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, struct xdp_desc *descs,
 347                                         u32 max_entries)
 348 {
 349         u32 nb_pkts = 0;
 350
 351         while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts]))
 352                 nb_pkts++;
 353
 354         xsk_tx_release(pool);
 355         return nb_pkts;
 356 }
 357
 358 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, struct xdp_desc *descs,
 359                                    u32 max_entries)
 360 {
 361         struct xdp_sock *xs;
 362         u32 nb_pkts;
 363
 364         rcu_read_lock();
 365         if (!list_is_singular(&pool->xsk_tx_list)) {
 366                 /* Fallback to the non-batched version */
 367                 rcu_read_unlock();
 368                 return xsk_tx_peek_release_fallback(pool, descs, max_entries);
 369         }
 370
 371         xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
 372         if (!xs) {
 373                 nb_pkts = 0;
 374                 goto out;
 375         }
 376
 377         nb_pkts = xskq_cons_peek_desc_batch(xs->tx, descs, pool, max_entries);
 378         if (!nb_pkts) {
 379                 xs->tx->queue_empty_descs++;
 380                 goto out;
 381         }
 382
 383         /* This is the backpressure mechanism for the Tx path. Try to
 384          * reserve space in the completion queue for all packets, but
 385          * if there are fewer slots available, just process that many
 386          * packets. This avoids having to implement any buffering in
 387          * the Tx path.
 388          */
 389         nb_pkts = xskq_prod_reserve_addr_batch(pool->cq, descs, nb_pkts);
 390         if (!nb_pkts)
 391                 goto out;
 392
 393         xskq_cons_release_n(xs->tx, nb_pkts);
 394         __xskq_cons_release(xs->tx);
 395         xs->sk.sk_write_space(&xs->sk);
 396
 397 out:
 398         rcu_read_unlock();
 399         return nb_pkts;
 400 }
 401 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
 402
 403 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
 404 {
 405         struct net_device *dev = xs->dev;
 406         int err;
 407
 408         rcu_read_lock();
 409         err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 410         rcu_read_unlock();
 411
 412         return err;
 413 }
 414
 415 static int xsk_zc_xmit(struct xdp_sock *xs)
 416 {
 417         return xsk_wakeup(xs, XDP_WAKEUP_TX);
 418 }
 419
 420 static void xsk_destruct_skb(struct sk_buff *skb)
 421 {
 422         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 423         struct xdp_sock *xs = xdp_sk(skb->sk);
 424         unsigned long flags;
 425
 426         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 427         xskq_prod_submit_addr(xs->pool->cq, addr);
 428         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 429
 430         sock_wfree(skb);
 431 }
 432
 433 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
 434                                               struct xdp_desc *desc)
 435 {
 436         struct xsk_buff_pool *pool = xs->pool;
 437         u32 hr, len, ts, offset, copy, copied;
 438         struct sk_buff *skb;
 439         struct page *page;
 440         void *buffer;
 441         int err, i;
 442         u64 addr;
 443
 444         hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
 445
 446         skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
 447         if (unlikely(!skb))
 448                 return ERR_PTR(err);
 449
 450         skb_reserve(skb, hr);
 451
 452         addr = desc->addr;
 453         len = desc->len;
 454         ts = pool->unaligned ? len : pool->chunk_size;
 455
 456         buffer = xsk_buff_raw_get_data(pool, addr);
 457         offset = offset_in_page(buffer);
 458         addr = buffer - pool->addrs;
 459
 460         for (copied = 0, i = 0; copied < len; i++) {
 461                 page = pool->umem->pgs[addr >> PAGE_SHIFT];
 462                 get_page(page);
 463
 464                 copy = min_t(u32, PAGE_SIZE - offset, len - copied);
 465                 skb_fill_page_desc(skb, i, page, offset, copy);
 466
 467                 copied += copy;
 468                 addr += copy;
 469                 offset = 0;
 470         }
 471
 472         skb->len += len;
 473         skb->data_len += len;
 474         skb->truesize += ts;
 475
 476         refcount_add(ts, &xs->sk.sk_wmem_alloc);
 477
 478         return skb;
 479 }
 480
 481 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
 482                                      struct xdp_desc *desc)
 483 {
 484         struct net_device *dev = xs->dev;
 485         struct sk_buff *skb;
 486
 487         if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
 488                 skb = xsk_build_skb_zerocopy(xs, desc);
 489                 if (IS_ERR(skb))
 490                         return skb;
 491         } else {
 492                 u32 hr, tr, len;
 493                 void *buffer;
 494                 int err;
 495
 496                 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
 497                 tr = dev->needed_tailroom;
 498                 len = desc->len;
 499
 500                 skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
 501                 if (unlikely(!skb))
 502                         return ERR_PTR(err);
 503
 504                 skb_reserve(skb, hr);
 505                 skb_put(skb, len);
 506
 507                 buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
 508                 err = skb_store_bits(skb, 0, buffer, len);
 509                 if (unlikely(err)) {
 510                         kfree_skb(skb);
 511                         return ERR_PTR(err);
 512                 }
 513         }
 514
 515         skb->dev = dev;
 516         skb->priority = xs->sk.sk_priority;
 517         skb->mark = xs->sk.sk_mark;
 518         skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
 519         skb->destructor = xsk_destruct_skb;
 520
 521         return skb;
 522 }
 523
 524 static int xsk_generic_xmit(struct sock *sk)
 525 {
 526         struct xdp_sock *xs = xdp_sk(sk);
 527         u32 max_batch = TX_BATCH_SIZE;
 528         bool sent_frame = false;
 529         struct xdp_desc desc;
 530         struct sk_buff *skb;
 531         unsigned long flags;
 532         int err = 0;
 533
 534         mutex_lock(&xs->mutex);
 535
 536         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 537                 goto out;
 538
 539         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 540                 if (max_batch-- == 0) {
 541                         err = -EAGAIN;
 542                         goto out;
 543                 }
 544
 545                 skb = xsk_build_skb(xs, &desc);
 546                 if (IS_ERR(skb)) {
 547                         err = PTR_ERR(skb);
 548                         goto out;
 549                 }
 550
 551                 /* This is the backpressure mechanism for the Tx path.
 552                  * Reserve space in the completion queue and only proceed
 553                  * if there is space in it. This avoids having to implement
 554                  * any buffering in the Tx path.
 555                  */
 556                 spin_lock_irqsave(&xs->pool->cq_lock, flags);
 557                 if (xskq_prod_reserve(xs->pool->cq)) {
 558                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 559                         kfree_skb(skb);
 560                         goto out;
 561                 }
 562                 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 563
 564                 err = __dev_direct_xmit(skb, xs->queue_id);
 565                 if  (err == NETDEV_TX_BUSY) {
 566                         /* Tell user-space to retry the send */
 567                         skb->destructor = sock_wfree;
 568                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 569                         xskq_prod_cancel(xs->pool->cq);
 570                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 571                         /* Free skb without triggering the perf drop trace */
 572                         consume_skb(skb);
 573                         err = -EAGAIN;
 574                         goto out;
 575                 }
 576
 577                 xskq_cons_release(xs->tx);
 578                 /* Ignore NET_XMIT_CN as packet might have been sent */
 579                 if (err == NET_XMIT_DROP) {
 580                         /* SKB completed but not sent */
 581                         err = -EBUSY;
 582                         goto out;
 583                 }
 584
 585                 sent_frame = true;
 586         }
 587
 588         xs->tx->queue_empty_descs++;
 589
 590 out:
 591         if (sent_frame)
 592                 if (xsk_tx_writeable(xs))
 593                         sk->sk_write_space(sk);
 594
 595         mutex_unlock(&xs->mutex);
 596         return err;
 597 }
 598
 599 static int __xsk_sendmsg(struct sock *sk)
 600 {
 601         struct xdp_sock *xs = xdp_sk(sk);
 602
 603         if (unlikely(!(xs->dev->flags & IFF_UP)))
 604                 return -ENETDOWN;
 605         if (unlikely(!xs->tx))
 606                 return -ENOBUFS;
 607
 608         return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
 609 }
 610
 611 static bool xsk_no_wakeup(struct sock *sk)
 612 {
 613 #ifdef CONFIG_NET_RX_BUSY_POLL
 614         /* Prefer busy-polling, skip the wakeup. */
 615         return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
 616                 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
 617 #else
 618         return false;
 619 #endif
 620 }
 621
 622 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 623 {
 624         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 625         struct sock *sk = sock->sk;
 626         struct xdp_sock *xs = xdp_sk(sk);
 627         struct xsk_buff_pool *pool;
 628
 629         if (unlikely(!xsk_is_bound(xs)))
 630                 return -ENXIO;
 631         if (unlikely(need_wait))
 632                 return -EOPNOTSUPP;
 633
 634         if (sk_can_busy_loop(sk))
 635                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 636
 637         if (xsk_no_wakeup(sk))
 638                 return 0;
 639
 640         pool = xs->pool;
 641         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
 642                 return __xsk_sendmsg(sk);
 643         return 0;
 644 }
 645
 646 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 647 {
 648         bool need_wait = !(flags & MSG_DONTWAIT);
 649         struct sock *sk = sock->sk;
 650         struct xdp_sock *xs = xdp_sk(sk);
 651
 652         if (unlikely(!xsk_is_bound(xs)))
 653                 return -ENXIO;
 654         if (unlikely(!(xs->dev->flags & IFF_UP)))
 655                 return -ENETDOWN;
 656         if (unlikely(!xs->rx))
 657                 return -ENOBUFS;
 658         if (unlikely(need_wait))
 659                 return -EOPNOTSUPP;
 660
 661         if (sk_can_busy_loop(sk))
 662                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 663
 664         if (xsk_no_wakeup(sk))
 665                 return 0;
 666
 667         if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
 668                 return xsk_wakeup(xs, XDP_WAKEUP_RX);
 669         return 0;
 670 }
 671
 672 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 673                              struct poll_table_struct *wait)
 674 {
 675         __poll_t mask = 0;
 676         struct sock *sk = sock->sk;
 677         struct xdp_sock *xs = xdp_sk(sk);
 678         struct xsk_buff_pool *pool;
 679
 680         sock_poll_wait(file, sock, wait);
 681
 682         if (unlikely(!xsk_is_bound(xs)))
 683                 return mask;
 684
 685         pool = xs->pool;
 686
 687         if (pool->cached_need_wakeup) {
 688                 if (xs->zc)
 689                         xsk_wakeup(xs, pool->cached_need_wakeup);
 690                 else
 691                         /* Poll needs to drive Tx also in copy mode */
 692                         __xsk_sendmsg(sk);
 693         }
 694
 695         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 696                 mask |= EPOLLIN | EPOLLRDNORM;
 697         if (xs->tx && xsk_tx_writeable(xs))
 698                 mask |= EPOLLOUT | EPOLLWRNORM;
 699
 700         return mask;
 701 }
 702
 703 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 704                           bool umem_queue)
 705 {
 706         struct xsk_queue *q;
 707
 708         if (entries == 0 || *queue || !is_power_of_2(entries))
 709                 return -EINVAL;
 710
 711         q = xskq_create(entries, umem_queue);
 712         if (!q)
 713                 return -ENOMEM;
 714
 715         /* Make sure queue is ready before it can be seen by others */
 716         smp_wmb();
 717         WRITE_ONCE(*queue, q);
 718         return 0;
 719 }
 720
 721 static void xsk_unbind_dev(struct xdp_sock *xs)
 722 {
 723         struct net_device *dev = xs->dev;
 724
 725         if (xs->state != XSK_BOUND)
 726                 return;
 727         WRITE_ONCE(xs->state, XSK_UNBOUND);
 728
 729         /* Wait for driver to stop using the xdp socket. */
 730         xp_del_xsk(xs->pool, xs);
 731         xs->dev = NULL;
 732         synchronize_net();
 733         dev_put(dev);
 734 }
 735
 736 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 737                                               struct xdp_sock __rcu ***map_entry)
 738 {
 739         struct xsk_map *map = NULL;
 740         struct xsk_map_node *node;
 741
 742         *map_entry = NULL;
 743
 744         spin_lock_bh(&xs->map_list_lock);
 745         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 746                                         node);
 747         if (node) {
 748                 bpf_map_inc(&node->map->map);
 749                 map = node->map;
 750                 *map_entry = node->map_entry;
 751         }
 752         spin_unlock_bh(&xs->map_list_lock);
 753         return map;
 754 }
 755
 756 static void xsk_delete_from_maps(struct xdp_sock *xs)
 757 {
 758         /* This function removes the current XDP socket from all the
 759          * maps it resides in. We need to take extra care here, due to
 760          * the two locks involved. Each map has a lock synchronizing
 761          * updates to the entries, and each socket has a lock that
 762          * synchronizes access to the list of maps (map_list). For
 763          * deadlock avoidance the locks need to be taken in the order
 764          * "map lock"->"socket map list lock". We start off by
 765          * accessing the socket map list, and take a reference to the
 766          * map to guarantee existence between the
 767          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 768          * calls. Then we ask the map to remove the socket, which
 769          * tries to remove the socket from the map. Note that there
 770          * might be updates to the map between
 771          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 772          */
 773         struct xdp_sock __rcu **map_entry = NULL;
 774         struct xsk_map *map;
 775
 776         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 777                 xsk_map_try_sock_delete(map, xs, map_entry);
 778                 bpf_map_put(&map->map);
 779         }
 780 }
 781
 782 static int xsk_release(struct socket *sock)
 783 {
 784         struct sock *sk = sock->sk;
 785         struct xdp_sock *xs = xdp_sk(sk);
 786         struct net *net;
 787
 788         if (!sk)
 789                 return 0;
 790
 791         net = sock_net(sk);
 792
 793         mutex_lock(&net->xdp.lock);
 794         sk_del_node_init_rcu(sk);
 795         mutex_unlock(&net->xdp.lock);
 796
 797         sock_prot_inuse_add(net, sk->sk_prot, -1);
 798
 799         xsk_delete_from_maps(xs);
 800         mutex_lock(&xs->mutex);
 801         xsk_unbind_dev(xs);
 802         mutex_unlock(&xs->mutex);
 803
 804         xskq_destroy(xs->rx);
 805         xskq_destroy(xs->tx);
 806         xskq_destroy(xs->fq_tmp);
 807         xskq_destroy(xs->cq_tmp);
 808
 809         sock_orphan(sk);
 810         sock->sk = NULL;
 811
 812         sk_refcnt_debug_release(sk);
 813         sock_put(sk);
 814
 815         return 0;
 816 }
 817
 818 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 819 {
 820         struct socket *sock;
 821         int err;
 822
 823         sock = sockfd_lookup(fd, &err);
 824         if (!sock)
 825                 return ERR_PTR(-ENOTSOCK);
 826
 827         if (sock->sk->sk_family != PF_XDP) {
 828                 sockfd_put(sock);
 829                 return ERR_PTR(-ENOPROTOOPT);
 830         }
 831
 832         return sock;
 833 }
 834
 835 static bool xsk_validate_queues(struct xdp_sock *xs)
 836 {
 837         return xs->fq_tmp && xs->cq_tmp;
 838 }
 839
 840 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 841 {
 842         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 843         struct sock *sk = sock->sk;
 844         struct xdp_sock *xs = xdp_sk(sk);
 845         struct net_device *dev;
 846         u32 flags, qid;
 847         int err = 0;
 848
 849         if (addr_len < sizeof(struct sockaddr_xdp))
 850                 return -EINVAL;
 851         if (sxdp->sxdp_family != AF_XDP)
 852                 return -EINVAL;
 853
 854         flags = sxdp->sxdp_flags;
 855         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 856                       XDP_USE_NEED_WAKEUP))
 857                 return -EINVAL;
 858
 859         rtnl_lock();
 860         mutex_lock(&xs->mutex);
 861         if (xs->state != XSK_READY) {
 862                 err = -EBUSY;
 863                 goto out_release;
 864         }
 865
 866         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 867         if (!dev) {
 868                 err = -ENODEV;
 869                 goto out_release;
 870         }
 871
 872         if (!xs->rx && !xs->tx) {
 873                 err = -EINVAL;
 874                 goto out_unlock;
 875         }
 876
 877         qid = sxdp->sxdp_queue_id;
 878
 879         if (flags & XDP_SHARED_UMEM) {
 880                 struct xdp_sock *umem_xs;
 881                 struct socket *sock;
 882
 883                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 884                     (flags & XDP_USE_NEED_WAKEUP)) {
 885                         /* Cannot specify flags for shared sockets. */
 886                         err = -EINVAL;
 887                         goto out_unlock;
 888                 }
 889
 890                 if (xs->umem) {
 891                         /* We have already our own. */
 892                         err = -EINVAL;
 893                         goto out_unlock;
 894                 }
 895
 896                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 897                 if (IS_ERR(sock)) {
 898                         err = PTR_ERR(sock);
 899                         goto out_unlock;
 900                 }
 901
 902                 umem_xs = xdp_sk(sock->sk);
 903                 if (!xsk_is_bound(umem_xs)) {
 904                         err = -EBADF;
 905                         sockfd_put(sock);
 906                         goto out_unlock;
 907                 }
 908
 909                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 910                         /* Share the umem with another socket on another qid
 911                          * and/or device.
 912                          */
 913                         xs->pool = xp_create_and_assign_umem(xs,
 914                                                              umem_xs->umem);
 915                         if (!xs->pool) {
 916                                 err = -ENOMEM;
 917                                 sockfd_put(sock);
 918                                 goto out_unlock;
 919                         }
 920
 921                         err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
 922                                                    dev, qid);
 923                         if (err) {
 924                                 xp_destroy(xs->pool);
 925                                 xs->pool = NULL;
 926                                 sockfd_put(sock);
 927                                 goto out_unlock;
 928                         }
 929                 } else {
 930                         /* Share the buffer pool with the other socket. */
 931                         if (xs->fq_tmp || xs->cq_tmp) {
 932                                 /* Do not allow setting your own fq or cq. */
 933                                 err = -EINVAL;
 934                                 sockfd_put(sock);
 935                                 goto out_unlock;
 936                         }
 937
 938                         xp_get_pool(umem_xs->pool);
 939                         xs->pool = umem_xs->pool;
 940                 }
 941
 942                 xdp_get_umem(umem_xs->umem);
 943                 WRITE_ONCE(xs->umem, umem_xs->umem);
 944                 sockfd_put(sock);
 945         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 946                 err = -EINVAL;
 947                 goto out_unlock;
 948         } else {
 949                 /* This xsk has its own umem. */
 950                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 951                 if (!xs->pool) {
 952                         err = -ENOMEM;
 953                         goto out_unlock;
 954                 }
 955
 956                 err = xp_assign_dev(xs->pool, dev, qid, flags);
 957                 if (err) {
 958                         xp_destroy(xs->pool);
 959                         xs->pool = NULL;
 960                         goto out_unlock;
 961                 }
 962         }
 963
 964         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
 965         xs->fq_tmp = NULL;
 966         xs->cq_tmp = NULL;
 967
 968         xs->dev = dev;
 969         xs->zc = xs->umem->zc;
 970         xs->queue_id = qid;
 971         xp_add_xsk(xs->pool, xs);
 972
 973 out_unlock:
 974         if (err) {
 975                 dev_put(dev);
 976         } else {
 977                 /* Matches smp_rmb() in bind() for shared umem
 978                  * sockets, and xsk_is_bound().
 979                  */
 980                 smp_wmb();
 981                 WRITE_ONCE(xs->state, XSK_BOUND);
 982         }
 983 out_release:
 984         mutex_unlock(&xs->mutex);
 985         rtnl_unlock();
 986         return err;
 987 }
 988
 989 struct xdp_umem_reg_v1 {
 990         __u64 addr; /* Start of packet data area */
 991         __u64 len; /* Length of packet data area */
 992         __u32 chunk_size;
 993         __u32 headroom;
 994 };
 995
 996 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 997                           sockptr_t optval, unsigned int optlen)
 998 {
 999         struct sock *sk = sock->sk;
1000         struct xdp_sock *xs = xdp_sk(sk);
1001         int err;
1002
1003         if (level != SOL_XDP)
1004                 return -ENOPROTOOPT;
1005
1006         switch (optname) {
1007         case XDP_RX_RING:
1008         case XDP_TX_RING:
1009         {
1010                 struct xsk_queue **q;
1011                 int entries;
1012
1013                 if (optlen < sizeof(entries))
1014                         return -EINVAL;
1015                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1016                         return -EFAULT;
1017
1018                 mutex_lock(&xs->mutex);
1019                 if (xs->state != XSK_READY) {
1020                         mutex_unlock(&xs->mutex);
1021                         return -EBUSY;
1022                 }
1023                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1024                 err = xsk_init_queue(entries, q, false);
1025                 if (!err && optname == XDP_TX_RING)
1026                         /* Tx needs to be explicitly woken up the first time */
1027                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1028                 mutex_unlock(&xs->mutex);
1029                 return err;
1030         }
1031         case XDP_UMEM_REG:
1032         {
1033                 size_t mr_size = sizeof(struct xdp_umem_reg);
1034                 struct xdp_umem_reg mr = {};
1035                 struct xdp_umem *umem;
1036
1037                 if (optlen < sizeof(struct xdp_umem_reg_v1))
1038                         return -EINVAL;
1039                 else if (optlen < sizeof(mr))
1040                         mr_size = sizeof(struct xdp_umem_reg_v1);
1041
1042                 if (copy_from_sockptr(&mr, optval, mr_size))
1043                         return -EFAULT;
1044
1045                 mutex_lock(&xs->mutex);
1046                 if (xs->state != XSK_READY || xs->umem) {
1047                         mutex_unlock(&xs->mutex);
1048                         return -EBUSY;
1049                 }
1050
1051                 umem = xdp_umem_create(&mr);
1052                 if (IS_ERR(umem)) {
1053                         mutex_unlock(&xs->mutex);
1054                         return PTR_ERR(umem);
1055                 }
1056
1057                 /* Make sure umem is ready before it can be seen by others */
1058                 smp_wmb();
1059                 WRITE_ONCE(xs->umem, umem);
1060                 mutex_unlock(&xs->mutex);
1061                 return 0;
1062         }
1063         case XDP_UMEM_FILL_RING:
1064         case XDP_UMEM_COMPLETION_RING:
1065         {
1066                 struct xsk_queue **q;
1067                 int entries;
1068
1069                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1070                         return -EFAULT;
1071
1072                 mutex_lock(&xs->mutex);
1073                 if (xs->state != XSK_READY) {
1074                         mutex_unlock(&xs->mutex);
1075                         return -EBUSY;
1076                 }
1077
1078                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1079                         &xs->cq_tmp;
1080                 err = xsk_init_queue(entries, q, true);
1081                 mutex_unlock(&xs->mutex);
1082                 return err;
1083         }
1084         default:
1085                 break;
1086         }
1087
1088         return -ENOPROTOOPT;
1089 }
1090
1091 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1092 {
1093         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1094         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1095         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1096 }
1097
1098 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1099 {
1100         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1101         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1102         ring->desc = offsetof(struct xdp_umem_ring, desc);
1103 }
1104
1105 struct xdp_statistics_v1 {
1106         __u64 rx_dropped;
1107         __u64 rx_invalid_descs;
1108         __u64 tx_invalid_descs;
1109 };
1110
1111 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1112                           char __user *optval, int __user *optlen)
1113 {
1114         struct sock *sk = sock->sk;
1115         struct xdp_sock *xs = xdp_sk(sk);
1116         int len;
1117
1118         if (level != SOL_XDP)
1119                 return -ENOPROTOOPT;
1120
1121         if (get_user(len, optlen))
1122                 return -EFAULT;
1123         if (len < 0)
1124                 return -EINVAL;
1125
1126         switch (optname) {
1127         case XDP_STATISTICS:
1128         {
1129                 struct xdp_statistics stats = {};
1130                 bool extra_stats = true;
1131                 size_t stats_size;
1132
1133                 if (len < sizeof(struct xdp_statistics_v1)) {
1134                         return -EINVAL;
1135                 } else if (len < sizeof(stats)) {
1136                         extra_stats = false;
1137                         stats_size = sizeof(struct xdp_statistics_v1);
1138                 } else {
1139                         stats_size = sizeof(stats);
1140                 }
1141
1142                 mutex_lock(&xs->mutex);
1143                 stats.rx_dropped = xs->rx_dropped;
1144                 if (extra_stats) {
1145                         stats.rx_ring_full = xs->rx_queue_full;
1146                         stats.rx_fill_ring_empty_descs =
1147                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1148                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1149                 } else {
1150                         stats.rx_dropped += xs->rx_queue_full;
1151                 }
1152                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1153                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1154                 mutex_unlock(&xs->mutex);
1155
1156                 if (copy_to_user(optval, &stats, stats_size))
1157                         return -EFAULT;
1158                 if (put_user(stats_size, optlen))
1159                         return -EFAULT;
1160
1161                 return 0;
1162         }
1163         case XDP_MMAP_OFFSETS:
1164         {
1165                 struct xdp_mmap_offsets off;
1166                 struct xdp_mmap_offsets_v1 off_v1;
1167                 bool flags_supported = true;
1168                 void *to_copy;
1169
1170                 if (len < sizeof(off_v1))
1171                         return -EINVAL;
1172                 else if (len < sizeof(off))
1173                         flags_supported = false;
1174
1175                 if (flags_supported) {
1176                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1177                          * except for the flags field added to the end.
1178                          */
1179                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1180                                                &off.rx);
1181                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1182                                                &off.tx);
1183                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1184                                                &off.fr);
1185                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1186                                                &off.cr);
1187                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1188                                                 ptrs.flags);
1189                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1190                                                 ptrs.flags);
1191                         off.fr.flags = offsetof(struct xdp_umem_ring,
1192                                                 ptrs.flags);
1193                         off.cr.flags = offsetof(struct xdp_umem_ring,
1194                                                 ptrs.flags);
1195
1196                         len = sizeof(off);
1197                         to_copy = &off;
1198                 } else {
1199                         xsk_enter_rxtx_offsets(&off_v1.rx);
1200                         xsk_enter_rxtx_offsets(&off_v1.tx);
1201                         xsk_enter_umem_offsets(&off_v1.fr);
1202                         xsk_enter_umem_offsets(&off_v1.cr);
1203
1204                         len = sizeof(off_v1);
1205                         to_copy = &off_v1;
1206                 }
1207
1208                 if (copy_to_user(optval, to_copy, len))
1209                         return -EFAULT;
1210                 if (put_user(len, optlen))
1211                         return -EFAULT;
1212
1213                 return 0;
1214         }
1215         case XDP_OPTIONS:
1216         {
1217                 struct xdp_options opts = {};
1218
1219                 if (len < sizeof(opts))
1220                         return -EINVAL;
1221
1222                 mutex_lock(&xs->mutex);
1223                 if (xs->zc)
1224                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1225                 mutex_unlock(&xs->mutex);
1226
1227                 len = sizeof(opts);
1228                 if (copy_to_user(optval, &opts, len))
1229                         return -EFAULT;
1230                 if (put_user(len, optlen))
1231                         return -EFAULT;
1232
1233                 return 0;
1234         }
1235         default:
1236                 break;
1237         }
1238
1239         return -EOPNOTSUPP;
1240 }
1241
1242 static int xsk_mmap(struct file *file, struct socket *sock,
1243                     struct vm_area_struct *vma)
1244 {
1245         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1246         unsigned long size = vma->vm_end - vma->vm_start;
1247         struct xdp_sock *xs = xdp_sk(sock->sk);
1248         struct xsk_queue *q = NULL;
1249         unsigned long pfn;
1250         struct page *qpg;
1251
1252         if (READ_ONCE(xs->state) != XSK_READY)
1253                 return -EBUSY;
1254
1255         if (offset == XDP_PGOFF_RX_RING) {
1256                 q = READ_ONCE(xs->rx);
1257         } else if (offset == XDP_PGOFF_TX_RING) {
1258                 q = READ_ONCE(xs->tx);
1259         } else {
1260                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1261                 smp_rmb();
1262                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1263                         q = READ_ONCE(xs->fq_tmp);
1264                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1265                         q = READ_ONCE(xs->cq_tmp);
1266         }
1267
1268         if (!q)
1269                 return -EINVAL;
1270
1271         /* Matches the smp_wmb() in xsk_init_queue */
1272         smp_rmb();
1273         qpg = virt_to_head_page(q->ring);
1274         if (size > page_size(qpg))
1275                 return -EINVAL;
1276
1277         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1278         return remap_pfn_range(vma, vma->vm_start, pfn,
1279                                size, vma->vm_page_prot);
1280 }
1281
1282 static int xsk_notifier(struct notifier_block *this,
1283                         unsigned long msg, void *ptr)
1284 {
1285         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1286         struct net *net = dev_net(dev);
1287         struct sock *sk;
1288
1289         switch (msg) {
1290         case NETDEV_UNREGISTER:
1291                 mutex_lock(&net->xdp.lock);
1292                 sk_for_each(sk, &net->xdp.list) {
1293                         struct xdp_sock *xs = xdp_sk(sk);
1294
1295                         mutex_lock(&xs->mutex);
1296                         if (xs->dev == dev) {
1297                                 sk->sk_err = ENETDOWN;
1298                                 if (!sock_flag(sk, SOCK_DEAD))
1299                                         sk_error_report(sk);
1300
1301                                 xsk_unbind_dev(xs);
1302
1303                                 /* Clear device references. */
1304                                 xp_clear_dev(xs->pool);
1305                         }
1306                         mutex_unlock(&xs->mutex);
1307                 }
1308                 mutex_unlock(&net->xdp.lock);
1309                 break;
1310         }
1311         return NOTIFY_DONE;
1312 }
1313
1314 static struct proto xsk_proto = {
1315         .name =         "XDP",
1316         .owner =        THIS_MODULE,
1317         .obj_size =     sizeof(struct xdp_sock),
1318 };
1319
1320 static const struct proto_ops xsk_proto_ops = {
1321         .family         = PF_XDP,
1322         .owner          = THIS_MODULE,
1323         .release        = xsk_release,
1324         .bind           = xsk_bind,
1325         .connect        = sock_no_connect,
1326         .socketpair     = sock_no_socketpair,
1327         .accept         = sock_no_accept,
1328         .getname        = sock_no_getname,
1329         .poll           = xsk_poll,
1330         .ioctl          = sock_no_ioctl,
1331         .listen         = sock_no_listen,
1332         .shutdown       = sock_no_shutdown,
1333         .setsockopt     = xsk_setsockopt,
1334         .getsockopt     = xsk_getsockopt,
1335         .sendmsg        = xsk_sendmsg,
1336         .recvmsg        = xsk_recvmsg,
1337         .mmap           = xsk_mmap,
1338         .sendpage       = sock_no_sendpage,
1339 };
1340
1341 static void xsk_destruct(struct sock *sk)
1342 {
1343         struct xdp_sock *xs = xdp_sk(sk);
1344
1345         if (!sock_flag(sk, SOCK_DEAD))
1346                 return;
1347
1348         if (!xp_put_pool(xs->pool))
1349                 xdp_put_umem(xs->umem, !xs->pool);
1350
1351         sk_refcnt_debug_dec(sk);
1352 }
1353
1354 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1355                       int kern)
1356 {
1357         struct xdp_sock *xs;
1358         struct sock *sk;
1359
1360         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1361                 return -EPERM;
1362         if (sock->type != SOCK_RAW)
1363                 return -ESOCKTNOSUPPORT;
1364
1365         if (protocol)
1366                 return -EPROTONOSUPPORT;
1367
1368         sock->state = SS_UNCONNECTED;
1369
1370         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1371         if (!sk)
1372                 return -ENOBUFS;
1373
1374         sock->ops = &xsk_proto_ops;
1375
1376         sock_init_data(sock, sk);
1377
1378         sk->sk_family = PF_XDP;
1379
1380         sk->sk_destruct = xsk_destruct;
1381         sk_refcnt_debug_inc(sk);
1382
1383         sock_set_flag(sk, SOCK_RCU_FREE);
1384
1385         xs = xdp_sk(sk);
1386         xs->state = XSK_READY;
1387         mutex_init(&xs->mutex);
1388         spin_lock_init(&xs->rx_lock);
1389
1390         INIT_LIST_HEAD(&xs->map_list);
1391         spin_lock_init(&xs->map_list_lock);
1392
1393         mutex_lock(&net->xdp.lock);
1394         sk_add_node_rcu(sk, &net->xdp.list);
1395         mutex_unlock(&net->xdp.lock);
1396
1397         sock_prot_inuse_add(net, &xsk_proto, 1);
1398
1399         return 0;
1400 }
1401
1402 static const struct net_proto_family xsk_family_ops = {
1403         .family = PF_XDP,
1404         .create = xsk_create,
1405         .owner  = THIS_MODULE,
1406 };
1407
1408 static struct notifier_block xsk_netdev_notifier = {
1409         .notifier_call  = xsk_notifier,
1410 };
1411
1412 static int __net_init xsk_net_init(struct net *net)
1413 {
1414         mutex_init(&net->xdp.lock);
1415         INIT_HLIST_HEAD(&net->xdp.list);
1416         return 0;
1417 }
1418
1419 static void __net_exit xsk_net_exit(struct net *net)
1420 {
1421         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1422 }
1423
1424 static struct pernet_operations xsk_net_ops = {
1425         .init = xsk_net_init,
1426         .exit = xsk_net_exit,
1427 };
1428
1429 static int __init xsk_init(void)
1430 {
1431         int err, cpu;
1432
1433         err = proto_register(&xsk_proto, 0 /* no slab */);
1434         if (err)
1435                 goto out;
1436
1437         err = sock_register(&xsk_family_ops);
1438         if (err)
1439                 goto out_proto;
1440
1441         err = register_pernet_subsys(&xsk_net_ops);
1442         if (err)
1443                 goto out_sk;
1444
1445         err = register_netdevice_notifier(&xsk_netdev_notifier);
1446         if (err)
1447                 goto out_pernet;
1448
1449         for_each_possible_cpu(cpu)
1450                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1451         return 0;
1452
1453 out_pernet:
1454         unregister_pernet_subsys(&xsk_net_ops);
1455 out_sk:
1456         sock_unregister(PF_XDP);
1457 out_proto:
1458         proto_unregister(&xsk_proto);
1459 out:
1460         return err;
1461 }
1462
1463 fs_initcall(xsk_init);