 47 EXPORT_SYMBOL(xsk_set_rx_need_wakeup); 48 49 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool) 50 { 51 struct xdp_sock *xs; 52 53 if (pool->cached_need_wakeup & XDP_WAKEUP_TX) 54 return; 55 56 rcu_read_lock(); 57 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) { 58 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 59 } 60 rcu_read_unlock(); 61 62 pool->cached_need_wakeup |= XDP_WAKEUP_TX; 63 } 64 EXPORT_SYMBOL(xsk_set_tx_need_wakeup); 65 66 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool) 67 { 68 if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX)) 69 return; 70 71 pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP; 72 pool->cached_need_wakeup &= ~XDP_WAKEUP_RX; 73 } 74 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup); 75 76 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool) 77 { 78 struct xdp_sock *xs; 79 80 if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX)) 81 return; 82 83 rcu_read_lock(); 84 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) { 85 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP; 86 } 87 rcu_read_unlock(); 88 89 pool->cached_need_wakeup &= ~XDP_WAKEUP_TX; 90 } 91 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup); 92 93 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool) 94 { 95 return pool->uses_need_wakeup; 96 } 97 EXPORT_SYMBOL(xsk_uses_need_wakeup); 98 99 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev, 100 u16 queue_id) 101 { 102 if (queue_id < dev->real_num_rx_queues) 103 return dev->_rx[queue_id].pool; 104 if (queue_id < dev->real_num_tx_queues) 105 return dev->_tx[queue_id].pool; 106 107 return NULL; 108 } 109 EXPORT_SYMBOL(xsk_get_pool_from_qid); 110 111 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id) 112 { 113 if (queue_id < dev->num_rx_queues) 114 dev->_rx[queue_id].pool = NULL; 115 if (queue_id < dev->num_tx_queues) 116 dev->_tx[queue_id].pool = NULL; 117 } 118 119 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do 120 * not know if the device has more tx queues than rx, or the opposite. 121 * This might also change during run time. 122 */ 123 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool, 124 u16 queue_id) 125 { 126 if (queue_id >= max_t(unsigned int, 127 dev->real_num_rx_queues, 128 dev->real_num_tx_queues)) 129 return -EINVAL; 130 131 if (queue_id < dev->real_num_rx_queues) 132 dev->_rx[queue_id].pool = pool; 133 if (queue_id < dev->real_num_tx_queues) 134 dev->_tx[queue_id].pool = pool; 135 136 return 0; 137 } 138 139 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff_xsk *xskb, u32 len, 140 u32 flags) 141 { 142 u64 addr; 143 int err; 144 145 addr = xp_get_handle(xskb, xskb->pool); 146 err = xskq_prod_reserve_desc(xs->rx, addr, len, flags); 147 if (err) { 148 xs->rx_queue_full++; 149 return err; 150 } 151 152 xp_release(xskb); 153 return 0; 154 } 155 156 static int xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 157 { 158 struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp); 159 u32 frags = xdp_buff_has_frags(xdp); 160 struct xdp_buff_xsk *pos, *tmp; 161 struct list_head *xskb_list; 162 u32 contd = 0; 163 int err; 164 165 if (frags) 166 contd = XDP_PKT_CONTD; 167 168 err = __xsk_rcv_zc(xs, xskb, len, contd); 169 if (err) 170 goto err; 171 if (likely(!frags)) 172 return 0; 173 174 xskb_list = &xskb->pool->xskb_list; 175 list_for_each_entry_safe(pos, tmp, xskb_list, list_node) { 176 if (list_is_singular(xskb_list)) 177 contd = 0; 178 len = pos->xdp.data_end - pos->xdp.data; 179 err = __xsk_rcv_zc(xs, pos, len, contd); 180 if (err) 181 goto err; 182 list_del(&pos->list_node); 183 } 184 185 return 0; 186 err: 187 xsk_buff_free(xdp); 188 return err; 189 } 190 191 static void *xsk_copy_xdp_start(struct xdp_buff *from) 192 { 193 if (unlikely(xdp_data_meta_unsupported(from))) 194 return from->data; 195 else 196 return from->data_meta; 197 } 198 199 static u32 xsk_copy_xdp(void *to, void **from, u32 to_len, 200 u32 *from_len, skb_frag_t **frag, u32 rem) 201 { 202 u32 copied = 0; 203 204 while (1) { 205 u32 copy_len = min_t(u32, *from_len, to_len); 206 207 memcpy(to, *from, copy_len); 208 copied += copy_len; 209 if (rem == copied) 210 return copied; 211 212 if (*from_len == copy_len) { 213 *from = skb_frag_address(*frag); 214 *from_len = skb_frag_size((*frag)++); 215 } else { 216 *from += copy_len; 217 *from_len -= copy_len; 218 } 219 if (to_len == copy_len) 220 return copied; 221 222 to_len -= copy_len; 223 to += copy_len; 224 } 225 } 226 227 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 228 { 229 u32 frame_size = xsk_pool_get_rx_frame_size(xs->pool); 230 void *copy_from = xsk_copy_xdp_start(xdp), *copy_to; 231 u32 from_len, meta_len, rem, num_desc; 232 struct xdp_buff_xsk *xskb; 233 struct xdp_buff *xsk_xdp; 234 skb_frag_t *frag; 235 236 from_len = xdp->data_end - copy_from; 237 meta_len = xdp->data - copy_from; 238 rem = len + meta_len; 239 240 if (len <= frame_size && !xdp_buff_has_frags(xdp)) { 241 int err; 242 243 xsk_xdp = xsk_buff_alloc(xs->pool); 244 if (!xsk_xdp) { 245 xs->rx_dropped++; 246 return -ENOMEM; 247 } 248 memcpy(xsk_xdp->data - meta_len, copy_from, rem); 249 xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp); 250 err = __xsk_rcv_zc(xs, xskb, len, 0); 251 if (err) { 252 xsk_buff_free(xsk_xdp); 253 return err; 254 } 255 256 return 0; 257 } 258 259 num_desc = (len - 1) / frame_size + 1; 260 261 if (!xsk_buff_can_alloc(xs->pool, num_desc)) { 262 xs->rx_dropped++; 263 return -ENOMEM; 264 } 265 if (xskq_prod_nb_free(xs->rx, num_desc) < num_desc) { 266 xs->rx_queue_full++; 267 return -ENOBUFS; 268 } 269 270 if (xdp_buff_has_frags(xdp)) { 271 struct skb_shared_info *sinfo; 272 273 sinfo = xdp_get_shared_info_from_buff(xdp); 274 frag = &sinfo->frags[0]; 275 } 276 277 do { 278 u32 to_len = frame_size + meta_len; 279 u32 copied; 280 281 xsk_xdp = xsk_buff_alloc(xs->pool); 282 copy_to = xsk_xdp->data - meta_len; 283 284 copied = xsk_copy_xdp(copy_to, ©_from, to_len, &from_len, &frag, rem); 285 rem -= copied; 286 287 xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp); 288 __xsk_rcv_zc(xs, xskb, copied - meta_len, rem ? XDP_PKT_CONTD : 0); 289 meta_len = 0; 290 } while (rem); 291 292 return 0; 293 } 294 295 static bool xsk_tx_writeable(struct xdp_sock *xs) 296 { 297 if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2) 298 return false; 299 300 return true; 301 } 302 303 static bool xsk_is_bound(struct xdp_sock *xs) 304 { 305 if (READ_ONCE(xs->state) == XSK_BOUND) { 306 /* Matches smp_wmb() in bind(). */ 307 smp_rmb(); 308 return true; 309 } 310 return false; 311 } 312 313 static int xsk_rcv_check(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len) 314 { 315 if (!xsk_is_bound(xs)) 316 return -ENXIO; 317 318 if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index) 319 return -EINVAL; 320 321 if (len > xsk_pool_get_rx_frame_size(xs->pool) && !xs->sg) { 322 xs->rx_dropped++; 323 return -ENOSPC; 324 } 325 326 return 0; 327 } 328 329 static void xsk_flush(struct xdp_sock *xs) 330 { 331 xskq_prod_submit(xs->rx); 332 __xskq_cons_release(xs->pool->fq); 333 sock_def_readable(&xs->sk); 334 } 335 336 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 337 { 338 u32 len = xdp_get_buff_len(xdp); 339 int err; 340 341 spin_lock_bh(&xs->rx_lock); 342 err = xsk_rcv_check(xs, xdp, len); 343 if (!err) { 344 err = __xsk_rcv(xs, xdp, len); 345 xsk_flush(xs); 346 } 347 spin_unlock_bh(&xs->rx_lock); 348 return err; 349 } 350 351 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp) 352 { 353 u32 len = xdp_get_buff_len(xdp); 354 int err; 355 356 err = xsk_rcv_check(xs, xdp, len); 357 if (err) 358 return err; 359 360 if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) { 361 len = xdp->data_end - xdp->data; 362 return xsk_rcv_zc(xs, xdp, len); 363 } 364 365 err = __xsk_rcv(xs, xdp, len); 366 if (!err) 367 xdp_return_buff(xdp); 368 return err; 369 } 370 371 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp) 372 { 373 int err; 374 375 err = xsk_rcv(xs, xdp); 376 if (err) 377 return err; 378 379 if (!xs->flush_node.prev) { 380 struct list_head *flush_list = bpf_net_ctx_get_xskmap_flush_list(); 381 382 list_add(&xs->flush_node, flush_list); 383 } 384 385 return 0; 386 } 387 388 void __xsk_map_flush(struct list_head *flush_list) 389 { 390 struct xdp_sock *xs, *tmp; 391 392 list_for_each_entry_safe(xs, tmp, flush_list, flush_node) { 393 xsk_flush(xs); 394 __list_del_clearprev(&xs->flush_node); 395 } 396 } 397 398 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries) 399 { 400 xskq_prod_submit_n(pool->cq, nb_entries); 401 } 402 EXPORT_SYMBOL(xsk_tx_completed); 403 404 void xsk_tx_release(struct xsk_buff_pool *pool) 405 { 406 struct xdp_sock *xs; 407 408 rcu_read_lock(); 409 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) { 410 __xskq_cons_release(xs->tx); 411 if (xsk_tx_writeable(xs)) 412 xs->sk.sk_write_space(&xs->sk); 413 } 414 rcu_read_unlock(); 415 } 416 EXPORT_SYMBOL(xsk_tx_release); 417 418 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc) 419 { 420 bool budget_exhausted = false; 421 struct xdp_sock *xs; 422 423 rcu_read_lock(); 424 again: 425 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) { 426 if (xs->tx_budget_spent >= MAX_PER_SOCKET_BUDGET) { 427 budget_exhausted = true; 428 continue; 429 } 430 431 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) { 432 if (xskq_has_descs(xs->tx)) 433 xskq_cons_release(xs->tx); 434 continue; 435 } 436 437 xs->tx_budget_spent++; 438 439 /* This is the backpressure mechanism for the Tx path. 440 * Reserve space in the completion queue and only proceed 441 * if there is space in it. This avoids having to implement 442 * any buffering in the Tx path. 443 */ 444 if (xskq_prod_reserve_addr(pool->cq, desc->addr)) 445 goto out; 446 447 xskq_cons_release(xs->tx); 448 rcu_read_unlock(); 449 return true; 450 } 451 452 if (budget_exhausted) { 453 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) 454 xs->tx_budget_spent = 0; 455 456 budget_exhausted = false; 457 goto again; 458 } 459 460 out: 461 rcu_read_unlock(); 462 return false; 463 } 464 EXPORT_SYMBOL(xsk_tx_peek_desc); 465 466 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, u32 max_entries) 467 { 468 struct xdp_desc *descs = pool->tx_descs; 469 u32 nb_pkts = 0; 470 471 while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts])) 472 nb_pkts++; 473 474 xsk_tx_release(pool); 475 return nb_pkts; 476 } 477 478 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, u32 nb_pkts) 479 { 480 struct xdp_sock *xs; 481 482 rcu_read_lock(); 483 if (!list_is_singular(&pool->xsk_tx_list)) { 484 /* Fallback to the non-batched version */ 485 rcu_read_unlock(); 486 return xsk_tx_peek_release_fallback(pool, nb_pkts); 487 } 488 489 xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list); 490 if (!xs) { 491 nb_pkts = 0; 492 goto out; 493 } 494 495 nb_pkts = xskq_cons_nb_entries(xs->tx, nb_pkts); 496 497 /* This is the backpressure mechanism for the Tx path. Try to 498 * reserve space in the completion queue for all packets, but 499 * if there are fewer slots available, just process that many 500 * packets. This avoids having to implement any buffering in 501 * the Tx path. 502 */ 503 nb_pkts = xskq_prod_nb_free(pool->cq, nb_pkts); 504 if (!nb_pkts) 505 goto out; 506 507 nb_pkts = xskq_cons_read_desc_batch(xs->tx, pool, nb_pkts); 508 if (!nb_pkts) { 509 xs->tx->queue_empty_descs++; 510 goto out; 511 } 512 513 __xskq_cons_release(xs->tx); 514 xskq_prod_write_addr_batch(pool->cq, pool->tx_descs, nb_pkts); 515 xs->sk.sk_write_space(&xs->sk); 516 517 out: 518 rcu_read_unlock(); 519 return nb_pkts; 520 } 521 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch); 522 523 static int xsk_wakeup(struct xdp_sock *xs, u8 flags) 524 { 525 struct net_device *dev = xs->dev; 526 527 return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags); 528 } 529 530 static int xsk_cq_reserve_addr_locked(struct xsk_buff_pool *pool, u64 addr) 531 { 532 unsigned long flags; 533 int ret; 534 535 spin_lock_irqsave(&pool->cq_lock, flags); 536 ret = xskq_prod_reserve_addr(pool->cq, addr); 537 spin_unlock_irqrestore(&pool->cq_lock, flags); 538 539 return ret; 540 } 541 542 static void xsk_cq_submit_locked(struct xsk_buff_pool *pool, u32 n) 543 { 544 unsigned long flags; 545 546 spin_lock_irqsave(&pool->cq_lock, flags); 547 xskq_prod_submit_n(pool->cq, n); 548 spin_unlock_irqrestore(&pool->cq_lock, flags); 549 } 550 551 static void xsk_cq_cancel_locked(struct xsk_buff_pool *pool, u32 n) 552 { 553 unsigned long flags; 554 555 spin_lock_irqsave(&pool->cq_lock, flags); 556 xskq_prod_cancel_n(pool->cq, n); 557 spin_unlock_irqrestore(&pool->cq_lock, flags); 558 } 559 560 static u32 xsk_get_num_desc(struct sk_buff *skb) 561 { 562 return skb ? (long)skb_shinfo(skb)->destructor_arg : 0; 563 } 564 565 static void xsk_destruct_skb(struct sk_buff *skb) 566 { 567 struct xsk_tx_metadata_compl *compl = &skb_shinfo(skb)->xsk_meta; 568 569 if (compl->tx_timestamp) { 570 /* sw completion timestamp, not a real one */ 571 *compl->tx_timestamp = ktime_get_tai_fast_ns(); 572 } 573 574 xsk_cq_submit_locked(xdp_sk(skb->sk)->pool, xsk_get_num_desc(skb)); 575 sock_wfree(skb); 576 } 577 578 static void xsk_set_destructor_arg(struct sk_buff *skb) 579 { 580 long num = xsk_get_num_desc(xdp_sk(skb->sk)->skb) + 1; 581 582 skb_shinfo(skb)->destructor_arg = (void *)num; 583 } 584 585 static void xsk_consume_skb(struct sk_buff *skb) 586 { 587 struct xdp_sock *xs = xdp_sk(skb->sk); 588 589 skb->destructor = sock_wfree; 590 xsk_cq_cancel_locked(xs->pool, xsk_get_num_desc(skb)); 591 /* Free skb without triggering the perf drop trace */ 592 consume_skb(skb); 593 xs->skb = NULL; 594 } 595 596 static void xsk_drop_skb(struct sk_buff *skb) 597 { 598 xdp_sk(skb->sk)->tx->invalid_descs += xsk_get_num_desc(skb); 599 xsk_consume_skb(skb); 600 } 601 602 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs, 603 struct xdp_desc *desc) 604 { 605 struct xsk_buff_pool *pool = xs->pool; 606 u32 hr, len, ts, offset, copy, copied; 607 struct sk_buff *skb = xs->skb; 608 struct page *page; 609 void *buffer; 610 int err, i; 611 u64 addr; 612 613 if (!skb) { 614 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom)); 615 616 skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err); 617 if (unlikely(!skb)) 618 return ERR_PTR(err); 619 620 skb_reserve(skb, hr); 621 } 622 623 addr = desc->addr; 624 len = desc->len; 625 ts = pool->unaligned ? len : pool->chunk_size; 626 627 buffer = xsk_buff_raw_get_data(pool, addr); 628 offset = offset_in_page(buffer); 629 addr = buffer - pool->addrs; 630 631 for (copied = 0, i = skb_shinfo(skb)->nr_frags; copied < len; i++) { 632 if (unlikely(i >= MAX_SKB_FRAGS)) 633 return ERR_PTR(-EOVERFLOW); 634 635 page = pool->umem->pgs[addr >> PAGE_SHIFT]; 636 get_page(page); 637 638 copy = min_t(u32, PAGE_SIZE - offset, len - copied); 639 skb_fill_page_desc(skb, i, page, offset, copy); 640 641 copied += copy; 642 addr += copy; 643 offset = 0; 644 } 645 646 skb->len += len; 647 skb->data_len += len; 648 skb->truesize += ts; 649 650 refcount_add(ts, &xs->sk.sk_wmem_alloc); 651 652 return skb; 653 } 654 655 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs, 656 struct xdp_desc *desc) 657 { 658 struct xsk_tx_metadata *meta = NULL; 659 struct net_device *dev = xs->dev; 660 struct sk_buff *skb = xs->skb; 661 bool first_frag = false; 662 int err; 663 664 if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) { 665 skb = xsk_build_skb_zerocopy(xs, desc); 666 if (IS_ERR(skb)) { 667 err = PTR_ERR(skb); 668 goto free_err; 669 } 670 } else { 671 u32 hr, tr, len; 672 void *buffer; 673 674 buffer = xsk_buff_raw_get_data(xs->pool, desc->addr); 675 len = desc->len; 676 677 if (!skb) { 678 first_frag = true; 679 680 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom)); 681 tr = dev->needed_tailroom; 682 skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err); 683 if (unlikely(!skb)) 684 goto free_err; 685 686 skb_reserve(skb, hr); 687 skb_put(skb, len); 688 689 err = skb_store_bits(skb, 0, buffer, len); 690 if (unlikely(err)) 691 goto free_err; 692 } else { 693 int nr_frags = skb_shinfo(skb)->nr_frags; 694 struct page *page; 695 u8 *vaddr; 696 697 if (unlikely(nr_frags == (MAX_SKB_FRAGS - 1) && xp_mb_desc(desc))) { 698 err = -EOVERFLOW; 699 goto free_err; 700 } 701 702 page = alloc_page(xs->sk.sk_allocation); 703 if (unlikely(!page)) { 704 err = -EAGAIN; 705 goto free_err; 706 } 707 708 vaddr = kmap_local_page(page); 709 memcpy(vaddr, buffer, len); 710 kunmap_local(vaddr); 711 712 skb_add_rx_frag(skb, nr_frags, page, 0, len, PAGE_SIZE); 713 refcount_add(PAGE_SIZE, &xs->sk.sk_wmem_alloc); 714 } 715 716 if (first_frag && desc->options & XDP_TX_METADATA) { 717 if (unlikely(xs->pool->tx_metadata_len == 0)) { 718 err = -EINVAL; 719 goto free_err; 720 } 721 722 meta = buffer - xs->pool->tx_metadata_len; 723 if (unlikely(!xsk_buff_valid_tx_metadata(meta))) { 724 err = -EINVAL; 725 goto free_err; 726 } 727 728 if (meta->flags & XDP_TXMD_FLAGS_CHECKSUM) { 729 if (unlikely(meta->request.csum_start + 730 meta->request.csum_offset + 731 sizeof(__sum16) > len)) { 732 err = -EINVAL; 733 goto free_err; 734 } 735 736 skb->csum_start = hr + meta->request.csum_start; 737 skb->csum_offset = meta->request.csum_offset; 738 skb->ip_summed = CHECKSUM_PARTIAL; 739 740 if (unlikely(xs->pool->tx_sw_csum)) { 741 err = skb_checksum_help(skb); 742 if (err) 743 goto free_err; 744 } 745 } 746 747 if (meta->flags & XDP_TXMD_FLAGS_LAUNCH_TIME) 748 skb->skb_mstamp_ns = meta->request.launch_time; 749 } 750 } 751 752 skb->dev = dev; 753 skb->priority = READ_ONCE(xs->sk.sk_priority); 754 skb->mark = READ_ONCE(xs->sk.sk_mark); 755 skb->destructor = xsk_destruct_skb; 756 xsk_tx_metadata_to_compl(meta, &skb_shinfo(skb)->xsk_meta); 757 xsk_set_destructor_arg(skb); 758 759 return skb; 760 761 free_err: 762 if (first_frag && skb) 763 kfree_skb(skb); 764 765 if (err == -EOVERFLOW) { 766 /* Drop the packet */ 767 xsk_set_destructor_arg(xs->skb); 768 xsk_drop_skb(xs->skb); 769 xskq_cons_release(xs->tx); 770 } else { 771 /* Let application retry */ 772 xsk_cq_cancel_locked(xs->pool, 1); 773 } 774 775 return ERR_PTR(err); 776 } 777 778 static int __xsk_generic_xmit(struct sock *sk) 779 { 780 struct xdp_sock *xs = xdp_sk(sk); 781 u32 max_batch = TX_BATCH_SIZE; 782 bool sent_frame = false; 783 struct xdp_desc desc; 784 struct sk_buff *skb; 785 int err = 0; 786 787 mutex_lock(&xs->mutex); 788 789 /* Since we dropped the RCU read lock, the socket state might have changed. */ 790 if (unlikely(!xsk_is_bound(xs))) { 791 err = -ENXIO; 792 goto out; 793 } 794 795 if (xs->queue_id >= xs->dev->real_num_tx_queues) 796 goto out; 797 798 while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) { 799 if (max_batch-- == 0) { 800 err = -EAGAIN; 801 goto out; 802 } 803 804 /* This is the backpressure mechanism for the Tx path. 805 * Reserve space in the completion queue and only proceed 806 * if there is space in it. This avoids having to implement 807 * any buffering in the Tx path. 808 */ 809 if (xsk_cq_reserve_addr_locked(xs->pool, desc.addr)) 810 goto out; 811 812 skb = xsk_build_skb(xs, &desc); 813 if (IS_ERR(skb)) { 814 err = PTR_ERR(skb); 815 if (err != -EOVERFLOW) 816 goto out; 817 err = 0; 818 continue; 819 } 820 821 xskq_cons_release(xs->tx); 822 823 if (xp_mb_desc(&desc)) { 824 xs->skb = skb; 825 continue; 826 } 827 828 err = __dev_direct_xmit(skb, xs->queue_id); 829 if (err == NETDEV_TX_BUSY) { 830 /* Tell user-space to retry the send */ 831 xskq_cons_cancel_n(xs->tx, xsk_get_num_desc(skb)); 832 xsk_consume_skb(skb); 833 err = -EAGAIN; 834 goto out; 835 } 836 837 /* Ignore NET_XMIT_CN as packet might have been sent */ 838 if (err == NET_XMIT_DROP) { 839 /* SKB completed but not sent */ 840 err = -EBUSY; 841 xs->skb = NULL; 842 goto out; 843 } 844 845 sent_frame = true; 846 xs->skb = NULL; 847 } 848 849 if (xskq_has_descs(xs->tx)) { 850 if (xs->skb) 851 xsk_drop_skb(xs->skb); 852 xskq_cons_release(xs->tx); 853 } 854 855 out: 856 if (sent_frame) 857 if (xsk_tx_writeable(xs)) 858 sk->sk_write_space(sk); 859 860 mutex_unlock(&xs->mutex); 861 return err; 862 } 863 864 static int xsk_generic_xmit(struct sock *sk) 865 { 866 int ret; 867 868 /* Drop the RCU lock since the SKB path might sleep. */ 869 rcu_read_unlock(); 870 ret = __xsk_generic_xmit(sk); 871 /* Reaquire RCU lock before going into common code. */ 872 rcu_read_lock(); 873 874 return ret; 875 } 876 877 static bool xsk_no_wakeup(struct sock *sk) 878 { 879 #ifdef CONFIG_NET_RX_BUSY_POLL 880 /* Prefer busy-polling, skip the wakeup. */ 881 return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) && 882 napi_id_valid(READ_ONCE(sk->sk_napi_id)); 883 #else 884 return false; 885 #endif 886 } 887 888 static int xsk_check_common(struct xdp_sock *xs) 889 { 890 if (unlikely(!xsk_is_bound(xs))) 891 return -ENXIO; 892 if (unlikely(!(xs->dev->flags & IFF_UP))) 893 return -ENETDOWN; 894 895 return 0; 896 } 897 898 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) 899 { 900 bool need_wait = !(m->msg_flags & MSG_DONTWAIT); 901 struct sock *sk = sock->sk; 902 struct xdp_sock *xs = xdp_sk(sk); 903 struct xsk_buff_pool *pool; 904 int err; 905 906 err = xsk_check_common(xs); 907 if (err) 908 return err; 909 if (unlikely(need_wait)) 910 return -EOPNOTSUPP; 911 if (unlikely(!xs->tx)) 912 return -ENOBUFS; 913 914 if (sk_can_busy_loop(sk)) 915 sk_busy_loop(sk, 1); /* only support non-blocking sockets */ 916 917 if (xs->zc && xsk_no_wakeup(sk)) 918 return 0; 919 920 pool = xs->pool; 921 if (pool->cached_need_wakeup & XDP_WAKEUP_TX) { 922 if (xs->zc) 923 return xsk_wakeup(xs, XDP_WAKEUP_TX); 924 return xsk_generic_xmit(sk); 925 } 926 return 0; 927 } 928 929 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) 930 { 931 int ret; 932 933 rcu_read_lock(); 934 ret = __xsk_sendmsg(sock, m, total_len); 935 rcu_read_unlock(); 936 937 return ret; 938 } 939 940 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags) 941 { 942 bool need_wait = !(flags & MSG_DONTWAIT); 943 struct sock *sk = sock->sk; 944 struct xdp_sock *xs = xdp_sk(sk); 945 int err; 946 947 err = xsk_check_common(xs); 948 if (err) 949 return err; 950 if (unlikely(!xs->rx)) 951 return -ENOBUFS; 952 if (unlikely(need_wait)) 953 return -EOPNOTSUPP; 954 955 if (sk_can_busy_loop(sk)) 956 sk_busy_loop(sk, 1); /* only support non-blocking sockets */ 957 958 if (xsk_no_wakeup(sk)) 959 return 0; 960 961 if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc) 962 return xsk_wakeup(xs, XDP_WAKEUP_RX); 963 return 0; 964 } 965 966 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags) 967 { 968 int ret; 969 970 rcu_read_lock(); 971 ret = __xsk_recvmsg(sock, m, len, flags); 972 rcu_read_unlock(); 973 974 return ret; 975 } 976 977 static __poll_t xsk_poll(struct file *file, struct socket *sock, 978 struct poll_table_struct *wait) 979 { 980 __poll_t mask = 0; 981 struct sock *sk = sock->sk; 982 struct xdp_sock *xs = xdp_sk(sk); 983 struct xsk_buff_pool *pool; 984 985 sock_poll_wait(file, sock, wait); 986 987 rcu_read_lock(); 988 if (xsk_check_common(xs)) 989 goto out; 990 991 pool = xs->pool; 992 993 if (pool->cached_need_wakeup) { 994 if (xs->zc) 995 xsk_wakeup(xs, pool->cached_need_wakeup); 996 else if (xs->tx) 997 /* Poll needs to drive Tx also in copy mode */ 998 xsk_generic_xmit(sk); 999 } 1000 1001 if (xs->rx && !xskq_prod_is_empty(xs->rx)) 1002 mask |= EPOLLIN | EPOLLRDNORM; 1003 if (xs->tx && xsk_tx_writeable(xs)) 1004 mask |= EPOLLOUT | EPOLLWRNORM; 1005 out: 1006 rcu_read_unlock(); 1007 return mask; 1008 } 1009 1010 static int xsk_init_queue(u32 entries, struct xsk_queue **queue, 1011 bool umem_queue) 1012 { 1013 struct xsk_queue *q; 1014 1015 if (entries == 0 || *queue || !is_power_of_2(entries)) 1016 return -EINVAL; 1017 1018 q = xskq_create(entries, umem_queue); 1019 if (!q) 1020 return -ENOMEM; 1021 1022 /* Make sure queue is ready before it can be seen by others */ 1023 smp_wmb(); 1024 WRITE_ONCE(*queue, q); 1025 return 0; 1026 } 1027 1028 static void xsk_unbind_dev(struct xdp_sock *xs) 1029 { 1030 struct net_device *dev = xs->dev; 1031 1032 if (xs->state != XSK_BOUND) 1033 return; 1034 WRITE_ONCE(xs->state, XSK_UNBOUND); 1035 1036 /* Wait for driver to stop using the xdp socket. */ 1037 xp_del_xsk(xs->pool, xs); 1038 synchronize_net(); 1039 dev_put(dev); 1040 } 1041 1042 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs, 1043 struct xdp_sock __rcu ***map_entry) 1044 { 1045 struct xsk_map *map = NULL; 1046 struct xsk_map_node *node; 1047 1048 *map_entry = NULL; 1049 1050 spin_lock_bh(&xs->map_list_lock); 1051 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node, 1052 node); 1053 if (node) { 1054 bpf_map_inc(&node->map->map); 1055 map = node->map; 1056 *map_entry = node->map_entry; 1057 } 1058 spin_unlock_bh(&xs->map_list_lock); 1059 return map; 1060 } 1061 1062 static void xsk_delete_from_maps(struct xdp_sock *xs) 1063 { 1064 /* This function removes the current XDP socket from all the 1065 * maps it resides in. We need to take extra care here, due to 1066 * the two locks involved. Each map has a lock synchronizing 1067 * updates to the entries, and each socket has a lock that 1068 * synchronizes access to the list of maps (map_list). For 1069 * deadlock avoidance the locks need to be taken in the order 1070 * "map lock"->"socket map list lock". We start off by 1071 * accessing the socket map list, and take a reference to the 1072 * map to guarantee existence between the 1073 * xsk_get_map_list_entry() and xsk_map_try_sock_delete() 1074 * calls. Then we ask the map to remove the socket, which 1075 * tries to remove the socket from the map. Note that there 1076 * might be updates to the map between 1077 * xsk_get_map_list_entry() and xsk_map_try_sock_delete(). 1078 */ 1079 struct xdp_sock __rcu **map_entry = NULL; 1080 struct xsk_map *map; 1081 1082 while ((map = xsk_get_map_list_entry(xs, &map_entry))) { 1083 xsk_map_try_sock_delete(map, xs, map_entry); 1084 bpf_map_put(&map->map); 1085 } 1086 } 1087 1088 static int xsk_release(struct socket *sock) 1089 { 1090 struct sock *sk = sock->sk; 1091 struct xdp_sock *xs = xdp_sk(sk); 1092 struct net *net; 1093 1094 if (!sk) 1095 return 0; 1096 1097 net = sock_net(sk); 1098 1099 if (xs->skb) 1100 xsk_drop_skb(xs->skb); 1101 1102 mutex_lock(&net->xdp.lock); 1103 sk_del_node_init_rcu(sk); 1104 mutex_unlock(&net->xdp.lock); 1105 1106 sock_prot_inuse_add(net, sk->sk_prot, -1); 1107 1108 xsk_delete_from_maps(xs); 1109 mutex_lock(&xs->mutex); 1110 xsk_unbind_dev(xs); 1111 mutex_unlock(&xs->mutex); 1112 1113 xskq_destroy(xs->rx); 1114 xskq_destroy(xs->tx); 1115 xskq_destroy(xs->fq_tmp); 1116 xskq_destroy(xs->cq_tmp); 1117 1118 sock_orphan(sk); 1119 sock->sk = NULL; 1120 1121 sock_put(sk); 1122 1123 return 0; 1124 } 1125 1126 static struct socket *xsk_lookup_xsk_from_fd(int fd) 1127 { 1128 struct socket *sock; 1129 int err; 1130 1131 sock = sockfd_lookup(fd, &err); 1132 if (!sock) 1133 return ERR_PTR(-ENOTSOCK); 1134 1135 if (sock->sk->sk_family != PF_XDP) { 1136 sockfd_put(sock); 1137 return ERR_PTR(-ENOPROTOOPT); 1138 } 1139 1140 return sock; 1141 } 1142 1143 static bool xsk_validate_queues(struct xdp_sock *xs) 1144 { 1145 return xs->fq_tmp && xs->cq_tmp; 1146 } 1147 1148 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len) 1149 { 1150 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr; 1151 struct sock *sk = sock->sk; 1152 struct xdp_sock *xs = xdp_sk(sk); 1153 struct net_device *dev; 1154 int bound_dev_if; 1155 u32 flags, qid; 1156 int err = 0; 1157 1158 if (addr_len < sizeof(struct sockaddr_xdp)) 1159 return -EINVAL; 1160 if (sxdp->sxdp_family != AF_XDP) 1161 return -EINVAL; 1162 1163 flags = sxdp->sxdp_flags; 1164 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY | 1165 XDP_USE_NEED_WAKEUP | XDP_USE_SG)) 1166 return -EINVAL; 1167 1168 bound_dev_if = READ_ONCE(sk->sk_bound_dev_if); 1169 if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex) 1170 return -EINVAL; 1171 1172 rtnl_lock(); 1173 mutex_lock(&xs->mutex); 1174 if (xs->state != XSK_READY) { 1175 err = -EBUSY; 1176 goto out_release; 1177 } 1178 1179 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex); 1180 if (!dev) { 1181 err = -ENODEV; 1182 goto out_release; 1183 } 1184 1185 netdev_lock_ops(dev); 1186 1187 if (!xs->rx && !xs->tx) { 1188 err = -EINVAL; 1189 goto out_unlock; 1190 } 1191 1192 qid = sxdp->sxdp_queue_id; 1193 1194 if (flags & XDP_SHARED_UMEM) { 1195 struct xdp_sock *umem_xs; 1196 struct socket *sock; 1197 1198 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) || 1199 (flags & XDP_USE_NEED_WAKEUP) || (flags & XDP_USE_SG)) { 1200 /* Cannot specify flags for shared sockets. */ 1201 err = -EINVAL; 1202 goto out_unlock; 1203 } 1204 1205 if (xs->umem) { 1206 /* We have already our own. */ 1207 err = -EINVAL; 1208 goto out_unlock; 1209 } 1210 1211 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd); 1212 if (IS_ERR(sock)) { 1213 err = PTR_ERR(sock); 1214 goto out_unlock; 1215 } 1216 1217 umem_xs = xdp_sk(sock->sk); 1218 if (!xsk_is_bound(umem_xs)) { 1219 err = -EBADF; 1220 sockfd_put(sock); 1221 goto out_unlock; 1222 } 1223 1224 if (umem_xs->queue_id != qid || umem_xs->dev != dev) { 1225 /* Share the umem with another socket on another qid 1226 * and/or device. 1227 */ 1228 xs->pool = xp_create_and_assign_umem(xs, 1229 umem_xs->umem); 1230 if (!xs->pool) { 1231 err = -ENOMEM; 1232 sockfd_put(sock); 1233 goto out_unlock; 1234 } 1235 1236 err = xp_assign_dev_shared(xs->pool, umem_xs, dev, 1237 qid); 1238 if (err) { 1239 xp_destroy(xs->pool); 1240 xs->pool = NULL; 1241 sockfd_put(sock); 1242 goto out_unlock; 1243 } 1244 } else { 1245 /* Share the buffer pool with the other socket. */ 1246 if (xs->fq_tmp || xs->cq_tmp) { 1247 /* Do not allow setting your own fq or cq. */ 1248 err = -EINVAL; 1249 sockfd_put(sock); 1250 goto out_unlock; 1251 } 1252 1253 xp_get_pool(umem_xs->pool); 1254 xs->pool = umem_xs->pool; 1255 1256 /* If underlying shared umem was created without Tx 1257 * ring, allocate Tx descs array that Tx batching API 1258 * utilizes 1259 */ 1260 if (xs->tx && !xs->pool->tx_descs) { 1261 err = xp_alloc_tx_descs(xs->pool, xs); 1262 if (err) { 1263 xp_put_pool(xs->pool); 1264 xs->pool = NULL; 1265 sockfd_put(sock); 1266 goto out_unlock; 1267 } 1268 } 1269 } 1270 1271 xdp_get_umem(umem_xs->umem); 1272 WRITE_ONCE(xs->umem, umem_xs->umem); 1273 sockfd_put(sock); 1274 } else if (!xs->umem || !xsk_validate_queues(xs)) { 1275 err = -EINVAL; 1276 goto out_unlock; 1277 } else { 1278 /* This xsk has its own umem. */ 1279 xs->pool = xp_create_and_assign_umem(xs, xs->umem); 1280 if (!xs->pool) { 1281 err = -ENOMEM; 1282 goto out_unlock; 1283 } 1284 1285 err = xp_assign_dev(xs->pool, dev, qid, flags); 1286 if (err) { 1287 xp_destroy(xs->pool); 1288 xs->pool = NULL; 1289 goto out_unlock; 1290 } 1291 } 1292 1293 /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */ 1294 xs->fq_tmp = NULL; 1295 xs->cq_tmp = NULL; 1296 1297 xs->dev = dev; 1298 xs->zc = xs->umem->zc; 1299 xs->sg = !!(xs->umem->flags & XDP_UMEM_SG_FLAG); 1300 xs->queue_id = qid; 1301 xp_add_xsk(xs->pool, xs); 1302 1303 if (xs->zc && qid < dev->real_num_rx_queues) { 1304 struct netdev_rx_queue *rxq; 1305 1306 rxq = __netif_get_rx_queue(dev, qid); 1307 if (rxq->napi) 1308 __sk_mark_napi_id_once(sk, rxq->napi->napi_id); 1309 } 1310 1311 out_unlock: 1312 if (err) { 1313 dev_put(dev); 1314 } else { 1315 /* Matches smp_rmb() in bind() for shared umem 1316 * sockets, and xsk_is_bound(). 1317 */ 1318 smp_wmb(); 1319 WRITE_ONCE(xs->state, XSK_BOUND); 1320 } 1321 netdev_unlock_ops(dev); 1322 out_release: 1323 mutex_unlock(&xs->mutex); 1324 rtnl_unlock(); 1325 return err; 1326 } 1327 1328 struct xdp_umem_reg_v1 { 1329 __u64 addr; /* Start of packet data area */ 1330 __u64 len; /* Length of packet data area */ 1331 __u32 chunk_size; 1332 __u32 headroom; 1333 }; 1334 1335 static int xsk_setsockopt(struct socket *sock, int level, int optname, 1336 sockptr_t optval, unsigned int optlen) 1337 { 1338 struct sock *sk = sock->sk; 1339 struct xdp_sock *xs = xdp_sk(sk); 1340 int err; 1341 1342 if (level != SOL_XDP) 1343 return -ENOPROTOOPT; 1344 1345 switch (optname) { 1346 case XDP_RX_RING: 1347 case XDP_TX_RING: 1348 { 1349 struct xsk_queue **q; 1350 int entries; 1351 1352 if (optlen < sizeof(entries)) 1353 return -EINVAL; 1354 if (copy_from_sockptr(&entries, optval, sizeof(entries))) 1355 return -EFAULT; 1356 1357 mutex_lock(&xs->mutex); 1358 if (xs->state != XSK_READY) { 1359 mutex_unlock(&xs->mutex); 1360 return -EBUSY; 1361 } 1362 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx; 1363 err = xsk_init_queue(entries, q, false); 1364 if (!err && optname == XDP_TX_RING) 1365 /* Tx needs to be explicitly woken up the first time */ 1366 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 1367 mutex_unlock(&xs->mutex); 1368 return err; 1369 } 1370 case XDP_UMEM_REG: 1371 { 1372 size_t mr_size = sizeof(struct xdp_umem_reg); 1373 struct xdp_umem_reg mr = {}; 1374 struct xdp_umem *umem; 1375 1376 if (optlen < sizeof(struct xdp_umem_reg_v1)) 1377 return -EINVAL; 1378 else if (optlen < sizeof(mr)) 1379 mr_size = sizeof(struct xdp_umem_reg_v1); 1380 1381 BUILD_BUG_ON(sizeof(struct xdp_umem_reg_v1) >= sizeof(struct xdp_umem_reg)); 1382 1383 /* Make sure the last field of the struct doesn't have 1384 * uninitialized padding. All padding has to be explicit 1385 * and has to be set to zero by the userspace to make 1386 * struct xdp_umem_reg extensible in the future. 1387 */ 1388 BUILD_BUG_ON(offsetof(struct xdp_umem_reg, tx_metadata_len) + 1389 sizeof_field(struct xdp_umem_reg, tx_metadata_len) != 1390 sizeof(struct xdp_umem_reg)); 1391 1392 if (copy_from_sockptr(&mr, optval, mr_size)) 1393 return -EFAULT; 1394 1395 mutex_lock(&xs->mutex); 1396 if (xs->state != XSK_READY || xs->umem) { 1397 mutex_unlock(&xs->mutex); 1398 return -EBUSY; 1399 } 1400 1401 umem = xdp_umem_create(&mr); 1402 if (IS_ERR(umem)) { 1403 mutex_unlock(&xs->mutex); 1404 return PTR_ERR(umem); 1405 } 1406 1407 /* Make sure umem is ready before it can be seen by others */ 1408 smp_wmb(); 1409 WRITE_ONCE(xs->umem, umem); 1410 mutex_unlock(&xs->mutex); 1411 return 0; 1412 } 1413 case XDP_UMEM_FILL_RING: 1414 case XDP_UMEM_COMPLETION_RING: 1415 { 1416 struct xsk_queue **q; 1417 int entries; 1418 1419 if (optlen < sizeof(entries)) 1420 return -EINVAL; 1421 if (copy_from_sockptr(&entries, optval, sizeof(entries))) 1422 return -EFAULT; 1423 1424 mutex_lock(&xs->mutex); 1425 if (xs->state != XSK_READY) { 1426 mutex_unlock(&xs->mutex); 1427 return -EBUSY; 1428 } 1429 1430 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp : 1431 &xs->cq_tmp; 1432 err = xsk_init_queue(entries, q, true); 1433 mutex_unlock(&xs->mutex); 1434 return err; 1435 } 1436 default: 1437 break; 1438 } 1439 1440 return -ENOPROTOOPT; 1441 } 1442 1443 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring) 1444 { 1445 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer); 1446 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer); 1447 ring->desc = offsetof(struct xdp_rxtx_ring, desc); 1448 } 1449 1450 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring) 1451 { 1452 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer); 1453 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer); 1454 ring->desc = offsetof(struct xdp_umem_ring, desc); 1455 } 1456 1457 struct xdp_statistics_v1 { 1458 __u64 rx_dropped; 1459 __u64 rx_invalid_descs; 1460 __u64 tx_invalid_descs; 1461 }; 1462 1463 static int xsk_getsockopt(struct socket *sock, int level, int optname, 1464 char __user *optval, int __user *optlen) 1465 { 1466 struct sock *sk = sock->sk; 1467 struct xdp_sock *xs = xdp_sk(sk); 1468 int len; 1469 1470 if (level != SOL_XDP) 1471 return -ENOPROTOOPT; 1472 1473 if (get_user(len, optlen)) 1474 return -EFAULT; 1475 if (len < 0) 1476 return -EINVAL; 1477 1478 switch (optname) { 1479 case XDP_STATISTICS: 1480 { 1481 struct xdp_statistics stats = {}; 1482 bool extra_stats = true; 1483 size_t stats_size; 1484 1485 if (len < sizeof(struct xdp_statistics_v1)) { 1486 return -EINVAL; 1487 } else if (len < sizeof(stats)) { 1488 extra_stats = false; 1489 stats_size = sizeof(struct xdp_statistics_v1); 1490 } else { 1491 stats_size = sizeof(stats); 1492 } 1493 1494 mutex_lock(&xs->mutex); 1495 stats.rx_dropped = xs->rx_dropped; 1496 if (extra_stats) { 1497 stats.rx_ring_full = xs->rx_queue_full; 1498 stats.rx_fill_ring_empty_descs = 1499 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0; 1500 stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx); 1501 } else { 1502 stats.rx_dropped += xs->rx_queue_full; 1503 } 1504 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx); 1505 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx); 1506 mutex_unlock(&xs->mutex); 1507 1508 if (copy_to_user(optval, &stats, stats_size)) 1509 return -EFAULT; 1510 if (put_user(stats_size, optlen)) 1511 return -EFAULT; 1512 1513 return 0; 1514 } 1515 case XDP_MMAP_OFFSETS: 1516 { 1517 struct xdp_mmap_offsets off; 1518 struct xdp_mmap_offsets_v1 off_v1; 1519 bool flags_supported = true; 1520 void *to_copy; 1521 1522 if (len < sizeof(off_v1)) 1523 return -EINVAL; 1524 else if (len < sizeof(off)) 1525 flags_supported = false; 1526 1527 if (flags_supported) { 1528 /* xdp_ring_offset is identical to xdp_ring_offset_v1 1529 * except for the flags field added to the end. 1530 */ 1531 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 1532 &off.rx); 1533 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 1534 &off.tx); 1535 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 1536 &off.fr); 1537 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 1538 &off.cr); 1539 off.rx.flags = offsetof(struct xdp_rxtx_ring, 1540 ptrs.flags); 1541 off.tx.flags = offsetof(struct xdp_rxtx_ring, 1542 ptrs.flags); 1543 off.fr.flags = offsetof(struct xdp_umem_ring, 1544 ptrs.flags); 1545 off.cr.flags = offsetof(struct xdp_umem_ring, 1546 ptrs.flags); 1547 1548 len = sizeof(off); 1549 to_copy = &off; 1550 } else { 1551 xsk_enter_rxtx_offsets(&off_v1.rx); 1552 xsk_enter_rxtx_offsets(&off_v1.tx); 1553 xsk_enter_umem_offsets(&off_v1.fr); 1554 xsk_enter_umem_offsets(&off_v1.cr); 1555 1556 len = sizeof(off_v1); 1557 to_copy = &off_v1; 1558 } 1559 1560 if (copy_to_user(optval, to_copy, len)) 1561 return -EFAULT; 1562 if (put_user(len, optlen)) 1563 return -EFAULT; 1564 1565 return 0; 1566 } 1567 case XDP_OPTIONS: 1568 { 1569 struct xdp_options opts = {}; 1570 1571 if (len < sizeof(opts)) 1572 return -EINVAL; 1573 1574 mutex_lock(&xs->mutex); 1575 if (xs->zc) 1576 opts.flags |= XDP_OPTIONS_ZEROCOPY; 1577 mutex_unlock(&xs->mutex); 1578 1579 len = sizeof(opts); 1580 if (copy_to_user(optval, &opts, len)) 1581 return -EFAULT; 1582 if (put_user(len, optlen)) 1583 return -EFAULT; 1584 1585 return 0; 1586 } 1587 default: 1588 break; 1589 } 1590 1591 return -EOPNOTSUPP; 1592 } 1593 1594 static int xsk_mmap(struct file *file, struct socket *sock, 1595 struct vm_area_struct *vma) 1596 { 1597 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT; 1598 unsigned long size = vma->vm_end - vma->vm_start; 1599 struct xdp_sock *xs = xdp_sk(sock->sk); 1600 int state = READ_ONCE(xs->state); 1601 struct xsk_queue *q = NULL; 1602 1603 if (state != XSK_READY && state != XSK_BOUND) 1604 return -EBUSY; 1605 1606 if (offset == XDP_PGOFF_RX_RING) { 1607 q = READ_ONCE(xs->rx); 1608 } else if (offset == XDP_PGOFF_TX_RING) { 1609 q = READ_ONCE(xs->tx); 1610 } else { 1611 /* Matches the smp_wmb() in XDP_UMEM_REG */ 1612 smp_rmb(); 1613 if (offset == XDP_UMEM_PGOFF_FILL_RING) 1614 q = state == XSK_READY ? READ_ONCE(xs->fq_tmp) : 1615 READ_ONCE(xs->pool->fq); 1616 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING) 1617 q = state == XSK_READY ? READ_ONCE(xs->cq_tmp) : 1618 READ_ONCE(xs->pool->cq); 1619 } 1620 1621 if (!q) 1622 return -EINVAL; 1623 1624 /* Matches the smp_wmb() in xsk_init_queue */ 1625 smp_rmb(); 1626 if (size > q->ring_vmalloc_size) 1627 return -EINVAL; 1628 1629 return remap_vmalloc_range(vma, q->ring, 0); 1630 } 1631 1632 static int xsk_notifier(struct notifier_block *this, 1633 unsigned long msg, void *ptr) 1634 { 1635 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 1636 struct net *net = dev_net(dev); 1637 struct sock *sk; 1638 1639 switch (msg) { 1640 case NETDEV_UNREGISTER: 1641 mutex_lock(&net->xdp.lock); 1642 sk_for_each(sk, &net->xdp.list) { 1643 struct xdp_sock *xs = xdp_sk(sk); 1644 1645 mutex_lock(&xs->mutex); 1646 if (xs->dev == dev) { 1647 sk->sk_err = ENETDOWN; 1648 if (!sock_flag(sk, SOCK_DEAD)) 1649 sk_error_report(sk); 1650 1651 xsk_unbind_dev(xs); 1652 1653 /* Clear device references. */ 1654 xp_clear_dev(xs->pool); 1655 } 1656 mutex_unlock(&xs->mutex); 1657 } 1658 mutex_unlock(&net->xdp.lock); 1659 break; 1660 } 1661 return NOTIFY_DONE; 1662 } 1663 1664 static struct proto xsk_proto = { 1665 .name = "XDP", 1666 .owner = THIS_MODULE, 1667 .obj_size = sizeof(struct xdp_sock), 1668 }; 1669 1670 static const struct proto_ops xsk_proto_ops = { 1671 .family = PF_XDP, 1672 .owner = THIS_MODULE, 1673 .release = xsk_release, 1674 .bind = xsk_bind, 1675 .connect = sock_no_connect, 1676 .socketpair = sock_no_socketpair, 1677 .accept = sock_no_accept, 1678 .getname = sock_no_getname, 1679 .poll = xsk_poll, 1680 .ioctl = sock_no_ioctl, 1681 .listen = sock_no_listen, 1682 .shutdown = sock_no_shutdown, 1683 .setsockopt = xsk_setsockopt, 1684 .getsockopt = xsk_getsockopt, 1685 .sendmsg = xsk_sendmsg, 1686 .recvmsg = xsk_recvmsg, 1687 .mmap = xsk_mmap, 1688 }; 1689 1690 static void xsk_destruct(struct sock *sk) 1691 { 1692 struct xdp_sock *xs = xdp_sk(sk); 1693 1694 if (!sock_flag(sk, SOCK_DEAD)) 1695 return; 1696 1697 if (!xp_put_pool(xs->pool)) 1698 xdp_put_umem(xs->umem, !xs->pool); 1699 } 1700 1701 static int xsk_create(struct net *net, struct socket *sock, int protocol, 1702 int kern) 1703 { 1704 struct xdp_sock *xs; 1705 struct sock *sk; 1706 1707 if (!ns_capable(net->user_ns, CAP_NET_RAW)) 1708 return -EPERM; 1709 if (sock->type != SOCK_RAW) 1710 return -ESOCKTNOSUPPORT; 1711 1712 if (protocol) 1713 return -EPROTONOSUPPORT; 1714 1715 sock->state = SS_UNCONNECTED; 1716 1717 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern); 1718 if (!sk) 1719 return -ENOBUFS; 1720 1721 sock->ops = &xsk_proto_ops; 1722 1723 sock_init_data(sock, sk); 1724 1725 sk->sk_family = PF_XDP; 1726 1727 sk->sk_destruct = xsk_destruct; 1728 1729 sock_set_flag(sk, SOCK_RCU_FREE); 1730 1731 xs = xdp_sk(sk); 1732 xs->state = XSK_READY; 1733 mutex_init(&xs->mutex); 1734 spin_lock_init(&xs->rx_lock); 1735 1736 INIT_LIST_HEAD(&xs->map_list); 1737 spin_lock_init(&xs->map_list_lock); 1738 1739 mutex_lock(&net->xdp.lock); 1740 sk_add_node_rcu(sk, &net->xdp.list); 1741 mutex_unlock(&net->xdp.lock); 1742 1743 sock_prot_inuse_add(net, &xsk_proto, 1); 1744 1745 return 0; 1746 } 1747 1748 static const struct net_proto_family xsk_family_ops = { 1749 .family = PF_XDP, 1750 .create = xsk_create, 1751 .owner = THIS_MODULE, 1752 }; 1753 1754 static struct notifier_block xsk_netdev_notifier = { 1755 .notifier_call = xsk_notifier, 1756 }; 1757 1758 static int __net_init xsk_net_init(struct net *net) 1759 { 1760 mutex_init(&net->xdp.lock); 1761 INIT_HLIST_HEAD(&net->xdp.list); 1762 return 0; 1763 } 1764 1765 static void __net_exit xsk_net_exit(struct net *net) 1766 { 1767 WARN_ON_ONCE(!hlist_empty(&net->xdp.list)); 1768 } 1769 1770 static struct pernet_operations xsk_net_ops = { 1771 .init = xsk_net_init, 1772 .exit = xsk_net_exit, 1773 }; 1774 1775 static int __init xsk_init(void) 1776 { 1777 int err; 1778 1779 err = proto_register(&xsk_proto, 0 /* no slab */); 1780 if (err) 1781 goto out; 1782 1783 err = sock_register(&xsk_family_ops); 1784 if (err) 1785 goto out_proto; 1786 1787 err = register_pernet_subsys(&xsk_net_ops); 1788 if (err) 1789 goto out_sk; 1790 1791 err = register_netdevice_notifier(&xsk_netdev_notifier); 1792 if (err) 1793 goto out_pernet; 1794 1795 return 0; 1796 1797 out_pernet: 1798 unregister_pernet_subsys(&xsk_net_ops); 1799 out_sk: 1800 sock_unregister(PF_XDP); 1801 out_proto: 1802 proto_unregister(&xsk_proto); 1803 out: 1804 return err; 1805 } 1806 1807 fs_initcall(xsk_init); 1808