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 -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         max_entries = xskq_cons_nb_entries(xs->tx, max_entries);
 377         nb_pkts = xskq_cons_read_desc_batch(xs->tx, 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, pool->tx_descs, nb_pkts);
 390         if (!nb_pkts)
 391                 goto out;
 392
 393         xskq_cons_release_n(xs->tx, max_entries);
 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
 407         return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 408 }
 409
 410 static void xsk_destruct_skb(struct sk_buff *skb)
 411 {
 412         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 413         struct xdp_sock *xs = xdp_sk(skb->sk);
 414         unsigned long flags;
 415
 416         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 417         xskq_prod_submit_addr(xs->pool->cq, addr);
 418         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 419
 420         sock_wfree(skb);
 421 }
 422
 423 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
 424                                               struct xdp_desc *desc)
 425 {
 426         struct xsk_buff_pool *pool = xs->pool;
 427         u32 hr, len, ts, offset, copy, copied;
 428         struct sk_buff *skb;
 429         struct page *page;
 430         void *buffer;
 431         int err, i;
 432         u64 addr;
 433
 434         hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
 435
 436         skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
 437         if (unlikely(!skb))
 438                 return ERR_PTR(err);
 439
 440         skb_reserve(skb, hr);
 441
 442         addr = desc->addr;
 443         len = desc->len;
 444         ts = pool->unaligned ? len : pool->chunk_size;
 445
 446         buffer = xsk_buff_raw_get_data(pool, addr);
 447         offset = offset_in_page(buffer);
 448         addr = buffer - pool->addrs;
 449
 450         for (copied = 0, i = 0; copied < len; i++) {
 451                 page = pool->umem->pgs[addr >> PAGE_SHIFT];
 452                 get_page(page);
 453
 454                 copy = min_t(u32, PAGE_SIZE - offset, len - copied);
 455                 skb_fill_page_desc(skb, i, page, offset, copy);
 456
 457                 copied += copy;
 458                 addr += copy;
 459                 offset = 0;
 460         }
 461
 462         skb->len += len;
 463         skb->data_len += len;
 464         skb->truesize += ts;
 465
 466         refcount_add(ts, &xs->sk.sk_wmem_alloc);
 467
 468         return skb;
 469 }
 470
 471 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
 472                                      struct xdp_desc *desc)
 473 {
 474         struct net_device *dev = xs->dev;
 475         struct sk_buff *skb;
 476
 477         if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
 478                 skb = xsk_build_skb_zerocopy(xs, desc);
 479                 if (IS_ERR(skb))
 480                         return skb;
 481         } else {
 482                 u32 hr, tr, len;
 483                 void *buffer;
 484                 int err;
 485
 486                 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
 487                 tr = dev->needed_tailroom;
 488                 len = desc->len;
 489
 490                 skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
 491                 if (unlikely(!skb))
 492                         return ERR_PTR(err);
 493
 494                 skb_reserve(skb, hr);
 495                 skb_put(skb, len);
 496
 497                 buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
 498                 err = skb_store_bits(skb, 0, buffer, len);
 499                 if (unlikely(err)) {
 500                         kfree_skb(skb);
 501                         return ERR_PTR(err);
 502                 }
 503         }
 504
 505         skb->dev = dev;
 506         skb->priority = xs->sk.sk_priority;
 507         skb->mark = xs->sk.sk_mark;
 508         skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
 509         skb->destructor = xsk_destruct_skb;
 510
 511         return skb;
 512 }
 513
 514 static int xsk_generic_xmit(struct sock *sk)
 515 {
 516         struct xdp_sock *xs = xdp_sk(sk);
 517         u32 max_batch = TX_BATCH_SIZE;
 518         bool sent_frame = false;
 519         struct xdp_desc desc;
 520         struct sk_buff *skb;
 521         unsigned long flags;
 522         int err = 0;
 523
 524         mutex_lock(&xs->mutex);
 525
 526         /* Since we dropped the RCU read lock, the socket state might have changed. */
 527         if (unlikely(!xsk_is_bound(xs))) {
 528                 err = -ENXIO;
 529                 goto out;
 530         }
 531
 532         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 533                 goto out;
 534
 535         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 536                 if (max_batch-- == 0) {
 537                         err = -EAGAIN;
 538                         goto out;
 539                 }
 540
 541                 skb = xsk_build_skb(xs, &desc);
 542                 if (IS_ERR(skb)) {
 543                         err = PTR_ERR(skb);
 544                         goto out;
 545                 }
 546
 547                 /* This is the backpressure mechanism for the Tx path.
 548                  * Reserve space in the completion queue and only proceed
 549                  * if there is space in it. This avoids having to implement
 550                  * any buffering in the Tx path.
 551                  */
 552                 spin_lock_irqsave(&xs->pool->cq_lock, flags);
 553                 if (xskq_prod_reserve(xs->pool->cq)) {
 554                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 555                         kfree_skb(skb);
 556                         goto out;
 557                 }
 558                 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 559
 560                 err = __dev_direct_xmit(skb, xs->queue_id);
 561                 if  (err == NETDEV_TX_BUSY) {
 562                         /* Tell user-space to retry the send */
 563                         skb->destructor = sock_wfree;
 564                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 565                         xskq_prod_cancel(xs->pool->cq);
 566                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 567                         /* Free skb without triggering the perf drop trace */
 568                         consume_skb(skb);
 569                         err = -EAGAIN;
 570                         goto out;
 571                 }
 572
 573                 xskq_cons_release(xs->tx);
 574                 /* Ignore NET_XMIT_CN as packet might have been sent */
 575                 if (err == NET_XMIT_DROP) {
 576                         /* SKB completed but not sent */
 577                         err = -EBUSY;
 578                         goto out;
 579                 }
 580
 581                 sent_frame = true;
 582         }
 583
 584         xs->tx->queue_empty_descs++;
 585
 586 out:
 587         if (sent_frame)
 588                 if (xsk_tx_writeable(xs))
 589                         sk->sk_write_space(sk);
 590
 591         mutex_unlock(&xs->mutex);
 592         return err;
 593 }
 594
 595 static int xsk_xmit(struct sock *sk)
 596 {
 597         struct xdp_sock *xs = xdp_sk(sk);
 598         int ret;
 599
 600         if (unlikely(!(xs->dev->flags & IFF_UP)))
 601                 return -ENETDOWN;
 602         if (unlikely(!xs->tx))
 603                 return -ENOBUFS;
 604
 605         if (xs->zc)
 606                 return xsk_wakeup(xs, XDP_WAKEUP_TX);
 607
 608         /* Drop the RCU lock since the SKB path might sleep. */
 609         rcu_read_unlock();
 610         ret = xsk_generic_xmit(sk);
 611         /* Reaquire RCU lock before going into common code. */
 612         rcu_read_lock();
 613
 614         return ret;
 615 }
 616
 617 static bool xsk_no_wakeup(struct sock *sk)
 618 {
 619 #ifdef CONFIG_NET_RX_BUSY_POLL
 620         /* Prefer busy-polling, skip the wakeup. */
 621         return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
 622                 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
 623 #else
 624         return false;
 625 #endif
 626 }
 627
 628 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 629 {
 630         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 631         struct sock *sk = sock->sk;
 632         struct xdp_sock *xs = xdp_sk(sk);
 633         struct xsk_buff_pool *pool;
 634
 635         if (unlikely(!xsk_is_bound(xs)))
 636                 return -ENXIO;
 637         if (unlikely(need_wait))
 638                 return -EOPNOTSUPP;
 639
 640         if (sk_can_busy_loop(sk))
 641                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 642
 643         if (xs->zc && xsk_no_wakeup(sk))
 644                 return 0;
 645
 646         pool = xs->pool;
 647         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
 648                 return xsk_xmit(sk);
 649         return 0;
 650 }
 651
 652 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 653 {
 654         int ret;
 655
 656         rcu_read_lock();
 657         ret = __xsk_sendmsg(sock, m, total_len);
 658         rcu_read_unlock();
 659
 660         return ret;
 661 }
 662
 663 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 664 {
 665         bool need_wait = !(flags & MSG_DONTWAIT);
 666         struct sock *sk = sock->sk;
 667         struct xdp_sock *xs = xdp_sk(sk);
 668
 669         if (unlikely(!xsk_is_bound(xs)))
 670                 return -ENXIO;
 671         if (unlikely(!(xs->dev->flags & IFF_UP)))
 672                 return -ENETDOWN;
 673         if (unlikely(!xs->rx))
 674                 return -ENOBUFS;
 675         if (unlikely(need_wait))
 676                 return -EOPNOTSUPP;
 677
 678         if (sk_can_busy_loop(sk))
 679                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 680
 681         if (xsk_no_wakeup(sk))
 682                 return 0;
 683
 684         if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
 685                 return xsk_wakeup(xs, XDP_WAKEUP_RX);
 686         return 0;
 687 }
 688
 689 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 690 {
 691         int ret;
 692
 693         rcu_read_lock();
 694         ret = __xsk_recvmsg(sock, m, len, flags);
 695         rcu_read_unlock();
 696
 697         return ret;
 698 }
 699
 700 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 701                              struct poll_table_struct *wait)
 702 {
 703         __poll_t mask = 0;
 704         struct sock *sk = sock->sk;
 705         struct xdp_sock *xs = xdp_sk(sk);
 706         struct xsk_buff_pool *pool;
 707
 708         sock_poll_wait(file, sock, wait);
 709
 710         rcu_read_lock();
 711         if (unlikely(!xsk_is_bound(xs))) {
 712                 rcu_read_unlock();
 713                 return mask;
 714         }
 715
 716         pool = xs->pool;
 717
 718         if (pool->cached_need_wakeup) {
 719                 if (xs->zc)
 720                         xsk_wakeup(xs, pool->cached_need_wakeup);
 721                 else
 722                         /* Poll needs to drive Tx also in copy mode */
 723                         xsk_xmit(sk);
 724         }
 725
 726         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 727                 mask |= EPOLLIN | EPOLLRDNORM;
 728         if (xs->tx && xsk_tx_writeable(xs))
 729                 mask |= EPOLLOUT | EPOLLWRNORM;
 730
 731         rcu_read_unlock();
 732         return mask;
 733 }
 734
 735 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 736                           bool umem_queue)
 737 {
 738         struct xsk_queue *q;
 739
 740         if (entries == 0 || *queue || !is_power_of_2(entries))
 741                 return -EINVAL;
 742
 743         q = xskq_create(entries, umem_queue);
 744         if (!q)
 745                 return -ENOMEM;
 746
 747         /* Make sure queue is ready before it can be seen by others */
 748         smp_wmb();
 749         WRITE_ONCE(*queue, q);
 750         return 0;
 751 }
 752
 753 static void xsk_unbind_dev(struct xdp_sock *xs)
 754 {
 755         struct net_device *dev = xs->dev;
 756
 757         if (xs->state != XSK_BOUND)
 758                 return;
 759         WRITE_ONCE(xs->state, XSK_UNBOUND);
 760
 761         /* Wait for driver to stop using the xdp socket. */
 762         xp_del_xsk(xs->pool, xs);
 763         synchronize_net();
 764         dev_put(dev);
 765 }
 766
 767 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 768                                               struct xdp_sock __rcu ***map_entry)
 769 {
 770         struct xsk_map *map = NULL;
 771         struct xsk_map_node *node;
 772
 773         *map_entry = NULL;
 774
 775         spin_lock_bh(&xs->map_list_lock);
 776         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 777                                         node);
 778         if (node) {
 779                 bpf_map_inc(&node->map->map);
 780                 map = node->map;
 781                 *map_entry = node->map_entry;
 782         }
 783         spin_unlock_bh(&xs->map_list_lock);
 784         return map;
 785 }
 786
 787 static void xsk_delete_from_maps(struct xdp_sock *xs)
 788 {
 789         /* This function removes the current XDP socket from all the
 790          * maps it resides in. We need to take extra care here, due to
 791          * the two locks involved. Each map has a lock synchronizing
 792          * updates to the entries, and each socket has a lock that
 793          * synchronizes access to the list of maps (map_list). For
 794          * deadlock avoidance the locks need to be taken in the order
 795          * "map lock"->"socket map list lock". We start off by
 796          * accessing the socket map list, and take a reference to the
 797          * map to guarantee existence between the
 798          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 799          * calls. Then we ask the map to remove the socket, which
 800          * tries to remove the socket from the map. Note that there
 801          * might be updates to the map between
 802          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 803          */
 804         struct xdp_sock __rcu **map_entry = NULL;
 805         struct xsk_map *map;
 806
 807         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 808                 xsk_map_try_sock_delete(map, xs, map_entry);
 809                 bpf_map_put(&map->map);
 810         }
 811 }
 812
 813 static int xsk_release(struct socket *sock)
 814 {
 815         struct sock *sk = sock->sk;
 816         struct xdp_sock *xs = xdp_sk(sk);
 817         struct net *net;
 818
 819         if (!sk)
 820                 return 0;
 821
 822         net = sock_net(sk);
 823
 824         mutex_lock(&net->xdp.lock);
 825         sk_del_node_init_rcu(sk);
 826         mutex_unlock(&net->xdp.lock);
 827
 828         sock_prot_inuse_add(net, sk->sk_prot, -1);
 829
 830         xsk_delete_from_maps(xs);
 831         mutex_lock(&xs->mutex);
 832         xsk_unbind_dev(xs);
 833         mutex_unlock(&xs->mutex);
 834
 835         xskq_destroy(xs->rx);
 836         xskq_destroy(xs->tx);
 837         xskq_destroy(xs->fq_tmp);
 838         xskq_destroy(xs->cq_tmp);
 839
 840         sock_orphan(sk);
 841         sock->sk = NULL;
 842
 843         sk_refcnt_debug_release(sk);
 844         sock_put(sk);
 845
 846         return 0;
 847 }
 848
 849 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 850 {
 851         struct socket *sock;
 852         int err;
 853
 854         sock = sockfd_lookup(fd, &err);
 855         if (!sock)
 856                 return ERR_PTR(-ENOTSOCK);
 857
 858         if (sock->sk->sk_family != PF_XDP) {
 859                 sockfd_put(sock);
 860                 return ERR_PTR(-ENOPROTOOPT);
 861         }
 862
 863         return sock;
 864 }
 865
 866 static bool xsk_validate_queues(struct xdp_sock *xs)
 867 {
 868         return xs->fq_tmp && xs->cq_tmp;
 869 }
 870
 871 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 872 {
 873         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 874         struct sock *sk = sock->sk;
 875         struct xdp_sock *xs = xdp_sk(sk);
 876         struct net_device *dev;
 877         u32 flags, qid;
 878         int err = 0;
 879
 880         if (addr_len < sizeof(struct sockaddr_xdp))
 881                 return -EINVAL;
 882         if (sxdp->sxdp_family != AF_XDP)
 883                 return -EINVAL;
 884
 885         flags = sxdp->sxdp_flags;
 886         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 887                       XDP_USE_NEED_WAKEUP))
 888                 return -EINVAL;
 889
 890         rtnl_lock();
 891         mutex_lock(&xs->mutex);
 892         if (xs->state != XSK_READY) {
 893                 err = -EBUSY;
 894                 goto out_release;
 895         }
 896
 897         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 898         if (!dev) {
 899                 err = -ENODEV;
 900                 goto out_release;
 901         }
 902
 903         if (!xs->rx && !xs->tx) {
 904                 err = -EINVAL;
 905                 goto out_unlock;
 906         }
 907
 908         qid = sxdp->sxdp_queue_id;
 909
 910         if (flags & XDP_SHARED_UMEM) {
 911                 struct xdp_sock *umem_xs;
 912                 struct socket *sock;
 913
 914                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 915                     (flags & XDP_USE_NEED_WAKEUP)) {
 916                         /* Cannot specify flags for shared sockets. */
 917                         err = -EINVAL;
 918                         goto out_unlock;
 919                 }
 920
 921                 if (xs->umem) {
 922                         /* We have already our own. */
 923                         err = -EINVAL;
 924                         goto out_unlock;
 925                 }
 926
 927                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 928                 if (IS_ERR(sock)) {
 929                         err = PTR_ERR(sock);
 930                         goto out_unlock;
 931                 }
 932
 933                 umem_xs = xdp_sk(sock->sk);
 934                 if (!xsk_is_bound(umem_xs)) {
 935                         err = -EBADF;
 936                         sockfd_put(sock);
 937                         goto out_unlock;
 938                 }
 939
 940                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 941                         /* Share the umem with another socket on another qid
 942                          * and/or device.
 943                          */
 944                         xs->pool = xp_create_and_assign_umem(xs,
 945                                                              umem_xs->umem);
 946                         if (!xs->pool) {
 947                                 err = -ENOMEM;
 948                                 sockfd_put(sock);
 949                                 goto out_unlock;
 950                         }
 951
 952                         err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
 953                                                    dev, qid);
 954                         if (err) {
 955                                 xp_destroy(xs->pool);
 956                                 xs->pool = NULL;
 957                                 sockfd_put(sock);
 958                                 goto out_unlock;
 959                         }
 960                 } else {
 961                         /* Share the buffer pool with the other socket. */
 962                         if (xs->fq_tmp || xs->cq_tmp) {
 963                                 /* Do not allow setting your own fq or cq. */
 964                                 err = -EINVAL;
 965                                 sockfd_put(sock);
 966                                 goto out_unlock;
 967                         }
 968
 969                         xp_get_pool(umem_xs->pool);
 970                         xs->pool = umem_xs->pool;
 971
 972                         /* If underlying shared umem was created without Tx
 973                          * ring, allocate Tx descs array that Tx batching API
 974                          * utilizes
 975                          */
 976                         if (xs->tx && !xs->pool->tx_descs) {
 977                                 err = xp_alloc_tx_descs(xs->pool, xs);
 978                                 if (err) {
 979                                         xp_put_pool(xs->pool);
 980                                         sockfd_put(sock);
 981                                         goto out_unlock;
 982                                 }
 983                         }
 984                 }
 985
 986                 xdp_get_umem(umem_xs->umem);
 987                 WRITE_ONCE(xs->umem, umem_xs->umem);
 988                 sockfd_put(sock);
 989         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 990                 err = -EINVAL;
 991                 goto out_unlock;
 992         } else {
 993                 /* This xsk has its own umem. */
 994                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 995                 if (!xs->pool) {
 996                         err = -ENOMEM;
 997                         goto out_unlock;
 998                 }
 999
1000                 err = xp_assign_dev(xs->pool, dev, qid, flags);
1001                 if (err) {
1002                         xp_destroy(xs->pool);
1003                         xs->pool = NULL;
1004                         goto out_unlock;
1005                 }
1006         }
1007
1008         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1009         xs->fq_tmp = NULL;
1010         xs->cq_tmp = NULL;
1011
1012         xs->dev = dev;
1013         xs->zc = xs->umem->zc;
1014         xs->queue_id = qid;
1015         xp_add_xsk(xs->pool, xs);
1016
1017 out_unlock:
1018         if (err) {
1019                 dev_put(dev);
1020         } else {
1021                 /* Matches smp_rmb() in bind() for shared umem
1022                  * sockets, and xsk_is_bound().
1023                  */
1024                 smp_wmb();
1025                 WRITE_ONCE(xs->state, XSK_BOUND);
1026         }
1027 out_release:
1028         mutex_unlock(&xs->mutex);
1029         rtnl_unlock();
1030         return err;
1031 }
1032
1033 struct xdp_umem_reg_v1 {
1034         __u64 addr; /* Start of packet data area */
1035         __u64 len; /* Length of packet data area */
1036         __u32 chunk_size;
1037         __u32 headroom;
1038 };
1039
1040 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1041                           sockptr_t optval, unsigned int optlen)
1042 {
1043         struct sock *sk = sock->sk;
1044         struct xdp_sock *xs = xdp_sk(sk);
1045         int err;
1046
1047         if (level != SOL_XDP)
1048                 return -ENOPROTOOPT;
1049
1050         switch (optname) {
1051         case XDP_RX_RING:
1052         case XDP_TX_RING:
1053         {
1054                 struct xsk_queue **q;
1055                 int entries;
1056
1057                 if (optlen < sizeof(entries))
1058                         return -EINVAL;
1059                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1060                         return -EFAULT;
1061
1062                 mutex_lock(&xs->mutex);
1063                 if (xs->state != XSK_READY) {
1064                         mutex_unlock(&xs->mutex);
1065                         return -EBUSY;
1066                 }
1067                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1068                 err = xsk_init_queue(entries, q, false);
1069                 if (!err && optname == XDP_TX_RING)
1070                         /* Tx needs to be explicitly woken up the first time */
1071                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1072                 mutex_unlock(&xs->mutex);
1073                 return err;
1074         }
1075         case XDP_UMEM_REG:
1076         {
1077                 size_t mr_size = sizeof(struct xdp_umem_reg);
1078                 struct xdp_umem_reg mr = {};
1079                 struct xdp_umem *umem;
1080
1081                 if (optlen < sizeof(struct xdp_umem_reg_v1))
1082                         return -EINVAL;
1083                 else if (optlen < sizeof(mr))
1084                         mr_size = sizeof(struct xdp_umem_reg_v1);
1085
1086                 if (copy_from_sockptr(&mr, optval, mr_size))
1087                         return -EFAULT;
1088
1089                 mutex_lock(&xs->mutex);
1090                 if (xs->state != XSK_READY || xs->umem) {
1091                         mutex_unlock(&xs->mutex);
1092                         return -EBUSY;
1093                 }
1094
1095                 umem = xdp_umem_create(&mr);
1096                 if (IS_ERR(umem)) {
1097                         mutex_unlock(&xs->mutex);
1098                         return PTR_ERR(umem);
1099                 }
1100
1101                 /* Make sure umem is ready before it can be seen by others */
1102                 smp_wmb();
1103                 WRITE_ONCE(xs->umem, umem);
1104                 mutex_unlock(&xs->mutex);
1105                 return 0;
1106         }
1107         case XDP_UMEM_FILL_RING:
1108         case XDP_UMEM_COMPLETION_RING:
1109         {
1110                 struct xsk_queue **q;
1111                 int entries;
1112
1113                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1114                         return -EFAULT;
1115
1116                 mutex_lock(&xs->mutex);
1117                 if (xs->state != XSK_READY) {
1118                         mutex_unlock(&xs->mutex);
1119                         return -EBUSY;
1120                 }
1121
1122                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1123                         &xs->cq_tmp;
1124                 err = xsk_init_queue(entries, q, true);
1125                 mutex_unlock(&xs->mutex);
1126                 return err;
1127         }
1128         default:
1129                 break;
1130         }
1131
1132         return -ENOPROTOOPT;
1133 }
1134
1135 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1136 {
1137         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1138         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1139         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1140 }
1141
1142 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1143 {
1144         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1145         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1146         ring->desc = offsetof(struct xdp_umem_ring, desc);
1147 }
1148
1149 struct xdp_statistics_v1 {
1150         __u64 rx_dropped;
1151         __u64 rx_invalid_descs;
1152         __u64 tx_invalid_descs;
1153 };
1154
1155 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1156                           char __user *optval, int __user *optlen)
1157 {
1158         struct sock *sk = sock->sk;
1159         struct xdp_sock *xs = xdp_sk(sk);
1160         int len;
1161
1162         if (level != SOL_XDP)
1163                 return -ENOPROTOOPT;
1164
1165         if (get_user(len, optlen))
1166                 return -EFAULT;
1167         if (len < 0)
1168                 return -EINVAL;
1169
1170         switch (optname) {
1171         case XDP_STATISTICS:
1172         {
1173                 struct xdp_statistics stats = {};
1174                 bool extra_stats = true;
1175                 size_t stats_size;
1176
1177                 if (len < sizeof(struct xdp_statistics_v1)) {
1178                         return -EINVAL;
1179                 } else if (len < sizeof(stats)) {
1180                         extra_stats = false;
1181                         stats_size = sizeof(struct xdp_statistics_v1);
1182                 } else {
1183                         stats_size = sizeof(stats);
1184                 }
1185
1186                 mutex_lock(&xs->mutex);
1187                 stats.rx_dropped = xs->rx_dropped;
1188                 if (extra_stats) {
1189                         stats.rx_ring_full = xs->rx_queue_full;
1190                         stats.rx_fill_ring_empty_descs =
1191                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1192                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1193                 } else {
1194                         stats.rx_dropped += xs->rx_queue_full;
1195                 }
1196                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1197                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1198                 mutex_unlock(&xs->mutex);
1199
1200                 if (copy_to_user(optval, &stats, stats_size))
1201                         return -EFAULT;
1202                 if (put_user(stats_size, optlen))
1203                         return -EFAULT;
1204
1205                 return 0;
1206         }
1207         case XDP_MMAP_OFFSETS:
1208         {
1209                 struct xdp_mmap_offsets off;
1210                 struct xdp_mmap_offsets_v1 off_v1;
1211                 bool flags_supported = true;
1212                 void *to_copy;
1213
1214                 if (len < sizeof(off_v1))
1215                         return -EINVAL;
1216                 else if (len < sizeof(off))
1217                         flags_supported = false;
1218
1219                 if (flags_supported) {
1220                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1221                          * except for the flags field added to the end.
1222                          */
1223                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1224                                                &off.rx);
1225                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1226                                                &off.tx);
1227                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1228                                                &off.fr);
1229                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1230                                                &off.cr);
1231                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1232                                                 ptrs.flags);
1233                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1234                                                 ptrs.flags);
1235                         off.fr.flags = offsetof(struct xdp_umem_ring,
1236                                                 ptrs.flags);
1237                         off.cr.flags = offsetof(struct xdp_umem_ring,
1238                                                 ptrs.flags);
1239
1240                         len = sizeof(off);
1241                         to_copy = &off;
1242                 } else {
1243                         xsk_enter_rxtx_offsets(&off_v1.rx);
1244                         xsk_enter_rxtx_offsets(&off_v1.tx);
1245                         xsk_enter_umem_offsets(&off_v1.fr);
1246                         xsk_enter_umem_offsets(&off_v1.cr);
1247
1248                         len = sizeof(off_v1);
1249                         to_copy = &off_v1;
1250                 }
1251
1252                 if (copy_to_user(optval, to_copy, len))
1253                         return -EFAULT;
1254                 if (put_user(len, optlen))
1255                         return -EFAULT;
1256
1257                 return 0;
1258         }
1259         case XDP_OPTIONS:
1260         {
1261                 struct xdp_options opts = {};
1262
1263                 if (len < sizeof(opts))
1264                         return -EINVAL;
1265
1266                 mutex_lock(&xs->mutex);
1267                 if (xs->zc)
1268                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1269                 mutex_unlock(&xs->mutex);
1270
1271                 len = sizeof(opts);
1272                 if (copy_to_user(optval, &opts, len))
1273                         return -EFAULT;
1274                 if (put_user(len, optlen))
1275                         return -EFAULT;
1276
1277                 return 0;
1278         }
1279         default:
1280                 break;
1281         }
1282
1283         return -EOPNOTSUPP;
1284 }
1285
1286 static int xsk_mmap(struct file *file, struct socket *sock,
1287                     struct vm_area_struct *vma)
1288 {
1289         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1290         unsigned long size = vma->vm_end - vma->vm_start;
1291         struct xdp_sock *xs = xdp_sk(sock->sk);
1292         struct xsk_queue *q = NULL;
1293         unsigned long pfn;
1294         struct page *qpg;
1295
1296         if (READ_ONCE(xs->state) != XSK_READY)
1297                 return -EBUSY;
1298
1299         if (offset == XDP_PGOFF_RX_RING) {
1300                 q = READ_ONCE(xs->rx);
1301         } else if (offset == XDP_PGOFF_TX_RING) {
1302                 q = READ_ONCE(xs->tx);
1303         } else {
1304                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1305                 smp_rmb();
1306                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1307                         q = READ_ONCE(xs->fq_tmp);
1308                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1309                         q = READ_ONCE(xs->cq_tmp);
1310         }
1311
1312         if (!q)
1313                 return -EINVAL;
1314
1315         /* Matches the smp_wmb() in xsk_init_queue */
1316         smp_rmb();
1317         qpg = virt_to_head_page(q->ring);
1318         if (size > page_size(qpg))
1319                 return -EINVAL;
1320
1321         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1322         return remap_pfn_range(vma, vma->vm_start, pfn,
1323                                size, vma->vm_page_prot);
1324 }
1325
1326 static int xsk_notifier(struct notifier_block *this,
1327                         unsigned long msg, void *ptr)
1328 {
1329         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1330         struct net *net = dev_net(dev);
1331         struct sock *sk;
1332
1333         switch (msg) {
1334         case NETDEV_UNREGISTER:
1335                 mutex_lock(&net->xdp.lock);
1336                 sk_for_each(sk, &net->xdp.list) {
1337                         struct xdp_sock *xs = xdp_sk(sk);
1338
1339                         mutex_lock(&xs->mutex);
1340                         if (xs->dev == dev) {
1341                                 sk->sk_err = ENETDOWN;
1342                                 if (!sock_flag(sk, SOCK_DEAD))
1343                                         sk_error_report(sk);
1344
1345                                 xsk_unbind_dev(xs);
1346
1347                                 /* Clear device references. */
1348                                 xp_clear_dev(xs->pool);
1349                         }
1350                         mutex_unlock(&xs->mutex);
1351                 }
1352                 mutex_unlock(&net->xdp.lock);
1353                 break;
1354         }
1355         return NOTIFY_DONE;
1356 }
1357
1358 static struct proto xsk_proto = {
1359         .name =         "XDP",
1360         .owner =        THIS_MODULE,
1361         .obj_size =     sizeof(struct xdp_sock),
1362 };
1363
1364 static const struct proto_ops xsk_proto_ops = {
1365         .family         = PF_XDP,
1366         .owner          = THIS_MODULE,
1367         .release        = xsk_release,
1368         .bind           = xsk_bind,
1369         .connect        = sock_no_connect,
1370         .socketpair     = sock_no_socketpair,
1371         .accept         = sock_no_accept,
1372         .getname        = sock_no_getname,
1373         .poll           = xsk_poll,
1374         .ioctl          = sock_no_ioctl,
1375         .listen         = sock_no_listen,
1376         .shutdown       = sock_no_shutdown,
1377         .setsockopt     = xsk_setsockopt,
1378         .getsockopt     = xsk_getsockopt,
1379         .sendmsg        = xsk_sendmsg,
1380         .recvmsg        = xsk_recvmsg,
1381         .mmap           = xsk_mmap,
1382         .sendpage       = sock_no_sendpage,
1383 };
1384
1385 static void xsk_destruct(struct sock *sk)
1386 {
1387         struct xdp_sock *xs = xdp_sk(sk);
1388
1389         if (!sock_flag(sk, SOCK_DEAD))
1390                 return;
1391
1392         if (!xp_put_pool(xs->pool))
1393                 xdp_put_umem(xs->umem, !xs->pool);
1394
1395         sk_refcnt_debug_dec(sk);
1396 }
1397
1398 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1399                       int kern)
1400 {
1401         struct xdp_sock *xs;
1402         struct sock *sk;
1403
1404         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1405                 return -EPERM;
1406         if (sock->type != SOCK_RAW)
1407                 return -ESOCKTNOSUPPORT;
1408
1409         if (protocol)
1410                 return -EPROTONOSUPPORT;
1411
1412         sock->state = SS_UNCONNECTED;
1413
1414         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1415         if (!sk)
1416                 return -ENOBUFS;
1417
1418         sock->ops = &xsk_proto_ops;
1419
1420         sock_init_data(sock, sk);
1421
1422         sk->sk_family = PF_XDP;
1423
1424         sk->sk_destruct = xsk_destruct;
1425         sk_refcnt_debug_inc(sk);
1426
1427         sock_set_flag(sk, SOCK_RCU_FREE);
1428
1429         xs = xdp_sk(sk);
1430         xs->state = XSK_READY;
1431         mutex_init(&xs->mutex);
1432         spin_lock_init(&xs->rx_lock);
1433
1434         INIT_LIST_HEAD(&xs->map_list);
1435         spin_lock_init(&xs->map_list_lock);
1436
1437         mutex_lock(&net->xdp.lock);
1438         sk_add_node_rcu(sk, &net->xdp.list);
1439         mutex_unlock(&net->xdp.lock);
1440
1441         sock_prot_inuse_add(net, &xsk_proto, 1);
1442
1443         return 0;
1444 }
1445
1446 static const struct net_proto_family xsk_family_ops = {
1447         .family = PF_XDP,
1448         .create = xsk_create,
1449         .owner  = THIS_MODULE,
1450 };
1451
1452 static struct notifier_block xsk_netdev_notifier = {
1453         .notifier_call  = xsk_notifier,
1454 };
1455
1456 static int __net_init xsk_net_init(struct net *net)
1457 {
1458         mutex_init(&net->xdp.lock);
1459         INIT_HLIST_HEAD(&net->xdp.list);
1460         return 0;
1461 }
1462
1463 static void __net_exit xsk_net_exit(struct net *net)
1464 {
1465         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1466 }
1467
1468 static struct pernet_operations xsk_net_ops = {
1469         .init = xsk_net_init,
1470         .exit = xsk_net_exit,
1471 };
1472
1473 static int __init xsk_init(void)
1474 {
1475         int err, cpu;
1476
1477         err = proto_register(&xsk_proto, 0 /* no slab */);
1478         if (err)
1479                 goto out;
1480
1481         err = sock_register(&xsk_family_ops);
1482         if (err)
1483                 goto out_proto;
1484
1485         err = register_pernet_subsys(&xsk_net_ops);
1486         if (err)
1487                 goto out_sk;
1488
1489         err = register_netdevice_notifier(&xsk_netdev_notifier);
1490         if (err)
1491                 goto out_pernet;
1492
1493         for_each_possible_cpu(cpu)
1494                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1495         return 0;
1496
1497 out_pernet:
1498         unregister_pernet_subsys(&xsk_net_ops);
1499 out_sk:
1500         sock_unregister(PF_XDP);
1501 out_proto:
1502         proto_unregister(&xsk_proto);
1503 out:
1504         return err;
1505 }
1506
1507 fs_initcall(xsk_init);