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 err = xsk_cq_reserve_addr_locked(xs->pool, desc.addr); 810 if (err) { 811 err = -EAGAIN; 812 goto out; 813 } 814 815 skb = xsk_build_skb(xs, &desc); 816 if (IS_ERR(skb)) { 817 err = PTR_ERR(skb); 818 if (err != -EOVERFLOW) 819 goto out; 820 err = 0; 821 continue; 822 } 823 824 xskq_cons_release(xs->tx); 825 826 if (xp_mb_desc(&desc)) { 827 xs->skb = skb; 828 continue; 829 } 830 831 err = __dev_direct_xmit(skb, xs->queue_id); 832 if (err == NETDEV_TX_BUSY) { 833 /* Tell user-space to retry the send */ 834 xskq_cons_cancel_n(xs->tx, xsk_get_num_desc(skb)); 835 xsk_consume_skb(skb); 836 err = -EAGAIN; 837 goto out; 838 } 839 840 /* Ignore NET_XMIT_CN as packet might have been sent */ 841 if (err == NET_XMIT_DROP) { 842 /* SKB completed but not sent */ 843 err = -EBUSY; 844 xs->skb = NULL; 845 goto out; 846 } 847 848 sent_frame = true; 849 xs->skb = NULL; 850 } 851 852 if (xskq_has_descs(xs->tx)) { 853 if (xs->skb) 854 xsk_drop_skb(xs->skb); 855 xskq_cons_release(xs->tx); 856 } 857 858 out: 859 if (sent_frame) 860 if (xsk_tx_writeable(xs)) 861 sk->sk_write_space(sk); 862 863 mutex_unlock(&xs->mutex); 864 return err; 865 } 866 867 static int xsk_generic_xmit(struct sock *sk) 868 { 869 int ret; 870 871 /* Drop the RCU lock since the SKB path might sleep. */ 872 rcu_read_unlock(); 873 ret = __xsk_generic_xmit(sk); 874 /* Reaquire RCU lock before going into common code. */ 875 rcu_read_lock(); 876 877 return ret; 878 } 879 880 static bool xsk_no_wakeup(struct sock *sk) 881 { 882 #ifdef CONFIG_NET_RX_BUSY_POLL 883 /* Prefer busy-polling, skip the wakeup. */ 884 return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) && 885 napi_id_valid(READ_ONCE(sk->sk_napi_id)); 886 #else 887 return false; 888 #endif 889 } 890 891 static int xsk_check_common(struct xdp_sock *xs) 892 { 893 if (unlikely(!xsk_is_bound(xs))) 894 return -ENXIO; 895 if (unlikely(!(xs->dev->flags & IFF_UP))) 896 return -ENETDOWN; 897 898 return 0; 899 } 900 901 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) 902 { 903 bool need_wait = !(m->msg_flags & MSG_DONTWAIT); 904 struct sock *sk = sock->sk; 905 struct xdp_sock *xs = xdp_sk(sk); 906 struct xsk_buff_pool *pool; 907 int err; 908 909 err = xsk_check_common(xs); 910 if (err) 911 return err; 912 if (unlikely(need_wait)) 913 return -EOPNOTSUPP; 914 if (unlikely(!xs->tx)) 915 return -ENOBUFS; 916 917 if (sk_can_busy_loop(sk)) 918 sk_busy_loop(sk, 1); /* only support non-blocking sockets */ 919 920 if (xs->zc && xsk_no_wakeup(sk)) 921 return 0; 922 923 pool = xs->pool; 924 if (pool->cached_need_wakeup & XDP_WAKEUP_TX) { 925 if (xs->zc) 926 return xsk_wakeup(xs, XDP_WAKEUP_TX); 927 return xsk_generic_xmit(sk); 928 } 929 return 0; 930 } 931 932 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len) 933 { 934 int ret; 935 936 rcu_read_lock(); 937 ret = __xsk_sendmsg(sock, m, total_len); 938 rcu_read_unlock(); 939 940 return ret; 941 } 942 943 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags) 944 { 945 bool need_wait = !(flags & MSG_DONTWAIT); 946 struct sock *sk = sock->sk; 947 struct xdp_sock *xs = xdp_sk(sk); 948 int err; 949 950 err = xsk_check_common(xs); 951 if (err) 952 return err; 953 if (unlikely(!xs->rx)) 954 return -ENOBUFS; 955 if (unlikely(need_wait)) 956 return -EOPNOTSUPP; 957 958 if (sk_can_busy_loop(sk)) 959 sk_busy_loop(sk, 1); /* only support non-blocking sockets */ 960 961 if (xsk_no_wakeup(sk)) 962 return 0; 963 964 if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc) 965 return xsk_wakeup(xs, XDP_WAKEUP_RX); 966 return 0; 967 } 968 969 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags) 970 { 971 int ret; 972 973 rcu_read_lock(); 974 ret = __xsk_recvmsg(sock, m, len, flags); 975 rcu_read_unlock(); 976 977 return ret; 978 } 979 980 static __poll_t xsk_poll(struct file *file, struct socket *sock, 981 struct poll_table_struct *wait) 982 { 983 __poll_t mask = 0; 984 struct sock *sk = sock->sk; 985 struct xdp_sock *xs = xdp_sk(sk); 986 struct xsk_buff_pool *pool; 987 988 sock_poll_wait(file, sock, wait); 989 990 rcu_read_lock(); 991 if (xsk_check_common(xs)) 992 goto out; 993 994 pool = xs->pool; 995 996 if (pool->cached_need_wakeup) { 997 if (xs->zc) 998 xsk_wakeup(xs, pool->cached_need_wakeup); 999 else if (xs->tx) 1000 /* Poll needs to drive Tx also in copy mode */ 1001 xsk_generic_xmit(sk); 1002 } 1003 1004 if (xs->rx && !xskq_prod_is_empty(xs->rx)) 1005 mask |= EPOLLIN | EPOLLRDNORM; 1006 if (xs->tx && xsk_tx_writeable(xs)) 1007 mask |= EPOLLOUT | EPOLLWRNORM; 1008 out: 1009 rcu_read_unlock(); 1010 return mask; 1011 } 1012 1013 static int xsk_init_queue(u32 entries, struct xsk_queue **queue, 1014 bool umem_queue) 1015 { 1016 struct xsk_queue *q; 1017 1018 if (entries == 0 || *queue || !is_power_of_2(entries)) 1019 return -EINVAL; 1020 1021 q = xskq_create(entries, umem_queue); 1022 if (!q) 1023 return -ENOMEM; 1024 1025 /* Make sure queue is ready before it can be seen by others */ 1026 smp_wmb(); 1027 WRITE_ONCE(*queue, q); 1028 return 0; 1029 } 1030 1031 static void xsk_unbind_dev(struct xdp_sock *xs) 1032 { 1033 struct net_device *dev = xs->dev; 1034 1035 if (xs->state != XSK_BOUND) 1036 return; 1037 WRITE_ONCE(xs->state, XSK_UNBOUND); 1038 1039 /* Wait for driver to stop using the xdp socket. */ 1040 xp_del_xsk(xs->pool, xs); 1041 synchronize_net(); 1042 dev_put(dev); 1043 } 1044 1045 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs, 1046 struct xdp_sock __rcu ***map_entry) 1047 { 1048 struct xsk_map *map = NULL; 1049 struct xsk_map_node *node; 1050 1051 *map_entry = NULL; 1052 1053 spin_lock_bh(&xs->map_list_lock); 1054 node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node, 1055 node); 1056 if (node) { 1057 bpf_map_inc(&node->map->map); 1058 map = node->map; 1059 *map_entry = node->map_entry; 1060 } 1061 spin_unlock_bh(&xs->map_list_lock); 1062 return map; 1063 } 1064 1065 static void xsk_delete_from_maps(struct xdp_sock *xs) 1066 { 1067 /* This function removes the current XDP socket from all the 1068 * maps it resides in. We need to take extra care here, due to 1069 * the two locks involved. Each map has a lock synchronizing 1070 * updates to the entries, and each socket has a lock that 1071 * synchronizes access to the list of maps (map_list). For 1072 * deadlock avoidance the locks need to be taken in the order 1073 * "map lock"->"socket map list lock". We start off by 1074 * accessing the socket map list, and take a reference to the 1075 * map to guarantee existence between the 1076 * xsk_get_map_list_entry() and xsk_map_try_sock_delete() 1077 * calls. Then we ask the map to remove the socket, which 1078 * tries to remove the socket from the map. Note that there 1079 * might be updates to the map between 1080 * xsk_get_map_list_entry() and xsk_map_try_sock_delete(). 1081 */ 1082 struct xdp_sock __rcu **map_entry = NULL; 1083 struct xsk_map *map; 1084 1085 while ((map = xsk_get_map_list_entry(xs, &map_entry))) { 1086 xsk_map_try_sock_delete(map, xs, map_entry); 1087 bpf_map_put(&map->map); 1088 } 1089 } 1090 1091 static int xsk_release(struct socket *sock) 1092 { 1093 struct sock *sk = sock->sk; 1094 struct xdp_sock *xs = xdp_sk(sk); 1095 struct net *net; 1096 1097 if (!sk) 1098 return 0; 1099 1100 net = sock_net(sk); 1101 1102 if (xs->skb) 1103 xsk_drop_skb(xs->skb); 1104 1105 mutex_lock(&net->xdp.lock); 1106 sk_del_node_init_rcu(sk); 1107 mutex_unlock(&net->xdp.lock); 1108 1109 sock_prot_inuse_add(net, sk->sk_prot, -1); 1110 1111 xsk_delete_from_maps(xs); 1112 mutex_lock(&xs->mutex); 1113 xsk_unbind_dev(xs); 1114 mutex_unlock(&xs->mutex); 1115 1116 xskq_destroy(xs->rx); 1117 xskq_destroy(xs->tx); 1118 xskq_destroy(xs->fq_tmp); 1119 xskq_destroy(xs->cq_tmp); 1120 1121 sock_orphan(sk); 1122 sock->sk = NULL; 1123 1124 sock_put(sk); 1125 1126 return 0; 1127 } 1128 1129 static struct socket *xsk_lookup_xsk_from_fd(int fd) 1130 { 1131 struct socket *sock; 1132 int err; 1133 1134 sock = sockfd_lookup(fd, &err); 1135 if (!sock) 1136 return ERR_PTR(-ENOTSOCK); 1137 1138 if (sock->sk->sk_family != PF_XDP) { 1139 sockfd_put(sock); 1140 return ERR_PTR(-ENOPROTOOPT); 1141 } 1142 1143 return sock; 1144 } 1145 1146 static bool xsk_validate_queues(struct xdp_sock *xs) 1147 { 1148 return xs->fq_tmp && xs->cq_tmp; 1149 } 1150 1151 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len) 1152 { 1153 struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr; 1154 struct sock *sk = sock->sk; 1155 struct xdp_sock *xs = xdp_sk(sk); 1156 struct net_device *dev; 1157 int bound_dev_if; 1158 u32 flags, qid; 1159 int err = 0; 1160 1161 if (addr_len < sizeof(struct sockaddr_xdp)) 1162 return -EINVAL; 1163 if (sxdp->sxdp_family != AF_XDP) 1164 return -EINVAL; 1165 1166 flags = sxdp->sxdp_flags; 1167 if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY | 1168 XDP_USE_NEED_WAKEUP | XDP_USE_SG)) 1169 return -EINVAL; 1170 1171 bound_dev_if = READ_ONCE(sk->sk_bound_dev_if); 1172 if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex) 1173 return -EINVAL; 1174 1175 rtnl_lock(); 1176 mutex_lock(&xs->mutex); 1177 if (xs->state != XSK_READY) { 1178 err = -EBUSY; 1179 goto out_release; 1180 } 1181 1182 dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex); 1183 if (!dev) { 1184 err = -ENODEV; 1185 goto out_release; 1186 } 1187 1188 netdev_lock_ops(dev); 1189 1190 if (!xs->rx && !xs->tx) { 1191 err = -EINVAL; 1192 goto out_unlock; 1193 } 1194 1195 qid = sxdp->sxdp_queue_id; 1196 1197 if (flags & XDP_SHARED_UMEM) { 1198 struct xdp_sock *umem_xs; 1199 struct socket *sock; 1200 1201 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) || 1202 (flags & XDP_USE_NEED_WAKEUP) || (flags & XDP_USE_SG)) { 1203 /* Cannot specify flags for shared sockets. */ 1204 err = -EINVAL; 1205 goto out_unlock; 1206 } 1207 1208 if (xs->umem) { 1209 /* We have already our own. */ 1210 err = -EINVAL; 1211 goto out_unlock; 1212 } 1213 1214 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd); 1215 if (IS_ERR(sock)) { 1216 err = PTR_ERR(sock); 1217 goto out_unlock; 1218 } 1219 1220 umem_xs = xdp_sk(sock->sk); 1221 if (!xsk_is_bound(umem_xs)) { 1222 err = -EBADF; 1223 sockfd_put(sock); 1224 goto out_unlock; 1225 } 1226 1227 if (umem_xs->queue_id != qid || umem_xs->dev != dev) { 1228 /* Share the umem with another socket on another qid 1229 * and/or device. 1230 */ 1231 xs->pool = xp_create_and_assign_umem(xs, 1232 umem_xs->umem); 1233 if (!xs->pool) { 1234 err = -ENOMEM; 1235 sockfd_put(sock); 1236 goto out_unlock; 1237 } 1238 1239 err = xp_assign_dev_shared(xs->pool, umem_xs, dev, 1240 qid); 1241 if (err) { 1242 xp_destroy(xs->pool); 1243 xs->pool = NULL; 1244 sockfd_put(sock); 1245 goto out_unlock; 1246 } 1247 } else { 1248 /* Share the buffer pool with the other socket. */ 1249 if (xs->fq_tmp || xs->cq_tmp) { 1250 /* Do not allow setting your own fq or cq. */ 1251 err = -EINVAL; 1252 sockfd_put(sock); 1253 goto out_unlock; 1254 } 1255 1256 xp_get_pool(umem_xs->pool); 1257 xs->pool = umem_xs->pool; 1258 1259 /* If underlying shared umem was created without Tx 1260 * ring, allocate Tx descs array that Tx batching API 1261 * utilizes 1262 */ 1263 if (xs->tx && !xs->pool->tx_descs) { 1264 err = xp_alloc_tx_descs(xs->pool, xs); 1265 if (err) { 1266 xp_put_pool(xs->pool); 1267 xs->pool = NULL; 1268 sockfd_put(sock); 1269 goto out_unlock; 1270 } 1271 } 1272 } 1273 1274 xdp_get_umem(umem_xs->umem); 1275 WRITE_ONCE(xs->umem, umem_xs->umem); 1276 sockfd_put(sock); 1277 } else if (!xs->umem || !xsk_validate_queues(xs)) { 1278 err = -EINVAL; 1279 goto out_unlock; 1280 } else { 1281 /* This xsk has its own umem. */ 1282 xs->pool = xp_create_and_assign_umem(xs, xs->umem); 1283 if (!xs->pool) { 1284 err = -ENOMEM; 1285 goto out_unlock; 1286 } 1287 1288 err = xp_assign_dev(xs->pool, dev, qid, flags); 1289 if (err) { 1290 xp_destroy(xs->pool); 1291 xs->pool = NULL; 1292 goto out_unlock; 1293 } 1294 } 1295 1296 /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */ 1297 xs->fq_tmp = NULL; 1298 xs->cq_tmp = NULL; 1299 1300 xs->dev = dev; 1301 xs->zc = xs->umem->zc; 1302 xs->sg = !!(xs->umem->flags & XDP_UMEM_SG_FLAG); 1303 xs->queue_id = qid; 1304 xp_add_xsk(xs->pool, xs); 1305 1306 if (xs->zc && qid < dev->real_num_rx_queues) { 1307 struct netdev_rx_queue *rxq; 1308 1309 rxq = __netif_get_rx_queue(dev, qid); 1310 if (rxq->napi) 1311 __sk_mark_napi_id_once(sk, rxq->napi->napi_id); 1312 } 1313 1314 out_unlock: 1315 if (err) { 1316 dev_put(dev); 1317 } else { 1318 /* Matches smp_rmb() in bind() for shared umem 1319 * sockets, and xsk_is_bound(). 1320 */ 1321 smp_wmb(); 1322 WRITE_ONCE(xs->state, XSK_BOUND); 1323 } 1324 netdev_unlock_ops(dev); 1325 out_release: 1326 mutex_unlock(&xs->mutex); 1327 rtnl_unlock(); 1328 return err; 1329 } 1330 1331 struct xdp_umem_reg_v1 { 1332 __u64 addr; /* Start of packet data area */ 1333 __u64 len; /* Length of packet data area */ 1334 __u32 chunk_size; 1335 __u32 headroom; 1336 }; 1337 1338 static int xsk_setsockopt(struct socket *sock, int level, int optname, 1339 sockptr_t optval, unsigned int optlen) 1340 { 1341 struct sock *sk = sock->sk; 1342 struct xdp_sock *xs = xdp_sk(sk); 1343 int err; 1344 1345 if (level != SOL_XDP) 1346 return -ENOPROTOOPT; 1347 1348 switch (optname) { 1349 case XDP_RX_RING: 1350 case XDP_TX_RING: 1351 { 1352 struct xsk_queue **q; 1353 int entries; 1354 1355 if (optlen < sizeof(entries)) 1356 return -EINVAL; 1357 if (copy_from_sockptr(&entries, optval, sizeof(entries))) 1358 return -EFAULT; 1359 1360 mutex_lock(&xs->mutex); 1361 if (xs->state != XSK_READY) { 1362 mutex_unlock(&xs->mutex); 1363 return -EBUSY; 1364 } 1365 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx; 1366 err = xsk_init_queue(entries, q, false); 1367 if (!err && optname == XDP_TX_RING) 1368 /* Tx needs to be explicitly woken up the first time */ 1369 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP; 1370 mutex_unlock(&xs->mutex); 1371 return err; 1372 } 1373 case XDP_UMEM_REG: 1374 { 1375 size_t mr_size = sizeof(struct xdp_umem_reg); 1376 struct xdp_umem_reg mr = {}; 1377 struct xdp_umem *umem; 1378 1379 if (optlen < sizeof(struct xdp_umem_reg_v1)) 1380 return -EINVAL; 1381 else if (optlen < sizeof(mr)) 1382 mr_size = sizeof(struct xdp_umem_reg_v1); 1383 1384 BUILD_BUG_ON(sizeof(struct xdp_umem_reg_v1) >= sizeof(struct xdp_umem_reg)); 1385 1386 /* Make sure the last field of the struct doesn't have 1387 * uninitialized padding. All padding has to be explicit 1388 * and has to be set to zero by the userspace to make 1389 * struct xdp_umem_reg extensible in the future. 1390 */ 1391 BUILD_BUG_ON(offsetof(struct xdp_umem_reg, tx_metadata_len) + 1392 sizeof_field(struct xdp_umem_reg, tx_metadata_len) != 1393 sizeof(struct xdp_umem_reg)); 1394 1395 if (copy_from_sockptr(&mr, optval, mr_size)) 1396 return -EFAULT; 1397 1398 mutex_lock(&xs->mutex); 1399 if (xs->state != XSK_READY || xs->umem) { 1400 mutex_unlock(&xs->mutex); 1401 return -EBUSY; 1402 } 1403 1404 umem = xdp_umem_create(&mr); 1405 if (IS_ERR(umem)) { 1406 mutex_unlock(&xs->mutex); 1407 return PTR_ERR(umem); 1408 } 1409 1410 /* Make sure umem is ready before it can be seen by others */ 1411 smp_wmb(); 1412 WRITE_ONCE(xs->umem, umem); 1413 mutex_unlock(&xs->mutex); 1414 return 0; 1415 } 1416 case XDP_UMEM_FILL_RING: 1417 case XDP_UMEM_COMPLETION_RING: 1418 { 1419 struct xsk_queue **q; 1420 int entries; 1421 1422 if (optlen < sizeof(entries)) 1423 return -EINVAL; 1424 if (copy_from_sockptr(&entries, optval, sizeof(entries))) 1425 return -EFAULT; 1426 1427 mutex_lock(&xs->mutex); 1428 if (xs->state != XSK_READY) { 1429 mutex_unlock(&xs->mutex); 1430 return -EBUSY; 1431 } 1432 1433 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp : 1434 &xs->cq_tmp; 1435 err = xsk_init_queue(entries, q, true); 1436 mutex_unlock(&xs->mutex); 1437 return err; 1438 } 1439 default: 1440 break; 1441 } 1442 1443 return -ENOPROTOOPT; 1444 } 1445 1446 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring) 1447 { 1448 ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer); 1449 ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer); 1450 ring->desc = offsetof(struct xdp_rxtx_ring, desc); 1451 } 1452 1453 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring) 1454 { 1455 ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer); 1456 ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer); 1457 ring->desc = offsetof(struct xdp_umem_ring, desc); 1458 } 1459 1460 struct xdp_statistics_v1 { 1461 __u64 rx_dropped; 1462 __u64 rx_invalid_descs; 1463 __u64 tx_invalid_descs; 1464 }; 1465 1466 static int xsk_getsockopt(struct socket *sock, int level, int optname, 1467 char __user *optval, int __user *optlen) 1468 { 1469 struct sock *sk = sock->sk; 1470 struct xdp_sock *xs = xdp_sk(sk); 1471 int len; 1472 1473 if (level != SOL_XDP) 1474 return -ENOPROTOOPT; 1475 1476 if (get_user(len, optlen)) 1477 return -EFAULT; 1478 if (len < 0) 1479 return -EINVAL; 1480 1481 switch (optname) { 1482 case XDP_STATISTICS: 1483 { 1484 struct xdp_statistics stats = {}; 1485 bool extra_stats = true; 1486 size_t stats_size; 1487 1488 if (len < sizeof(struct xdp_statistics_v1)) { 1489 return -EINVAL; 1490 } else if (len < sizeof(stats)) { 1491 extra_stats = false; 1492 stats_size = sizeof(struct xdp_statistics_v1); 1493 } else { 1494 stats_size = sizeof(stats); 1495 } 1496 1497 mutex_lock(&xs->mutex); 1498 stats.rx_dropped = xs->rx_dropped; 1499 if (extra_stats) { 1500 stats.rx_ring_full = xs->rx_queue_full; 1501 stats.rx_fill_ring_empty_descs = 1502 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0; 1503 stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx); 1504 } else { 1505 stats.rx_dropped += xs->rx_queue_full; 1506 } 1507 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx); 1508 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx); 1509 mutex_unlock(&xs->mutex); 1510 1511 if (copy_to_user(optval, &stats, stats_size)) 1512 return -EFAULT; 1513 if (put_user(stats_size, optlen)) 1514 return -EFAULT; 1515 1516 return 0; 1517 } 1518 case XDP_MMAP_OFFSETS: 1519 { 1520 struct xdp_mmap_offsets off; 1521 struct xdp_mmap_offsets_v1 off_v1; 1522 bool flags_supported = true; 1523 void *to_copy; 1524 1525 if (len < sizeof(off_v1)) 1526 return -EINVAL; 1527 else if (len < sizeof(off)) 1528 flags_supported = false; 1529 1530 if (flags_supported) { 1531 /* xdp_ring_offset is identical to xdp_ring_offset_v1 1532 * except for the flags field added to the end. 1533 */ 1534 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 1535 &off.rx); 1536 xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *) 1537 &off.tx); 1538 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 1539 &off.fr); 1540 xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *) 1541 &off.cr); 1542 off.rx.flags = offsetof(struct xdp_rxtx_ring, 1543 ptrs.flags); 1544 off.tx.flags = offsetof(struct xdp_rxtx_ring, 1545 ptrs.flags); 1546 off.fr.flags = offsetof(struct xdp_umem_ring, 1547 ptrs.flags); 1548 off.cr.flags = offsetof(struct xdp_umem_ring, 1549 ptrs.flags); 1550 1551 len = sizeof(off); 1552 to_copy = &off; 1553 } else { 1554 xsk_enter_rxtx_offsets(&off_v1.rx); 1555 xsk_enter_rxtx_offsets(&off_v1.tx); 1556 xsk_enter_umem_offsets(&off_v1.fr); 1557 xsk_enter_umem_offsets(&off_v1.cr); 1558 1559 len = sizeof(off_v1); 1560 to_copy = &off_v1; 1561 } 1562 1563 if (copy_to_user(optval, to_copy, len)) 1564 return -EFAULT; 1565 if (put_user(len, optlen)) 1566 return -EFAULT; 1567 1568 return 0; 1569 } 1570 case XDP_OPTIONS: 1571 { 1572 struct xdp_options opts = {}; 1573 1574 if (len < sizeof(opts)) 1575 return -EINVAL; 1576 1577 mutex_lock(&xs->mutex); 1578 if (xs->zc) 1579 opts.flags |= XDP_OPTIONS_ZEROCOPY; 1580 mutex_unlock(&xs->mutex); 1581 1582 len = sizeof(opts); 1583 if (copy_to_user(optval, &opts, len)) 1584 return -EFAULT; 1585 if (put_user(len, optlen)) 1586 return -EFAULT; 1587 1588 return 0; 1589 } 1590 default: 1591 break; 1592 } 1593 1594 return -EOPNOTSUPP; 1595 } 1596 1597 static int xsk_mmap(struct file *file, struct socket *sock, 1598 struct vm_area_struct *vma) 1599 { 1600 loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT; 1601 unsigned long size = vma->vm_end - vma->vm_start; 1602 struct xdp_sock *xs = xdp_sk(sock->sk); 1603 int state = READ_ONCE(xs->state); 1604 struct xsk_queue *q = NULL; 1605 1606 if (state != XSK_READY && state != XSK_BOUND) 1607 return -EBUSY; 1608 1609 if (offset == XDP_PGOFF_RX_RING) { 1610 q = READ_ONCE(xs->rx); 1611 } else if (offset == XDP_PGOFF_TX_RING) { 1612 q = READ_ONCE(xs->tx); 1613 } else { 1614 /* Matches the smp_wmb() in XDP_UMEM_REG */ 1615 smp_rmb(); 1616 if (offset == XDP_UMEM_PGOFF_FILL_RING) 1617 q = state == XSK_READY ? READ_ONCE(xs->fq_tmp) : 1618 READ_ONCE(xs->pool->fq); 1619 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING) 1620 q = state == XSK_READY ? READ_ONCE(xs->cq_tmp) : 1621 READ_ONCE(xs->pool->cq); 1622 } 1623 1624 if (!q) 1625 return -EINVAL; 1626 1627 /* Matches the smp_wmb() in xsk_init_queue */ 1628 smp_rmb(); 1629 if (size > q->ring_vmalloc_size) 1630 return -EINVAL; 1631 1632 return remap_vmalloc_range(vma, q->ring, 0); 1633 } 1634 1635 static int xsk_notifier(struct notifier_block *this, 1636 unsigned long msg, void *ptr) 1637 { 1638 struct net_device *dev = netdev_notifier_info_to_dev(ptr); 1639 struct net *net = dev_net(dev); 1640 struct sock *sk; 1641 1642 switch (msg) { 1643 case NETDEV_UNREGISTER: 1644 mutex_lock(&net->xdp.lock); 1645 sk_for_each(sk, &net->xdp.list) { 1646 struct xdp_sock *xs = xdp_sk(sk); 1647 1648 mutex_lock(&xs->mutex); 1649 if (xs->dev == dev) { 1650 sk->sk_err = ENETDOWN; 1651 if (!sock_flag(sk, SOCK_DEAD)) 1652 sk_error_report(sk); 1653 1654 xsk_unbind_dev(xs); 1655 1656 /* Clear device references. */ 1657 xp_clear_dev(xs->pool); 1658 } 1659 mutex_unlock(&xs->mutex); 1660 } 1661 mutex_unlock(&net->xdp.lock); 1662 break; 1663 } 1664 return NOTIFY_DONE; 1665 } 1666 1667 static struct proto xsk_proto = { 1668 .name = "XDP", 1669 .owner = THIS_MODULE, 1670 .obj_size = sizeof(struct xdp_sock), 1671 }; 1672 1673 static const struct proto_ops xsk_proto_ops = { 1674 .family = PF_XDP, 1675 .owner = THIS_MODULE, 1676 .release = xsk_release, 1677 .bind = xsk_bind, 1678 .connect = sock_no_connect, 1679 .socketpair = sock_no_socketpair, 1680 .accept = sock_no_accept, 1681 .getname = sock_no_getname, 1682 .poll = xsk_poll, 1683 .ioctl = sock_no_ioctl, 1684 .listen = sock_no_listen, 1685 .shutdown = sock_no_shutdown, 1686 .setsockopt = xsk_setsockopt, 1687 .getsockopt = xsk_getsockopt, 1688 .sendmsg = xsk_sendmsg, 1689 .recvmsg = xsk_recvmsg, 1690 .mmap = xsk_mmap, 1691 }; 1692 1693 static void xsk_destruct(struct sock *sk) 1694 { 1695 struct xdp_sock *xs = xdp_sk(sk); 1696 1697 if (!sock_flag(sk, SOCK_DEAD)) 1698 return; 1699 1700 if (!xp_put_pool(xs->pool)) 1701 xdp_put_umem(xs->umem, !xs->pool); 1702 } 1703 1704 static int xsk_create(struct net *net, struct socket *sock, int protocol, 1705 int kern) 1706 { 1707 struct xdp_sock *xs; 1708 struct sock *sk; 1709 1710 if (!ns_capable(net->user_ns, CAP_NET_RAW)) 1711 return -EPERM; 1712 if (sock->type != SOCK_RAW) 1713 return -ESOCKTNOSUPPORT; 1714 1715 if (protocol) 1716 return -EPROTONOSUPPORT; 1717 1718 sock->state = SS_UNCONNECTED; 1719 1720 sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern); 1721 if (!sk) 1722 return -ENOBUFS; 1723 1724 sock->ops = &xsk_proto_ops; 1725 1726 sock_init_data(sock, sk); 1727 1728 sk->sk_family = PF_XDP; 1729 1730 sk->sk_destruct = xsk_destruct; 1731 1732 sock_set_flag(sk, SOCK_RCU_FREE); 1733 1734 xs = xdp_sk(sk); 1735 xs->state = XSK_READY; 1736 mutex_init(&xs->mutex); 1737 spin_lock_init(&xs->rx_lock); 1738 1739 INIT_LIST_HEAD(&xs->map_list); 1740 spin_lock_init(&xs->map_list_lock); 1741 1742 mutex_lock(&net->xdp.lock); 1743 sk_add_node_rcu(sk, &net->xdp.list); 1744 mutex_unlock(&net->xdp.lock); 1745 1746 sock_prot_inuse_add(net, &xsk_proto, 1); 1747 1748 return 0; 1749 } 1750 1751 static const struct net_proto_family xsk_family_ops = { 1752 .family = PF_XDP, 1753 .create = xsk_create, 1754 .owner = THIS_MODULE, 1755 }; 1756 1757 static struct notifier_block xsk_netdev_notifier = { 1758 .notifier_call = xsk_notifier, 1759 }; 1760 1761 static int __net_init xsk_net_init(struct net *net) 1762 { 1763 mutex_init(&net->xdp.lock); 1764 INIT_HLIST_HEAD(&net->xdp.list); 1765 return 0; 1766 } 1767 1768 static void __net_exit xsk_net_exit(struct net *net) 1769 { 1770 WARN_ON_ONCE(!hlist_empty(&net->xdp.list)); 1771 } 1772 1773 static struct pernet_operations xsk_net_ops = { 1774 .init = xsk_net_init, 1775 .exit = xsk_net_exit, 1776 }; 1777 1778 static int __init xsk_init(void) 1779 { 1780 int err; 1781 1782 err = proto_register(&xsk_proto, 0 /* no slab */); 1783 if (err) 1784 goto out; 1785 1786 err = sock_register(&xsk_family_ops); 1787 if (err) 1788 goto out_proto; 1789 1790 err = register_pernet_subsys(&xsk_net_ops); 1791 if (err) 1792 goto out_sk; 1793 1794 err = register_netdevice_notifier(&xsk_netdev_notifier); 1795 if (err) 1796 goto out_pernet; 1797 1798 return 0; 1799 1800 out_pernet: 1801 unregister_pernet_subsys(&xsk_net_ops); 1802 out_sk: 1803 sock_unregister(PF_XDP); 1804 out_proto: 1805 proto_unregister(&xsk_proto); 1806 out: 1807 return err; 1808 } 1809 1810 fs_initcall(xsk_init); 1811