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 <bjorn.topel@intel.com>
   9  *            Magnus Karlsson <magnus.karlsson@intel.com>
  10  */
  11
  12 #define pr_fmt(fmt) "AF_XDP: %s: " fmt, __func__
  13
  14 #include <linux/if_xdp.h>
  15 #include <linux/init.h>
  16 #include <linux/sched/mm.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/task.h>
  19 #include <linux/socket.h>
  20 #include <linux/file.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/net.h>
  23 #include <linux/netdevice.h>
  24 #include <linux/rculist.h>
  25 #include <net/xdp_sock_drv.h>
  26 #include <net/busy_poll.h>
  27 #include <net/xdp.h>
  28
  29 #include "xsk_queue.h"
  30 #include "xdp_umem.h"
  31 #include "xsk.h"
  32
  33 #define TX_BATCH_SIZE 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 -ENOMEM;
 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 -ENXIO;
 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
 406         return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 407 }
 408
 409 static void xsk_destruct_skb(struct sk_buff *skb)
 410 {
 411         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 412         struct xdp_sock *xs = xdp_sk(skb->sk);
 413         unsigned long flags;
 414
 415         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 416         xskq_prod_submit_addr(xs->pool->cq, addr);
 417         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 418
 419         sock_wfree(skb);
 420 }
 421
 422 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
 423                                               struct xdp_desc *desc)
 424 {
 425         struct xsk_buff_pool *pool = xs->pool;
 426         u32 hr, len, ts, offset, copy, copied;
 427         struct sk_buff *skb;
 428         struct page *page;
 429         void *buffer;
 430         int err, i;
 431         u64 addr;
 432
 433         hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
 434
 435         skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
 436         if (unlikely(!skb))
 437                 return ERR_PTR(err);
 438
 439         skb_reserve(skb, hr);
 440
 441         addr = desc->addr;
 442         len = desc->len;
 443         ts = pool->unaligned ? len : pool->chunk_size;
 444
 445         buffer = xsk_buff_raw_get_data(pool, addr);
 446         offset = offset_in_page(buffer);
 447         addr = buffer - pool->addrs;
 448
 449         for (copied = 0, i = 0; copied < len; i++) {
 450                 page = pool->umem->pgs[addr >> PAGE_SHIFT];
 451                 get_page(page);
 452
 453                 copy = min_t(u32, PAGE_SIZE - offset, len - copied);
 454                 skb_fill_page_desc(skb, i, page, offset, copy);
 455
 456                 copied += copy;
 457                 addr += copy;
 458                 offset = 0;
 459         }
 460
 461         skb->len += len;
 462         skb->data_len += len;
 463         skb->truesize += ts;
 464
 465         refcount_add(ts, &xs->sk.sk_wmem_alloc);
 466
 467         return skb;
 468 }
 469
 470 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
 471                                      struct xdp_desc *desc)
 472 {
 473         struct net_device *dev = xs->dev;
 474         struct sk_buff *skb;
 475
 476         if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
 477                 skb = xsk_build_skb_zerocopy(xs, desc);
 478                 if (IS_ERR(skb))
 479                         return skb;
 480         } else {
 481                 u32 hr, tr, len;
 482                 void *buffer;
 483                 int err;
 484
 485                 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
 486                 tr = dev->needed_tailroom;
 487                 len = desc->len;
 488
 489                 skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
 490                 if (unlikely(!skb))
 491                         return ERR_PTR(err);
 492
 493                 skb_reserve(skb, hr);
 494                 skb_put(skb, len);
 495
 496                 buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
 497                 err = skb_store_bits(skb, 0, buffer, len);
 498                 if (unlikely(err)) {
 499                         kfree_skb(skb);
 500                         return ERR_PTR(err);
 501                 }
 502         }
 503
 504         skb->dev = dev;
 505         skb->priority = xs->sk.sk_priority;
 506         skb->mark = xs->sk.sk_mark;
 507         skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
 508         skb->destructor = xsk_destruct_skb;
 509
 510         return skb;
 511 }
 512
 513 static int xsk_generic_xmit(struct sock *sk)
 514 {
 515         struct xdp_sock *xs = xdp_sk(sk);
 516         u32 max_batch = TX_BATCH_SIZE;
 517         bool sent_frame = false;
 518         struct xdp_desc desc;
 519         struct sk_buff *skb;
 520         unsigned long flags;
 521         int err = 0;
 522
 523         mutex_lock(&xs->mutex);
 524
 525         /* Since we dropped the RCU read lock, the socket state might have changed. */
 526         if (unlikely(!xsk_is_bound(xs))) {
 527                 err = -ENXIO;
 528                 goto out;
 529         }
 530
 531         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 532                 goto out;
 533
 534         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 535                 if (max_batch-- == 0) {
 536                         err = -EAGAIN;
 537                         goto out;
 538                 }
 539
 540                 skb = xsk_build_skb(xs, &desc);
 541                 if (IS_ERR(skb)) {
 542                         err = PTR_ERR(skb);
 543                         goto out;
 544                 }
 545
 546                 /* This is the backpressure mechanism for the Tx path.
 547                  * Reserve space in the completion queue and only proceed
 548                  * if there is space in it. This avoids having to implement
 549                  * any buffering in the Tx path.
 550                  */
 551                 spin_lock_irqsave(&xs->pool->cq_lock, flags);
 552                 if (xskq_prod_reserve(xs->pool->cq)) {
 553                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 554                         kfree_skb(skb);
 555                         goto out;
 556                 }
 557                 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 558
 559                 err = __dev_direct_xmit(skb, xs->queue_id);
 560                 if  (err == NETDEV_TX_BUSY) {
 561                         /* Tell user-space to retry the send */
 562                         skb->destructor = sock_wfree;
 563                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 564                         xskq_prod_cancel(xs->pool->cq);
 565                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 566                         /* Free skb without triggering the perf drop trace */
 567                         consume_skb(skb);
 568                         err = -EAGAIN;
 569                         goto out;
 570                 }
 571
 572                 xskq_cons_release(xs->tx);
 573                 /* Ignore NET_XMIT_CN as packet might have been sent */
 574                 if (err == NET_XMIT_DROP) {
 575                         /* SKB completed but not sent */
 576                         err = -EBUSY;
 577                         goto out;
 578                 }
 579
 580                 sent_frame = true;
 581         }
 582
 583         xs->tx->queue_empty_descs++;
 584
 585 out:
 586         if (sent_frame)
 587                 if (xsk_tx_writeable(xs))
 588                         sk->sk_write_space(sk);
 589
 590         mutex_unlock(&xs->mutex);
 591         return err;
 592 }
 593
 594 static int xsk_xmit(struct sock *sk)
 595 {
 596         struct xdp_sock *xs = xdp_sk(sk);
 597         int ret;
 598
 599         if (unlikely(!(xs->dev->flags & IFF_UP)))
 600                 return -ENETDOWN;
 601         if (unlikely(!xs->tx))
 602                 return -ENOBUFS;
 603
 604         if (xs->zc)
 605                 return xsk_wakeup(xs, XDP_WAKEUP_TX);
 606
 607         /* Drop the RCU lock since the SKB path might sleep. */
 608         rcu_read_unlock();
 609         ret = xsk_generic_xmit(sk);
 610         /* Reaquire RCU lock before going into common code. */
 611         rcu_read_lock();
 612
 613         return ret;
 614 }
 615
 616 static bool xsk_no_wakeup(struct sock *sk)
 617 {
 618 #ifdef CONFIG_NET_RX_BUSY_POLL
 619         /* Prefer busy-polling, skip the wakeup. */
 620         return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
 621                 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
 622 #else
 623         return false;
 624 #endif
 625 }
 626
 627 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 628 {
 629         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 630         struct sock *sk = sock->sk;
 631         struct xdp_sock *xs = xdp_sk(sk);
 632         struct xsk_buff_pool *pool;
 633
 634         if (unlikely(!xsk_is_bound(xs)))
 635                 return -ENXIO;
 636         if (unlikely(need_wait))
 637                 return -EOPNOTSUPP;
 638
 639         if (sk_can_busy_loop(sk))
 640                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 641
 642         if (xs->zc && xsk_no_wakeup(sk))
 643                 return 0;
 644
 645         pool = xs->pool;
 646         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
 647                 return xsk_xmit(sk);
 648         return 0;
 649 }
 650
 651 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 652 {
 653         int ret;
 654
 655         rcu_read_lock();
 656         ret = __xsk_sendmsg(sock, m, total_len);
 657         rcu_read_unlock();
 658
 659         return ret;
 660 }
 661
 662 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 663 {
 664         bool need_wait = !(flags & MSG_DONTWAIT);
 665         struct sock *sk = sock->sk;
 666         struct xdp_sock *xs = xdp_sk(sk);
 667
 668         if (unlikely(!xsk_is_bound(xs)))
 669                 return -ENXIO;
 670         if (unlikely(!(xs->dev->flags & IFF_UP)))
 671                 return -ENETDOWN;
 672         if (unlikely(!xs->rx))
 673                 return -ENOBUFS;
 674         if (unlikely(need_wait))
 675                 return -EOPNOTSUPP;
 676
 677         if (sk_can_busy_loop(sk))
 678                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 679
 680         if (xsk_no_wakeup(sk))
 681                 return 0;
 682
 683         if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
 684                 return xsk_wakeup(xs, XDP_WAKEUP_RX);
 685         return 0;
 686 }
 687
 688 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 689 {
 690         int ret;
 691
 692         rcu_read_lock();
 693         ret = __xsk_recvmsg(sock, m, len, flags);
 694         rcu_read_unlock();
 695
 696         return ret;
 697 }
 698
 699 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 700                              struct poll_table_struct *wait)
 701 {
 702         __poll_t mask = 0;
 703         struct sock *sk = sock->sk;
 704         struct xdp_sock *xs = xdp_sk(sk);
 705         struct xsk_buff_pool *pool;
 706
 707         sock_poll_wait(file, sock, wait);
 708
 709         rcu_read_lock();
 710         if (unlikely(!xsk_is_bound(xs))) {
 711                 rcu_read_unlock();
 712                 return mask;
 713         }
 714
 715         pool = xs->pool;
 716
 717         if (pool->cached_need_wakeup) {
 718                 if (xs->zc)
 719                         xsk_wakeup(xs, pool->cached_need_wakeup);
 720                 else
 721                         /* Poll needs to drive Tx also in copy mode */
 722                         xsk_xmit(sk);
 723         }
 724
 725         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 726                 mask |= EPOLLIN | EPOLLRDNORM;
 727         if (xs->tx && xsk_tx_writeable(xs))
 728                 mask |= EPOLLOUT | EPOLLWRNORM;
 729
 730         rcu_read_unlock();
 731         return mask;
 732 }
 733
 734 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 735                           bool umem_queue)
 736 {
 737         struct xsk_queue *q;
 738
 739         if (entries == 0 || *queue || !is_power_of_2(entries))
 740                 return -EINVAL;
 741
 742         q = xskq_create(entries, umem_queue);
 743         if (!q)
 744                 return -ENOMEM;
 745
 746         /* Make sure queue is ready before it can be seen by others */
 747         smp_wmb();
 748         WRITE_ONCE(*queue, q);
 749         return 0;
 750 }
 751
 752 static void xsk_unbind_dev(struct xdp_sock *xs)
 753 {
 754         struct net_device *dev = xs->dev;
 755
 756         if (xs->state != XSK_BOUND)
 757                 return;
 758         WRITE_ONCE(xs->state, XSK_UNBOUND);
 759
 760         /* Wait for driver to stop using the xdp socket. */
 761         xp_del_xsk(xs->pool, xs);
 762         synchronize_net();
 763         dev_put(dev);
 764 }
 765
 766 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 767                                               struct xdp_sock __rcu ***map_entry)
 768 {
 769         struct xsk_map *map = NULL;
 770         struct xsk_map_node *node;
 771
 772         *map_entry = NULL;
 773
 774         spin_lock_bh(&xs->map_list_lock);
 775         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 776                                         node);
 777         if (node) {
 778                 bpf_map_inc(&node->map->map);
 779                 map = node->map;
 780                 *map_entry = node->map_entry;
 781         }
 782         spin_unlock_bh(&xs->map_list_lock);
 783         return map;
 784 }
 785
 786 static void xsk_delete_from_maps(struct xdp_sock *xs)
 787 {
 788         /* This function removes the current XDP socket from all the
 789          * maps it resides in. We need to take extra care here, due to
 790          * the two locks involved. Each map has a lock synchronizing
 791          * updates to the entries, and each socket has a lock that
 792          * synchronizes access to the list of maps (map_list). For
 793          * deadlock avoidance the locks need to be taken in the order
 794          * "map lock"->"socket map list lock". We start off by
 795          * accessing the socket map list, and take a reference to the
 796          * map to guarantee existence between the
 797          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 798          * calls. Then we ask the map to remove the socket, which
 799          * tries to remove the socket from the map. Note that there
 800          * might be updates to the map between
 801          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 802          */
 803         struct xdp_sock __rcu **map_entry = NULL;
 804         struct xsk_map *map;
 805
 806         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 807                 xsk_map_try_sock_delete(map, xs, map_entry);
 808                 bpf_map_put(&map->map);
 809         }
 810 }
 811
 812 static int xsk_release(struct socket *sock)
 813 {
 814         struct sock *sk = sock->sk;
 815         struct xdp_sock *xs = xdp_sk(sk);
 816         struct net *net;
 817
 818         if (!sk)
 819                 return 0;
 820
 821         net = sock_net(sk);
 822
 823         mutex_lock(&net->xdp.lock);
 824         sk_del_node_init_rcu(sk);
 825         mutex_unlock(&net->xdp.lock);
 826
 827         sock_prot_inuse_add(net, sk->sk_prot, -1);
 828
 829         xsk_delete_from_maps(xs);
 830         mutex_lock(&xs->mutex);
 831         xsk_unbind_dev(xs);
 832         mutex_unlock(&xs->mutex);
 833
 834         xskq_destroy(xs->rx);
 835         xskq_destroy(xs->tx);
 836         xskq_destroy(xs->fq_tmp);
 837         xskq_destroy(xs->cq_tmp);
 838
 839         sock_orphan(sk);
 840         sock->sk = NULL;
 841
 842         sk_refcnt_debug_release(sk);
 843         sock_put(sk);
 844
 845         return 0;
 846 }
 847
 848 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 849 {
 850         struct socket *sock;
 851         int err;
 852
 853         sock = sockfd_lookup(fd, &err);
 854         if (!sock)
 855                 return ERR_PTR(-ENOTSOCK);
 856
 857         if (sock->sk->sk_family != PF_XDP) {
 858                 sockfd_put(sock);
 859                 return ERR_PTR(-ENOPROTOOPT);
 860         }
 861
 862         return sock;
 863 }
 864
 865 static bool xsk_validate_queues(struct xdp_sock *xs)
 866 {
 867         return xs->fq_tmp && xs->cq_tmp;
 868 }
 869
 870 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 871 {
 872         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 873         struct sock *sk = sock->sk;
 874         struct xdp_sock *xs = xdp_sk(sk);
 875         struct net_device *dev;
 876         u32 flags, qid;
 877         int err = 0;
 878
 879         if (addr_len < sizeof(struct sockaddr_xdp))
 880                 return -EINVAL;
 881         if (sxdp->sxdp_family != AF_XDP)
 882                 return -EINVAL;
 883
 884         flags = sxdp->sxdp_flags;
 885         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 886                       XDP_USE_NEED_WAKEUP))
 887                 return -EINVAL;
 888
 889         rtnl_lock();
 890         mutex_lock(&xs->mutex);
 891         if (xs->state != XSK_READY) {
 892                 err = -EBUSY;
 893                 goto out_release;
 894         }
 895
 896         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 897         if (!dev) {
 898                 err = -ENODEV;
 899                 goto out_release;
 900         }
 901
 902         if (!xs->rx && !xs->tx) {
 903                 err = -EINVAL;
 904                 goto out_unlock;
 905         }
 906
 907         qid = sxdp->sxdp_queue_id;
 908
 909         if (flags & XDP_SHARED_UMEM) {
 910                 struct xdp_sock *umem_xs;
 911                 struct socket *sock;
 912
 913                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 914                     (flags & XDP_USE_NEED_WAKEUP)) {
 915                         /* Cannot specify flags for shared sockets. */
 916                         err = -EINVAL;
 917                         goto out_unlock;
 918                 }
 919
 920                 if (xs->umem) {
 921                         /* We have already our own. */
 922                         err = -EINVAL;
 923                         goto out_unlock;
 924                 }
 925
 926                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 927                 if (IS_ERR(sock)) {
 928                         err = PTR_ERR(sock);
 929                         goto out_unlock;
 930                 }
 931
 932                 umem_xs = xdp_sk(sock->sk);
 933                 if (!xsk_is_bound(umem_xs)) {
 934                         err = -EBADF;
 935                         sockfd_put(sock);
 936                         goto out_unlock;
 937                 }
 938
 939                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 940                         /* Share the umem with another socket on another qid
 941                          * and/or device.
 942                          */
 943                         xs->pool = xp_create_and_assign_umem(xs,
 944                                                              umem_xs->umem);
 945                         if (!xs->pool) {
 946                                 err = -ENOMEM;
 947                                 sockfd_put(sock);
 948                                 goto out_unlock;
 949                         }
 950
 951                         err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
 952                                                    dev, qid);
 953                         if (err) {
 954                                 xp_destroy(xs->pool);
 955                                 xs->pool = NULL;
 956                                 sockfd_put(sock);
 957                                 goto out_unlock;
 958                         }
 959                 } else {
 960                         /* Share the buffer pool with the other socket. */
 961                         if (xs->fq_tmp || xs->cq_tmp) {
 962                                 /* Do not allow setting your own fq or cq. */
 963                                 err = -EINVAL;
 964                                 sockfd_put(sock);
 965                                 goto out_unlock;
 966                         }
 967
 968                         xp_get_pool(umem_xs->pool);
 969                         xs->pool = umem_xs->pool;
 970
 971                         /* If underlying shared umem was created without Tx
 972                          * ring, allocate Tx descs array that Tx batching API
 973                          * utilizes
 974                          */
 975                         if (xs->tx && !xs->pool->tx_descs) {
 976                                 err = xp_alloc_tx_descs(xs->pool, xs);
 977                                 if (err) {
 978                                         xp_put_pool(xs->pool);
 979                                         sockfd_put(sock);
 980                                         goto out_unlock;
 981                                 }
 982                         }
 983                 }
 984
 985                 xdp_get_umem(umem_xs->umem);
 986                 WRITE_ONCE(xs->umem, umem_xs->umem);
 987                 sockfd_put(sock);
 988         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 989                 err = -EINVAL;
 990                 goto out_unlock;
 991         } else {
 992                 /* This xsk has its own umem. */
 993                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 994                 if (!xs->pool) {
 995                         err = -ENOMEM;
 996                         goto out_unlock;
 997                 }
 998
 999                 err = xp_assign_dev(xs->pool, dev, qid, flags);
1000                 if (err) {
1001                         xp_destroy(xs->pool);
1002                         xs->pool = NULL;
1003                         goto out_unlock;
1004                 }
1005         }
1006
1007         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1008         xs->fq_tmp = NULL;
1009         xs->cq_tmp = NULL;
1010
1011         xs->dev = dev;
1012         xs->zc = xs->umem->zc;
1013         xs->queue_id = qid;
1014         xp_add_xsk(xs->pool, xs);
1015
1016 out_unlock:
1017         if (err) {
1018                 dev_put(dev);
1019         } else {
1020                 /* Matches smp_rmb() in bind() for shared umem
1021                  * sockets, and xsk_is_bound().
1022                  */
1023                 smp_wmb();
1024                 WRITE_ONCE(xs->state, XSK_BOUND);
1025         }
1026 out_release:
1027         mutex_unlock(&xs->mutex);
1028         rtnl_unlock();
1029         return err;
1030 }
1031
1032 struct xdp_umem_reg_v1 {
1033         __u64 addr; /* Start of packet data area */
1034         __u64 len; /* Length of packet data area */
1035         __u32 chunk_size;
1036         __u32 headroom;
1037 };
1038
1039 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1040                           sockptr_t optval, unsigned int optlen)
1041 {
1042         struct sock *sk = sock->sk;
1043         struct xdp_sock *xs = xdp_sk(sk);
1044         int err;
1045
1046         if (level != SOL_XDP)
1047                 return -ENOPROTOOPT;
1048
1049         switch (optname) {
1050         case XDP_RX_RING:
1051         case XDP_TX_RING:
1052         {
1053                 struct xsk_queue **q;
1054                 int entries;
1055
1056                 if (optlen < sizeof(entries))
1057                         return -EINVAL;
1058                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1059                         return -EFAULT;
1060
1061                 mutex_lock(&xs->mutex);
1062                 if (xs->state != XSK_READY) {
1063                         mutex_unlock(&xs->mutex);
1064                         return -EBUSY;
1065                 }
1066                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1067                 err = xsk_init_queue(entries, q, false);
1068                 if (!err && optname == XDP_TX_RING)
1069                         /* Tx needs to be explicitly woken up the first time */
1070                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1071                 mutex_unlock(&xs->mutex);
1072                 return err;
1073         }
1074         case XDP_UMEM_REG:
1075         {
1076                 size_t mr_size = sizeof(struct xdp_umem_reg);
1077                 struct xdp_umem_reg mr = {};
1078                 struct xdp_umem *umem;
1079
1080                 if (optlen < sizeof(struct xdp_umem_reg_v1))
1081                         return -EINVAL;
1082                 else if (optlen < sizeof(mr))
1083                         mr_size = sizeof(struct xdp_umem_reg_v1);
1084
1085                 if (copy_from_sockptr(&mr, optval, mr_size))
1086                         return -EFAULT;
1087
1088                 mutex_lock(&xs->mutex);
1089                 if (xs->state != XSK_READY || xs->umem) {
1090                         mutex_unlock(&xs->mutex);
1091                         return -EBUSY;
1092                 }
1093
1094                 umem = xdp_umem_create(&mr);
1095                 if (IS_ERR(umem)) {
1096                         mutex_unlock(&xs->mutex);
1097                         return PTR_ERR(umem);
1098                 }
1099
1100                 /* Make sure umem is ready before it can be seen by others */
1101                 smp_wmb();
1102                 WRITE_ONCE(xs->umem, umem);
1103                 mutex_unlock(&xs->mutex);
1104                 return 0;
1105         }
1106         case XDP_UMEM_FILL_RING:
1107         case XDP_UMEM_COMPLETION_RING:
1108         {
1109                 struct xsk_queue **q;
1110                 int entries;
1111
1112                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1113                         return -EFAULT;
1114
1115                 mutex_lock(&xs->mutex);
1116                 if (xs->state != XSK_READY) {
1117                         mutex_unlock(&xs->mutex);
1118                         return -EBUSY;
1119                 }
1120
1121                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1122                         &xs->cq_tmp;
1123                 err = xsk_init_queue(entries, q, true);
1124                 mutex_unlock(&xs->mutex);
1125                 return err;
1126         }
1127         default:
1128                 break;
1129         }
1130
1131         return -ENOPROTOOPT;
1132 }
1133
1134 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1135 {
1136         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1137         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1138         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1139 }
1140
1141 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1142 {
1143         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1144         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1145         ring->desc = offsetof(struct xdp_umem_ring, desc);
1146 }
1147
1148 struct xdp_statistics_v1 {
1149         __u64 rx_dropped;
1150         __u64 rx_invalid_descs;
1151         __u64 tx_invalid_descs;
1152 };
1153
1154 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1155                           char __user *optval, int __user *optlen)
1156 {
1157         struct sock *sk = sock->sk;
1158         struct xdp_sock *xs = xdp_sk(sk);
1159         int len;
1160
1161         if (level != SOL_XDP)
1162                 return -ENOPROTOOPT;
1163
1164         if (get_user(len, optlen))
1165                 return -EFAULT;
1166         if (len < 0)
1167                 return -EINVAL;
1168
1169         switch (optname) {
1170         case XDP_STATISTICS:
1171         {
1172                 struct xdp_statistics stats = {};
1173                 bool extra_stats = true;
1174                 size_t stats_size;
1175
1176                 if (len < sizeof(struct xdp_statistics_v1)) {
1177                         return -EINVAL;
1178                 } else if (len < sizeof(stats)) {
1179                         extra_stats = false;
1180                         stats_size = sizeof(struct xdp_statistics_v1);
1181                 } else {
1182                         stats_size = sizeof(stats);
1183                 }
1184
1185                 mutex_lock(&xs->mutex);
1186                 stats.rx_dropped = xs->rx_dropped;
1187                 if (extra_stats) {
1188                         stats.rx_ring_full = xs->rx_queue_full;
1189                         stats.rx_fill_ring_empty_descs =
1190                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1191                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1192                 } else {
1193                         stats.rx_dropped += xs->rx_queue_full;
1194                 }
1195                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1196                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1197                 mutex_unlock(&xs->mutex);
1198
1199                 if (copy_to_user(optval, &stats, stats_size))
1200                         return -EFAULT;
1201                 if (put_user(stats_size, optlen))
1202                         return -EFAULT;
1203
1204                 return 0;
1205         }
1206         case XDP_MMAP_OFFSETS:
1207         {
1208                 struct xdp_mmap_offsets off;
1209                 struct xdp_mmap_offsets_v1 off_v1;
1210                 bool flags_supported = true;
1211                 void *to_copy;
1212
1213                 if (len < sizeof(off_v1))
1214                         return -EINVAL;
1215                 else if (len < sizeof(off))
1216                         flags_supported = false;
1217
1218                 if (flags_supported) {
1219                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1220                          * except for the flags field added to the end.
1221                          */
1222                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1223                                                &off.rx);
1224                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1225                                                &off.tx);
1226                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1227                                                &off.fr);
1228                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1229                                                &off.cr);
1230                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1231                                                 ptrs.flags);
1232                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1233                                                 ptrs.flags);
1234                         off.fr.flags = offsetof(struct xdp_umem_ring,
1235                                                 ptrs.flags);
1236                         off.cr.flags = offsetof(struct xdp_umem_ring,
1237                                                 ptrs.flags);
1238
1239                         len = sizeof(off);
1240                         to_copy = &off;
1241                 } else {
1242                         xsk_enter_rxtx_offsets(&off_v1.rx);
1243                         xsk_enter_rxtx_offsets(&off_v1.tx);
1244                         xsk_enter_umem_offsets(&off_v1.fr);
1245                         xsk_enter_umem_offsets(&off_v1.cr);
1246
1247                         len = sizeof(off_v1);
1248                         to_copy = &off_v1;
1249                 }
1250
1251                 if (copy_to_user(optval, to_copy, len))
1252                         return -EFAULT;
1253                 if (put_user(len, optlen))
1254                         return -EFAULT;
1255
1256                 return 0;
1257         }
1258         case XDP_OPTIONS:
1259         {
1260                 struct xdp_options opts = {};
1261
1262                 if (len < sizeof(opts))
1263                         return -EINVAL;
1264
1265                 mutex_lock(&xs->mutex);
1266                 if (xs->zc)
1267                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1268                 mutex_unlock(&xs->mutex);
1269
1270                 len = sizeof(opts);
1271                 if (copy_to_user(optval, &opts, len))
1272                         return -EFAULT;
1273                 if (put_user(len, optlen))
1274                         return -EFAULT;
1275
1276                 return 0;
1277         }
1278         default:
1279                 break;
1280         }
1281
1282         return -EOPNOTSUPP;
1283 }
1284
1285 static int xsk_mmap(struct file *file, struct socket *sock,
1286                     struct vm_area_struct *vma)
1287 {
1288         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1289         unsigned long size = vma->vm_end - vma->vm_start;
1290         struct xdp_sock *xs = xdp_sk(sock->sk);
1291         struct xsk_queue *q = NULL;
1292         unsigned long pfn;
1293         struct page *qpg;
1294
1295         if (READ_ONCE(xs->state) != XSK_READY)
1296                 return -EBUSY;
1297
1298         if (offset == XDP_PGOFF_RX_RING) {
1299                 q = READ_ONCE(xs->rx);
1300         } else if (offset == XDP_PGOFF_TX_RING) {
1301                 q = READ_ONCE(xs->tx);
1302         } else {
1303                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1304                 smp_rmb();
1305                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1306                         q = READ_ONCE(xs->fq_tmp);
1307                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1308                         q = READ_ONCE(xs->cq_tmp);
1309         }
1310
1311         if (!q)
1312                 return -EINVAL;
1313
1314         /* Matches the smp_wmb() in xsk_init_queue */
1315         smp_rmb();
1316         qpg = virt_to_head_page(q->ring);
1317         if (size > page_size(qpg))
1318                 return -EINVAL;
1319
1320         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1321         return remap_pfn_range(vma, vma->vm_start, pfn,
1322                                size, vma->vm_page_prot);
1323 }
1324
1325 static int xsk_notifier(struct notifier_block *this,
1326                         unsigned long msg, void *ptr)
1327 {
1328         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1329         struct net *net = dev_net(dev);
1330         struct sock *sk;
1331
1332         switch (msg) {
1333         case NETDEV_UNREGISTER:
1334                 mutex_lock(&net->xdp.lock);
1335                 sk_for_each(sk, &net->xdp.list) {
1336                         struct xdp_sock *xs = xdp_sk(sk);
1337
1338                         mutex_lock(&xs->mutex);
1339                         if (xs->dev == dev) {
1340                                 sk->sk_err = ENETDOWN;
1341                                 if (!sock_flag(sk, SOCK_DEAD))
1342                                         sk_error_report(sk);
1343
1344                                 xsk_unbind_dev(xs);
1345
1346                                 /* Clear device references. */
1347                                 xp_clear_dev(xs->pool);
1348                         }
1349                         mutex_unlock(&xs->mutex);
1350                 }
1351                 mutex_unlock(&net->xdp.lock);
1352                 break;
1353         }
1354         return NOTIFY_DONE;
1355 }
1356
1357 static struct proto xsk_proto = {
1358         .name =         "XDP",
1359         .owner =        THIS_MODULE,
1360         .obj_size =     sizeof(struct xdp_sock),
1361 };
1362
1363 static const struct proto_ops xsk_proto_ops = {
1364         .family         = PF_XDP,
1365         .owner          = THIS_MODULE,
1366         .release        = xsk_release,
1367         .bind           = xsk_bind,
1368         .connect        = sock_no_connect,
1369         .socketpair     = sock_no_socketpair,
1370         .accept         = sock_no_accept,
1371         .getname        = sock_no_getname,
1372         .poll           = xsk_poll,
1373         .ioctl          = sock_no_ioctl,
1374         .listen         = sock_no_listen,
1375         .shutdown       = sock_no_shutdown,
1376         .setsockopt     = xsk_setsockopt,
1377         .getsockopt     = xsk_getsockopt,
1378         .sendmsg        = xsk_sendmsg,
1379         .recvmsg        = xsk_recvmsg,
1380         .mmap           = xsk_mmap,
1381         .sendpage       = sock_no_sendpage,
1382 };
1383
1384 static void xsk_destruct(struct sock *sk)
1385 {
1386         struct xdp_sock *xs = xdp_sk(sk);
1387
1388         if (!sock_flag(sk, SOCK_DEAD))
1389                 return;
1390
1391         if (!xp_put_pool(xs->pool))
1392                 xdp_put_umem(xs->umem, !xs->pool);
1393
1394         sk_refcnt_debug_dec(sk);
1395 }
1396
1397 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1398                       int kern)
1399 {
1400         struct xdp_sock *xs;
1401         struct sock *sk;
1402
1403         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1404                 return -EPERM;
1405         if (sock->type != SOCK_RAW)
1406                 return -ESOCKTNOSUPPORT;
1407
1408         if (protocol)
1409                 return -EPROTONOSUPPORT;
1410
1411         sock->state = SS_UNCONNECTED;
1412
1413         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1414         if (!sk)
1415                 return -ENOBUFS;
1416
1417         sock->ops = &xsk_proto_ops;
1418
1419         sock_init_data(sock, sk);
1420
1421         sk->sk_family = PF_XDP;
1422
1423         sk->sk_destruct = xsk_destruct;
1424         sk_refcnt_debug_inc(sk);
1425
1426         sock_set_flag(sk, SOCK_RCU_FREE);
1427
1428         xs = xdp_sk(sk);
1429         xs->state = XSK_READY;
1430         mutex_init(&xs->mutex);
1431         spin_lock_init(&xs->rx_lock);
1432
1433         INIT_LIST_HEAD(&xs->map_list);
1434         spin_lock_init(&xs->map_list_lock);
1435
1436         mutex_lock(&net->xdp.lock);
1437         sk_add_node_rcu(sk, &net->xdp.list);
1438         mutex_unlock(&net->xdp.lock);
1439
1440         sock_prot_inuse_add(net, &xsk_proto, 1);
1441
1442         return 0;
1443 }
1444
1445 static const struct net_proto_family xsk_family_ops = {
1446         .family = PF_XDP,
1447         .create = xsk_create,
1448         .owner  = THIS_MODULE,
1449 };
1450
1451 static struct notifier_block xsk_netdev_notifier = {
1452         .notifier_call  = xsk_notifier,
1453 };
1454
1455 static int __net_init xsk_net_init(struct net *net)
1456 {
1457         mutex_init(&net->xdp.lock);
1458         INIT_HLIST_HEAD(&net->xdp.list);
1459         return 0;
1460 }
1461
1462 static void __net_exit xsk_net_exit(struct net *net)
1463 {
1464         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1465 }
1466
1467 static struct pernet_operations xsk_net_ops = {
1468         .init = xsk_net_init,
1469         .exit = xsk_net_exit,
1470 };
1471
1472 static int __init xsk_init(void)
1473 {
1474         int err, cpu;
1475
1476         err = proto_register(&xsk_proto, 0 /* no slab */);
1477         if (err)
1478                 goto out;
1479
1480         err = sock_register(&xsk_family_ops);
1481         if (err)
1482                 goto out_proto;
1483
1484         err = register_pernet_subsys(&xsk_net_ops);
1485         if (err)
1486                 goto out_sk;
1487
1488         err = register_netdevice_notifier(&xsk_netdev_notifier);
1489         if (err)
1490                 goto out_pernet;
1491
1492         for_each_possible_cpu(cpu)
1493                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1494         return 0;
1495
1496 out_pernet:
1497         unregister_pernet_subsys(&xsk_net_ops);
1498 out_sk:
1499         sock_unregister(PF_XDP);
1500 out_proto:
1501         proto_unregister(&xsk_proto);
1502 out:
1503         return err;
1504 }
1505
1506 fs_initcall(xsk_init);