net/xdp/xsk.c

   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 <[email protected]>
   9  *            Magnus Karlsson <[email protected]>
  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 32
  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 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
 138 {
 139         struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
 140         u64 addr;
 141         int err;
 142
 143         addr = xp_get_handle(xskb);
 144         err = xskq_prod_reserve_desc(xs->rx, addr, len);
 145         if (err) {
 146                 xs->rx_queue_full++;
 147                 return err;
 148         }
 149
 150         xp_release(xskb);
 151         return 0;
 152 }
 153
 154 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len)
 155 {
 156         void *from_buf, *to_buf;
 157         u32 metalen;
 158
 159         if (unlikely(xdp_data_meta_unsupported(from))) {
 160                 from_buf = from->data;
 161                 to_buf = to->data;
 162                 metalen = 0;
 163         } else {
 164                 from_buf = from->data_meta;
 165                 metalen = from->data - from->data_meta;
 166                 to_buf = to->data - metalen;
 167         }
 168
 169         memcpy(to_buf, from_buf, len + metalen);
 170 }
 171
 172 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 173 {
 174         struct xdp_buff *xsk_xdp;
 175         int err;
 176         u32 len;
 177
 178         len = xdp->data_end - xdp->data;
 179         if (len > xsk_pool_get_rx_frame_size(xs->pool)) {
 180                 xs->rx_dropped++;
 181                 return -ENOSPC;
 182         }
 183
 184         xsk_xdp = xsk_buff_alloc(xs->pool);
 185         if (!xsk_xdp) {
 186                 xs->rx_dropped++;
 187                 return -ENOSPC;
 188         }
 189
 190         xsk_copy_xdp(xsk_xdp, xdp, len);
 191         err = __xsk_rcv_zc(xs, xsk_xdp, len);
 192         if (err) {
 193                 xsk_buff_free(xsk_xdp);
 194                 return err;
 195         }
 196         return 0;
 197 }
 198
 199 static bool xsk_tx_writeable(struct xdp_sock *xs)
 200 {
 201         if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
 202                 return false;
 203
 204         return true;
 205 }
 206
 207 static bool xsk_is_bound(struct xdp_sock *xs)
 208 {
 209         if (READ_ONCE(xs->state) == XSK_BOUND) {
 210                 /* Matches smp_wmb() in bind(). */
 211                 smp_rmb();
 212                 return true;
 213         }
 214         return false;
 215 }
 216
 217 static int xsk_rcv_check(struct xdp_sock *xs, struct xdp_buff *xdp)
 218 {
 219         if (!xsk_is_bound(xs))
 220                 return -EINVAL;
 221
 222         if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
 223                 return -EINVAL;
 224
 225         sk_mark_napi_id_once_xdp(&xs->sk, xdp);
 226         return 0;
 227 }
 228
 229 static void xsk_flush(struct xdp_sock *xs)
 230 {
 231         xskq_prod_submit(xs->rx);
 232         __xskq_cons_release(xs->pool->fq);
 233         sock_def_readable(&xs->sk);
 234 }
 235
 236 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 237 {
 238         int err;
 239
 240         spin_lock_bh(&xs->rx_lock);
 241         err = xsk_rcv_check(xs, xdp);
 242         if (!err) {
 243                 err = __xsk_rcv(xs, xdp);
 244                 xsk_flush(xs);
 245         }
 246         spin_unlock_bh(&xs->rx_lock);
 247         return err;
 248 }
 249
 250 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 251 {
 252         int err;
 253         u32 len;
 254
 255         err = xsk_rcv_check(xs, xdp);
 256         if (err)
 257                 return err;
 258
 259         if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) {
 260                 len = xdp->data_end - xdp->data;
 261                 return __xsk_rcv_zc(xs, xdp, len);
 262         }
 263
 264         err = __xsk_rcv(xs, xdp);
 265         if (!err)
 266                 xdp_return_buff(xdp);
 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);
 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, u32 max_entries)
 347 {
 348         struct xdp_desc *descs = pool->tx_descs;
 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, u32 max_entries)
 359 {
 360         struct xdp_sock *xs;
 361         u32 nb_pkts;
 362
 363         rcu_read_lock();
 364         if (!list_is_singular(&pool->xsk_tx_list)) {
 365                 /* Fallback to the non-batched version */
 366                 rcu_read_unlock();
 367                 return xsk_tx_peek_release_fallback(pool, max_entries);
 368         }
 369
 370         xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
 371         if (!xs) {
 372                 nb_pkts = 0;
 373                 goto out;
 374         }
 375
 376         nb_pkts = xskq_cons_peek_desc_batch(xs->tx, pool, max_entries);
 377         if (!nb_pkts) {
 378                 xs->tx->queue_empty_descs++;
 379                 goto out;
 380         }
 381
 382         /* This is the backpressure mechanism for the Tx path. Try to
 383          * reserve space in the completion queue for all packets, but
 384          * if there are fewer slots available, just process that many
 385          * packets. This avoids having to implement any buffering in
 386          * the Tx path.
 387          */
 388         nb_pkts = xskq_prod_reserve_addr_batch(pool->cq, pool->tx_descs, nb_pkts);
 389         if (!nb_pkts)
 390                 goto out;
 391
 392         xskq_cons_release_n(xs->tx, nb_pkts);
 393         __xskq_cons_release(xs->tx);
 394         xs->sk.sk_write_space(&xs->sk);
 395
 396 out:
 397         rcu_read_unlock();
 398         return nb_pkts;
 399 }
 400 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
 401
 402 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
 403 {
 404         struct net_device *dev = xs->dev;
 405         int err;
 406
 407         rcu_read_lock();
 408         err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 409         rcu_read_unlock();
 410
 411         return err;
 412 }
 413
 414 static int xsk_zc_xmit(struct xdp_sock *xs)
 415 {
 416         return xsk_wakeup(xs, XDP_WAKEUP_TX);
 417 }
 418
 419 static void xsk_destruct_skb(struct sk_buff *skb)
 420 {
 421         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 422         struct xdp_sock *xs = xdp_sk(skb->sk);
 423         unsigned long flags;
 424
 425         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 426         xskq_prod_submit_addr(xs->pool->cq, addr);
 427         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 428
 429         sock_wfree(skb);
 430 }
 431
 432 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
 433                                               struct xdp_desc *desc)
 434 {
 435         struct xsk_buff_pool *pool = xs->pool;
 436         u32 hr, len, ts, offset, copy, copied;
 437         struct sk_buff *skb;
 438         struct page *page;
 439         void *buffer;
 440         int err, i;
 441         u64 addr;
 442
 443         hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
 444
 445         skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
 446         if (unlikely(!skb))
 447                 return ERR_PTR(err);
 448
 449         skb_reserve(skb, hr);
 450
 451         addr = desc->addr;
 452         len = desc->len;
 453         ts = pool->unaligned ? len : pool->chunk_size;
 454
 455         buffer = xsk_buff_raw_get_data(pool, addr);
 456         offset = offset_in_page(buffer);
 457         addr = buffer - pool->addrs;
 458
 459         for (copied = 0, i = 0; copied < len; i++) {
 460                 page = pool->umem->pgs[addr >> PAGE_SHIFT];
 461                 get_page(page);
 462
 463                 copy = min_t(u32, PAGE_SIZE - offset, len - copied);
 464                 skb_fill_page_desc(skb, i, page, offset, copy);
 465
 466                 copied += copy;
 467                 addr += copy;
 468                 offset = 0;
 469         }
 470
 471         skb->len += len;
 472         skb->data_len += len;
 473         skb->truesize += ts;
 474
 475         refcount_add(ts, &xs->sk.sk_wmem_alloc);
 476
 477         return skb;
 478 }
 479
 480 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
 481                                      struct xdp_desc *desc)
 482 {
 483         struct net_device *dev = xs->dev;
 484         struct sk_buff *skb;
 485
 486         if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
 487                 skb = xsk_build_skb_zerocopy(xs, desc);
 488                 if (IS_ERR(skb))
 489                         return skb;
 490         } else {
 491                 u32 hr, tr, len;
 492                 void *buffer;
 493                 int err;
 494
 495                 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
 496                 tr = dev->needed_tailroom;
 497                 len = desc->len;
 498
 499                 skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
 500                 if (unlikely(!skb))
 501                         return ERR_PTR(err);
 502
 503                 skb_reserve(skb, hr);
 504                 skb_put(skb, len);
 505
 506                 buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
 507                 err = skb_store_bits(skb, 0, buffer, len);
 508                 if (unlikely(err)) {
 509                         kfree_skb(skb);
 510                         return ERR_PTR(err);
 511                 }
 512         }
 513
 514         skb->dev = dev;
 515         skb->priority = xs->sk.sk_priority;
 516         skb->mark = xs->sk.sk_mark;
 517         skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
 518         skb->destructor = xsk_destruct_skb;
 519
 520         return skb;
 521 }
 522
 523 static int xsk_generic_xmit(struct sock *sk)
 524 {
 525         struct xdp_sock *xs = xdp_sk(sk);
 526         u32 max_batch = TX_BATCH_SIZE;
 527         bool sent_frame = false;
 528         struct xdp_desc desc;
 529         struct sk_buff *skb;
 530         unsigned long flags;
 531         int err = 0;
 532
 533         mutex_lock(&xs->mutex);
 534
 535         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 536                 goto out;
 537
 538         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 539                 if (max_batch-- == 0) {
 540                         err = -EAGAIN;
 541                         goto out;
 542                 }
 543
 544                 skb = xsk_build_skb(xs, &desc);
 545                 if (IS_ERR(skb)) {
 546                         err = PTR_ERR(skb);
 547                         goto out;
 548                 }
 549
 550                 /* This is the backpressure mechanism for the Tx path.
 551                  * Reserve space in the completion queue and only proceed
 552                  * if there is space in it. This avoids having to implement
 553                  * any buffering in the Tx path.
 554                  */
 555                 spin_lock_irqsave(&xs->pool->cq_lock, flags);
 556                 if (xskq_prod_reserve(xs->pool->cq)) {
 557                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 558                         kfree_skb(skb);
 559                         goto out;
 560                 }
 561                 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 562
 563                 err = __dev_direct_xmit(skb, xs->queue_id);
 564                 if  (err == NETDEV_TX_BUSY) {
 565                         /* Tell user-space to retry the send */
 566                         skb->destructor = sock_wfree;
 567                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 568                         xskq_prod_cancel(xs->pool->cq);
 569                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 570                         /* Free skb without triggering the perf drop trace */
 571                         consume_skb(skb);
 572                         err = -EAGAIN;
 573                         goto out;
 574                 }
 575
 576                 xskq_cons_release(xs->tx);
 577                 /* Ignore NET_XMIT_CN as packet might have been sent */
 578                 if (err == NET_XMIT_DROP) {
 579                         /* SKB completed but not sent */
 580                         err = -EBUSY;
 581                         goto out;
 582                 }
 583
 584                 sent_frame = true;
 585         }
 586
 587         xs->tx->queue_empty_descs++;
 588
 589 out:
 590         if (sent_frame)
 591                 if (xsk_tx_writeable(xs))
 592                         sk->sk_write_space(sk);
 593
 594         mutex_unlock(&xs->mutex);
 595         return err;
 596 }
 597
 598 static int __xsk_sendmsg(struct sock *sk)
 599 {
 600         struct xdp_sock *xs = xdp_sk(sk);
 601
 602         if (unlikely(!(xs->dev->flags & IFF_UP)))
 603                 return -ENETDOWN;
 604         if (unlikely(!xs->tx))
 605                 return -ENOBUFS;
 606
 607         return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
 608 }
 609
 610 static bool xsk_no_wakeup(struct sock *sk)
 611 {
 612 #ifdef CONFIG_NET_RX_BUSY_POLL
 613         /* Prefer busy-polling, skip the wakeup. */
 614         return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
 615                 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
 616 #else
 617         return false;
 618 #endif
 619 }
 620
 621 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 622 {
 623         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 624         struct sock *sk = sock->sk;
 625         struct xdp_sock *xs = xdp_sk(sk);
 626         struct xsk_buff_pool *pool;
 627
 628         if (unlikely(!xsk_is_bound(xs)))
 629                 return -ENXIO;
 630         if (unlikely(need_wait))
 631                 return -EOPNOTSUPP;
 632
 633         if (sk_can_busy_loop(sk))
 634                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 635
 636         if (xsk_no_wakeup(sk))
 637                 return 0;
 638
 639         pool = xs->pool;
 640         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
 641                 return __xsk_sendmsg(sk);
 642         return 0;
 643 }
 644
 645 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 646 {
 647         bool need_wait = !(flags & MSG_DONTWAIT);
 648         struct sock *sk = sock->sk;
 649         struct xdp_sock *xs = xdp_sk(sk);
 650
 651         if (unlikely(!xsk_is_bound(xs)))
 652                 return -ENXIO;
 653         if (unlikely(!(xs->dev->flags & IFF_UP)))
 654                 return -ENETDOWN;
 655         if (unlikely(!xs->rx))
 656                 return -ENOBUFS;
 657         if (unlikely(need_wait))
 658                 return -EOPNOTSUPP;
 659
 660         if (sk_can_busy_loop(sk))
 661                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 662
 663         if (xsk_no_wakeup(sk))
 664                 return 0;
 665
 666         if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
 667                 return xsk_wakeup(xs, XDP_WAKEUP_RX);
 668         return 0;
 669 }
 670
 671 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 672                              struct poll_table_struct *wait)
 673 {
 674         __poll_t mask = 0;
 675         struct sock *sk = sock->sk;
 676         struct xdp_sock *xs = xdp_sk(sk);
 677         struct xsk_buff_pool *pool;
 678
 679         sock_poll_wait(file, sock, wait);
 680
 681         if (unlikely(!xsk_is_bound(xs)))
 682                 return mask;
 683
 684         pool = xs->pool;
 685
 686         if (pool->cached_need_wakeup) {
 687                 if (xs->zc)
 688                         xsk_wakeup(xs, pool->cached_need_wakeup);
 689                 else
 690                         /* Poll needs to drive Tx also in copy mode */
 691                         __xsk_sendmsg(sk);
 692         }
 693
 694         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 695                 mask |= EPOLLIN | EPOLLRDNORM;
 696         if (xs->tx && xsk_tx_writeable(xs))
 697                 mask |= EPOLLOUT | EPOLLWRNORM;
 698
 699         return mask;
 700 }
 701
 702 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 703                           bool umem_queue)
 704 {
 705         struct xsk_queue *q;
 706
 707         if (entries == 0 || *queue || !is_power_of_2(entries))
 708                 return -EINVAL;
 709
 710         q = xskq_create(entries, umem_queue);
 711         if (!q)
 712                 return -ENOMEM;
 713
 714         /* Make sure queue is ready before it can be seen by others */
 715         smp_wmb();
 716         WRITE_ONCE(*queue, q);
 717         return 0;
 718 }
 719
 720 static void xsk_unbind_dev(struct xdp_sock *xs)
 721 {
 722         struct net_device *dev = xs->dev;
 723
 724         if (xs->state != XSK_BOUND)
 725                 return;
 726         WRITE_ONCE(xs->state, XSK_UNBOUND);
 727
 728         /* Wait for driver to stop using the xdp socket. */
 729         xp_del_xsk(xs->pool, xs);
 730         xs->dev = NULL;
 731         synchronize_net();
 732         dev_put(dev);
 733 }
 734
 735 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 736                                               struct xdp_sock __rcu ***map_entry)
 737 {
 738         struct xsk_map *map = NULL;
 739         struct xsk_map_node *node;
 740
 741         *map_entry = NULL;
 742
 743         spin_lock_bh(&xs->map_list_lock);
 744         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 745                                         node);
 746         if (node) {
 747                 bpf_map_inc(&node->map->map);
 748                 map = node->map;
 749                 *map_entry = node->map_entry;
 750         }
 751         spin_unlock_bh(&xs->map_list_lock);
 752         return map;
 753 }
 754
 755 static void xsk_delete_from_maps(struct xdp_sock *xs)
 756 {
 757         /* This function removes the current XDP socket from all the
 758          * maps it resides in. We need to take extra care here, due to
 759          * the two locks involved. Each map has a lock synchronizing
 760          * updates to the entries, and each socket has a lock that
 761          * synchronizes access to the list of maps (map_list). For
 762          * deadlock avoidance the locks need to be taken in the order
 763          * "map lock"->"socket map list lock". We start off by
 764          * accessing the socket map list, and take a reference to the
 765          * map to guarantee existence between the
 766          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 767          * calls. Then we ask the map to remove the socket, which
 768          * tries to remove the socket from the map. Note that there
 769          * might be updates to the map between
 770          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 771          */
 772         struct xdp_sock __rcu **map_entry = NULL;
 773         struct xsk_map *map;
 774
 775         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 776                 xsk_map_try_sock_delete(map, xs, map_entry);
 777                 bpf_map_put(&map->map);
 778         }
 779 }
 780
 781 static int xsk_release(struct socket *sock)
 782 {
 783         struct sock *sk = sock->sk;
 784         struct xdp_sock *xs = xdp_sk(sk);
 785         struct net *net;
 786
 787         if (!sk)
 788                 return 0;
 789
 790         net = sock_net(sk);
 791
 792         mutex_lock(&net->xdp.lock);
 793         sk_del_node_init_rcu(sk);
 794         mutex_unlock(&net->xdp.lock);
 795
 796         sock_prot_inuse_add(net, sk->sk_prot, -1);
 797
 798         xsk_delete_from_maps(xs);
 799         mutex_lock(&xs->mutex);
 800         xsk_unbind_dev(xs);
 801         mutex_unlock(&xs->mutex);
 802
 803         xskq_destroy(xs->rx);
 804         xskq_destroy(xs->tx);
 805         xskq_destroy(xs->fq_tmp);
 806         xskq_destroy(xs->cq_tmp);
 807
 808         sock_orphan(sk);
 809         sock->sk = NULL;
 810
 811         sk_refcnt_debug_release(sk);
 812         sock_put(sk);
 813
 814         return 0;
 815 }
 816
 817 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 818 {
 819         struct socket *sock;
 820         int err;
 821
 822         sock = sockfd_lookup(fd, &err);
 823         if (!sock)
 824                 return ERR_PTR(-ENOTSOCK);
 825
 826         if (sock->sk->sk_family != PF_XDP) {
 827                 sockfd_put(sock);
 828                 return ERR_PTR(-ENOPROTOOPT);
 829         }
 830
 831         return sock;
 832 }
 833
 834 static bool xsk_validate_queues(struct xdp_sock *xs)
 835 {
 836         return xs->fq_tmp && xs->cq_tmp;
 837 }
 838
 839 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 840 {
 841         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 842         struct sock *sk = sock->sk;
 843         struct xdp_sock *xs = xdp_sk(sk);
 844         struct net_device *dev;
 845         u32 flags, qid;
 846         int err = 0;
 847
 848         if (addr_len < sizeof(struct sockaddr_xdp))
 849                 return -EINVAL;
 850         if (sxdp->sxdp_family != AF_XDP)
 851                 return -EINVAL;
 852
 853         flags = sxdp->sxdp_flags;
 854         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 855                       XDP_USE_NEED_WAKEUP))
 856                 return -EINVAL;
 857
 858         rtnl_lock();
 859         mutex_lock(&xs->mutex);
 860         if (xs->state != XSK_READY) {
 861                 err = -EBUSY;
 862                 goto out_release;
 863         }
 864
 865         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 866         if (!dev) {
 867                 err = -ENODEV;
 868                 goto out_release;
 869         }
 870
 871         if (!xs->rx && !xs->tx) {
 872                 err = -EINVAL;
 873                 goto out_unlock;
 874         }
 875
 876         qid = sxdp->sxdp_queue_id;
 877
 878         if (flags & XDP_SHARED_UMEM) {
 879                 struct xdp_sock *umem_xs;
 880                 struct socket *sock;
 881
 882                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 883                     (flags & XDP_USE_NEED_WAKEUP)) {
 884                         /* Cannot specify flags for shared sockets. */
 885                         err = -EINVAL;
 886                         goto out_unlock;
 887                 }
 888
 889                 if (xs->umem) {
 890                         /* We have already our own. */
 891                         err = -EINVAL;
 892                         goto out_unlock;
 893                 }
 894
 895                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 896                 if (IS_ERR(sock)) {
 897                         err = PTR_ERR(sock);
 898                         goto out_unlock;
 899                 }
 900
 901                 umem_xs = xdp_sk(sock->sk);
 902                 if (!xsk_is_bound(umem_xs)) {
 903                         err = -EBADF;
 904                         sockfd_put(sock);
 905                         goto out_unlock;
 906                 }
 907
 908                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 909                         /* Share the umem with another socket on another qid
 910                          * and/or device.
 911                          */
 912                         xs->pool = xp_create_and_assign_umem(xs,
 913                                                              umem_xs->umem);
 914                         if (!xs->pool) {
 915                                 err = -ENOMEM;
 916                                 sockfd_put(sock);
 917                                 goto out_unlock;
 918                         }
 919
 920                         err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
 921                                                    dev, qid);
 922                         if (err) {
 923                                 xp_destroy(xs->pool);
 924                                 xs->pool = NULL;
 925                                 sockfd_put(sock);
 926                                 goto out_unlock;
 927                         }
 928                 } else {
 929                         /* Share the buffer pool with the other socket. */
 930                         if (xs->fq_tmp || xs->cq_tmp) {
 931                                 /* Do not allow setting your own fq or cq. */
 932                                 err = -EINVAL;
 933                                 sockfd_put(sock);
 934                                 goto out_unlock;
 935                         }
 936
 937                         xp_get_pool(umem_xs->pool);
 938                         xs->pool = umem_xs->pool;
 939                 }
 940
 941                 xdp_get_umem(umem_xs->umem);
 942                 WRITE_ONCE(xs->umem, umem_xs->umem);
 943                 sockfd_put(sock);
 944         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 945                 err = -EINVAL;
 946                 goto out_unlock;
 947         } else {
 948                 /* This xsk has its own umem. */
 949                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 950                 if (!xs->pool) {
 951                         err = -ENOMEM;
 952                         goto out_unlock;
 953                 }
 954
 955                 err = xp_assign_dev(xs->pool, dev, qid, flags);
 956                 if (err) {
 957                         xp_destroy(xs->pool);
 958                         xs->pool = NULL;
 959                         goto out_unlock;
 960                 }
 961         }
 962
 963         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
 964         xs->fq_tmp = NULL;
 965         xs->cq_tmp = NULL;
 966
 967         xs->dev = dev;
 968         xs->zc = xs->umem->zc;
 969         xs->queue_id = qid;
 970         xp_add_xsk(xs->pool, xs);
 971
 972 out_unlock:
 973         if (err) {
 974                 dev_put(dev);
 975         } else {
 976                 /* Matches smp_rmb() in bind() for shared umem
 977                  * sockets, and xsk_is_bound().
 978                  */
 979                 smp_wmb();
 980                 WRITE_ONCE(xs->state, XSK_BOUND);
 981         }
 982 out_release:
 983         mutex_unlock(&xs->mutex);
 984         rtnl_unlock();
 985         return err;
 986 }
 987
 988 struct xdp_umem_reg_v1 {
 989         __u64 addr; /* Start of packet data area */
 990         __u64 len; /* Length of packet data area */
 991         __u32 chunk_size;
 992         __u32 headroom;
 993 };
 994
 995 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 996                           sockptr_t optval, unsigned int optlen)
 997 {
 998         struct sock *sk = sock->sk;
 999         struct xdp_sock *xs = xdp_sk(sk);
1000         int err;
1001
1002         if (level != SOL_XDP)
1003                 return -ENOPROTOOPT;
1004
1005         switch (optname) {
1006         case XDP_RX_RING:
1007         case XDP_TX_RING:
1008         {
1009                 struct xsk_queue **q;
1010                 int entries;
1011
1012                 if (optlen < sizeof(entries))
1013                         return -EINVAL;
1014                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1015                         return -EFAULT;
1016
1017                 mutex_lock(&xs->mutex);
1018                 if (xs->state != XSK_READY) {
1019                         mutex_unlock(&xs->mutex);
1020                         return -EBUSY;
1021                 }
1022                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1023                 err = xsk_init_queue(entries, q, false);
1024                 if (!err && optname == XDP_TX_RING)
1025                         /* Tx needs to be explicitly woken up the first time */
1026                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1027                 mutex_unlock(&xs->mutex);
1028                 return err;
1029         }
1030         case XDP_UMEM_REG:
1031         {
1032                 size_t mr_size = sizeof(struct xdp_umem_reg);
1033                 struct xdp_umem_reg mr = {};
1034                 struct xdp_umem *umem;
1035
1036                 if (optlen < sizeof(struct xdp_umem_reg_v1))
1037                         return -EINVAL;
1038                 else if (optlen < sizeof(mr))
1039                         mr_size = sizeof(struct xdp_umem_reg_v1);
1040
1041                 if (copy_from_sockptr(&mr, optval, mr_size))
1042                         return -EFAULT;
1043
1044                 mutex_lock(&xs->mutex);
1045                 if (xs->state != XSK_READY || xs->umem) {
1046                         mutex_unlock(&xs->mutex);
1047                         return -EBUSY;
1048                 }
1049
1050                 umem = xdp_umem_create(&mr);
1051                 if (IS_ERR(umem)) {
1052                         mutex_unlock(&xs->mutex);
1053                         return PTR_ERR(umem);
1054                 }
1055
1056                 /* Make sure umem is ready before it can be seen by others */
1057                 smp_wmb();
1058                 WRITE_ONCE(xs->umem, umem);
1059                 mutex_unlock(&xs->mutex);
1060                 return 0;
1061         }
1062         case XDP_UMEM_FILL_RING:
1063         case XDP_UMEM_COMPLETION_RING:
1064         {
1065                 struct xsk_queue **q;
1066                 int entries;
1067
1068                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1069                         return -EFAULT;
1070
1071                 mutex_lock(&xs->mutex);
1072                 if (xs->state != XSK_READY) {
1073                         mutex_unlock(&xs->mutex);
1074                         return -EBUSY;
1075                 }
1076
1077                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1078                         &xs->cq_tmp;
1079                 err = xsk_init_queue(entries, q, true);
1080                 mutex_unlock(&xs->mutex);
1081                 return err;
1082         }
1083         default:
1084                 break;
1085         }
1086
1087         return -ENOPROTOOPT;
1088 }
1089
1090 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1091 {
1092         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1093         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1094         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1095 }
1096
1097 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1098 {
1099         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1100         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1101         ring->desc = offsetof(struct xdp_umem_ring, desc);
1102 }
1103
1104 struct xdp_statistics_v1 {
1105         __u64 rx_dropped;
1106         __u64 rx_invalid_descs;
1107         __u64 tx_invalid_descs;
1108 };
1109
1110 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1111                           char __user *optval, int __user *optlen)
1112 {
1113         struct sock *sk = sock->sk;
1114         struct xdp_sock *xs = xdp_sk(sk);
1115         int len;
1116
1117         if (level != SOL_XDP)
1118                 return -ENOPROTOOPT;
1119
1120         if (get_user(len, optlen))
1121                 return -EFAULT;
1122         if (len < 0)
1123                 return -EINVAL;
1124
1125         switch (optname) {
1126         case XDP_STATISTICS:
1127         {
1128                 struct xdp_statistics stats = {};
1129                 bool extra_stats = true;
1130                 size_t stats_size;
1131
1132                 if (len < sizeof(struct xdp_statistics_v1)) {
1133                         return -EINVAL;
1134                 } else if (len < sizeof(stats)) {
1135                         extra_stats = false;
1136                         stats_size = sizeof(struct xdp_statistics_v1);
1137                 } else {
1138                         stats_size = sizeof(stats);
1139                 }
1140
1141                 mutex_lock(&xs->mutex);
1142                 stats.rx_dropped = xs->rx_dropped;
1143                 if (extra_stats) {
1144                         stats.rx_ring_full = xs->rx_queue_full;
1145                         stats.rx_fill_ring_empty_descs =
1146                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1147                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1148                 } else {
1149                         stats.rx_dropped += xs->rx_queue_full;
1150                 }
1151                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1152                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1153                 mutex_unlock(&xs->mutex);
1154
1155                 if (copy_to_user(optval, &stats, stats_size))
1156                         return -EFAULT;
1157                 if (put_user(stats_size, optlen))
1158                         return -EFAULT;
1159
1160                 return 0;
1161         }
1162         case XDP_MMAP_OFFSETS:
1163         {
1164                 struct xdp_mmap_offsets off;
1165                 struct xdp_mmap_offsets_v1 off_v1;
1166                 bool flags_supported = true;
1167                 void *to_copy;
1168
1169                 if (len < sizeof(off_v1))
1170                         return -EINVAL;
1171                 else if (len < sizeof(off))
1172                         flags_supported = false;
1173
1174                 if (flags_supported) {
1175                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1176                          * except for the flags field added to the end.
1177                          */
1178                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1179                                                &off.rx);
1180                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1181                                                &off.tx);
1182                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1183                                                &off.fr);
1184                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1185                                                &off.cr);
1186                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1187                                                 ptrs.flags);
1188                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1189                                                 ptrs.flags);
1190                         off.fr.flags = offsetof(struct xdp_umem_ring,
1191                                                 ptrs.flags);
1192                         off.cr.flags = offsetof(struct xdp_umem_ring,
1193                                                 ptrs.flags);
1194
1195                         len = sizeof(off);
1196                         to_copy = &off;
1197                 } else {
1198                         xsk_enter_rxtx_offsets(&off_v1.rx);
1199                         xsk_enter_rxtx_offsets(&off_v1.tx);
1200                         xsk_enter_umem_offsets(&off_v1.fr);
1201                         xsk_enter_umem_offsets(&off_v1.cr);
1202
1203                         len = sizeof(off_v1);
1204                         to_copy = &off_v1;
1205                 }
1206
1207                 if (copy_to_user(optval, to_copy, len))
1208                         return -EFAULT;
1209                 if (put_user(len, optlen))
1210                         return -EFAULT;
1211
1212                 return 0;
1213         }
1214         case XDP_OPTIONS:
1215         {
1216                 struct xdp_options opts = {};
1217
1218                 if (len < sizeof(opts))
1219                         return -EINVAL;
1220
1221                 mutex_lock(&xs->mutex);
1222                 if (xs->zc)
1223                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1224                 mutex_unlock(&xs->mutex);
1225
1226                 len = sizeof(opts);
1227                 if (copy_to_user(optval, &opts, len))
1228                         return -EFAULT;
1229                 if (put_user(len, optlen))
1230                         return -EFAULT;
1231
1232                 return 0;
1233         }
1234         default:
1235                 break;
1236         }
1237
1238         return -EOPNOTSUPP;
1239 }
1240
1241 static int xsk_mmap(struct file *file, struct socket *sock,
1242                     struct vm_area_struct *vma)
1243 {
1244         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1245         unsigned long size = vma->vm_end - vma->vm_start;
1246         struct xdp_sock *xs = xdp_sk(sock->sk);
1247         struct xsk_queue *q = NULL;
1248         unsigned long pfn;
1249         struct page *qpg;
1250
1251         if (READ_ONCE(xs->state) != XSK_READY)
1252                 return -EBUSY;
1253
1254         if (offset == XDP_PGOFF_RX_RING) {
1255                 q = READ_ONCE(xs->rx);
1256         } else if (offset == XDP_PGOFF_TX_RING) {
1257                 q = READ_ONCE(xs->tx);
1258         } else {
1259                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1260                 smp_rmb();
1261                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1262                         q = READ_ONCE(xs->fq_tmp);
1263                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1264                         q = READ_ONCE(xs->cq_tmp);
1265         }
1266
1267         if (!q)
1268                 return -EINVAL;
1269
1270         /* Matches the smp_wmb() in xsk_init_queue */
1271         smp_rmb();
1272         qpg = virt_to_head_page(q->ring);
1273         if (size > page_size(qpg))
1274                 return -EINVAL;
1275
1276         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1277         return remap_pfn_range(vma, vma->vm_start, pfn,
1278                                size, vma->vm_page_prot);
1279 }
1280
1281 static int xsk_notifier(struct notifier_block *this,
1282                         unsigned long msg, void *ptr)
1283 {
1284         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1285         struct net *net = dev_net(dev);
1286         struct sock *sk;
1287
1288         switch (msg) {
1289         case NETDEV_UNREGISTER:
1290                 mutex_lock(&net->xdp.lock);
1291                 sk_for_each(sk, &net->xdp.list) {
1292                         struct xdp_sock *xs = xdp_sk(sk);
1293
1294                         mutex_lock(&xs->mutex);
1295                         if (xs->dev == dev) {
1296                                 sk->sk_err = ENETDOWN;
1297                                 if (!sock_flag(sk, SOCK_DEAD))
1298                                         sk_error_report(sk);
1299
1300                                 xsk_unbind_dev(xs);
1301
1302                                 /* Clear device references. */
1303                                 xp_clear_dev(xs->pool);
1304                         }
1305                         mutex_unlock(&xs->mutex);
1306                 }
1307                 mutex_unlock(&net->xdp.lock);
1308                 break;
1309         }
1310         return NOTIFY_DONE;
1311 }
1312
1313 static struct proto xsk_proto = {
1314         .name =         "XDP",
1315         .owner =        THIS_MODULE,
1316         .obj_size =     sizeof(struct xdp_sock),
1317 };
1318
1319 static const struct proto_ops xsk_proto_ops = {
1320         .family         = PF_XDP,
1321         .owner          = THIS_MODULE,
1322         .release        = xsk_release,
1323         .bind           = xsk_bind,
1324         .connect        = sock_no_connect,
1325         .socketpair     = sock_no_socketpair,
1326         .accept         = sock_no_accept,
1327         .getname        = sock_no_getname,
1328         .poll           = xsk_poll,
1329         .ioctl          = sock_no_ioctl,
1330         .listen         = sock_no_listen,
1331         .shutdown       = sock_no_shutdown,
1332         .setsockopt     = xsk_setsockopt,
1333         .getsockopt     = xsk_getsockopt,
1334         .sendmsg        = xsk_sendmsg,
1335         .recvmsg        = xsk_recvmsg,
1336         .mmap           = xsk_mmap,
1337         .sendpage       = sock_no_sendpage,
1338 };
1339
1340 static void xsk_destruct(struct sock *sk)
1341 {
1342         struct xdp_sock *xs = xdp_sk(sk);
1343
1344         if (!sock_flag(sk, SOCK_DEAD))
1345                 return;
1346
1347         if (!xp_put_pool(xs->pool))
1348                 xdp_put_umem(xs->umem, !xs->pool);
1349
1350         sk_refcnt_debug_dec(sk);
1351 }
1352
1353 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1354                       int kern)
1355 {
1356         struct xdp_sock *xs;
1357         struct sock *sk;
1358
1359         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1360                 return -EPERM;
1361         if (sock->type != SOCK_RAW)
1362                 return -ESOCKTNOSUPPORT;
1363
1364         if (protocol)
1365                 return -EPROTONOSUPPORT;
1366
1367         sock->state = SS_UNCONNECTED;
1368
1369         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1370         if (!sk)
1371                 return -ENOBUFS;
1372
1373         sock->ops = &xsk_proto_ops;
1374
1375         sock_init_data(sock, sk);
1376
1377         sk->sk_family = PF_XDP;
1378
1379         sk->sk_destruct = xsk_destruct;
1380         sk_refcnt_debug_inc(sk);
1381
1382         sock_set_flag(sk, SOCK_RCU_FREE);
1383
1384         xs = xdp_sk(sk);
1385         xs->state = XSK_READY;
1386         mutex_init(&xs->mutex);
1387         spin_lock_init(&xs->rx_lock);
1388
1389         INIT_LIST_HEAD(&xs->map_list);
1390         spin_lock_init(&xs->map_list_lock);
1391
1392         mutex_lock(&net->xdp.lock);
1393         sk_add_node_rcu(sk, &net->xdp.list);
1394         mutex_unlock(&net->xdp.lock);
1395
1396         sock_prot_inuse_add(net, &xsk_proto, 1);
1397
1398         return 0;
1399 }
1400
1401 static const struct net_proto_family xsk_family_ops = {
1402         .family = PF_XDP,
1403         .create = xsk_create,
1404         .owner  = THIS_MODULE,
1405 };
1406
1407 static struct notifier_block xsk_netdev_notifier = {
1408         .notifier_call  = xsk_notifier,
1409 };
1410
1411 static int __net_init xsk_net_init(struct net *net)
1412 {
1413         mutex_init(&net->xdp.lock);
1414         INIT_HLIST_HEAD(&net->xdp.list);
1415         return 0;
1416 }
1417
1418 static void __net_exit xsk_net_exit(struct net *net)
1419 {
1420         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1421 }
1422
1423 static struct pernet_operations xsk_net_ops = {
1424         .init = xsk_net_init,
1425         .exit = xsk_net_exit,
1426 };
1427
1428 static int __init xsk_init(void)
1429 {
1430         int err, cpu;
1431
1432         err = proto_register(&xsk_proto, 0 /* no slab */);
1433         if (err)
1434                 goto out;
1435
1436         err = sock_register(&xsk_family_ops);
1437         if (err)
1438                 goto out_proto;
1439
1440         err = register_pernet_subsys(&xsk_net_ops);
1441         if (err)
1442                 goto out_sk;
1443
1444         err = register_netdevice_notifier(&xsk_netdev_notifier);
1445         if (err)
1446                 goto out_pernet;
1447
1448         for_each_possible_cpu(cpu)
1449                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1450         return 0;
1451
1452 out_pernet:
1453         unregister_pernet_subsys(&xsk_net_ops);
1454 out_sk:
1455         sock_unregister(PF_XDP);
1456 out_proto:
1457         proto_unregister(&xsk_proto);
1458 out:
1459         return err;
1460 }
1461
1462 fs_initcall(xsk_init);