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         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                 /* This is the backpressure mechanism for the Tx path.
 542                  * Reserve space in the completion queue and only proceed
 543                  * if there is space in it. This avoids having to implement
 544                  * any buffering in the Tx path.
 545                  */
 546                 spin_lock_irqsave(&xs->pool->cq_lock, flags);
 547                 if (xskq_prod_reserve(xs->pool->cq)) {
 548                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 549                         goto out;
 550                 }
 551                 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 552
 553                 skb = xsk_build_skb(xs, &desc);
 554                 if (IS_ERR(skb)) {
 555                         err = PTR_ERR(skb);
 556                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 557                         xskq_prod_cancel(xs->pool->cq);
 558                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 559                         goto out;
 560                 }
 561
 562                 err = __dev_direct_xmit(skb, xs->queue_id);
 563                 if  (err == NETDEV_TX_BUSY) {
 564                         /* Tell user-space to retry the send */
 565                         skb->destructor = sock_wfree;
 566                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 567                         xskq_prod_cancel(xs->pool->cq);
 568                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 569                         /* Free skb without triggering the perf drop trace */
 570                         consume_skb(skb);
 571                         err = -EAGAIN;
 572                         goto out;
 573                 }
 574
 575                 xskq_cons_release(xs->tx);
 576                 /* Ignore NET_XMIT_CN as packet might have been sent */
 577                 if (err == NET_XMIT_DROP) {
 578                         /* SKB completed but not sent */
 579                         err = -EBUSY;
 580                         goto out;
 581                 }
 582
 583                 sent_frame = true;
 584         }
 585
 586         xs->tx->queue_empty_descs++;
 587
 588 out:
 589         if (sent_frame)
 590                 if (xsk_tx_writeable(xs))
 591                         sk->sk_write_space(sk);
 592
 593         mutex_unlock(&xs->mutex);
 594         return err;
 595 }
 596
 597 static int xsk_xmit(struct sock *sk)
 598 {
 599         struct xdp_sock *xs = xdp_sk(sk);
 600         int ret;
 601
 602         if (unlikely(!(xs->dev->flags & IFF_UP)))
 603                 return -ENETDOWN;
 604         if (unlikely(!xs->tx))
 605                 return -ENOBUFS;
 606
 607         if (xs->zc)
 608                 return xsk_wakeup(xs, XDP_WAKEUP_TX);
 609
 610         /* Drop the RCU lock since the SKB path might sleep. */
 611         rcu_read_unlock();
 612         ret = xsk_generic_xmit(sk);
 613         /* Reaquire RCU lock before going into common code. */
 614         rcu_read_lock();
 615
 616         return ret;
 617 }
 618
 619 static bool xsk_no_wakeup(struct sock *sk)
 620 {
 621 #ifdef CONFIG_NET_RX_BUSY_POLL
 622         /* Prefer busy-polling, skip the wakeup. */
 623         return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
 624                 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
 625 #else
 626         return false;
 627 #endif
 628 }
 629
 630 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 631 {
 632         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 633         struct sock *sk = sock->sk;
 634         struct xdp_sock *xs = xdp_sk(sk);
 635         struct xsk_buff_pool *pool;
 636
 637         if (unlikely(!xsk_is_bound(xs)))
 638                 return -ENXIO;
 639         if (unlikely(need_wait))
 640                 return -EOPNOTSUPP;
 641
 642         if (sk_can_busy_loop(sk))
 643                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 644
 645         if (xs->zc && xsk_no_wakeup(sk))
 646                 return 0;
 647
 648         pool = xs->pool;
 649         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
 650                 return xsk_xmit(sk);
 651         return 0;
 652 }
 653
 654 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 655 {
 656         int ret;
 657
 658         rcu_read_lock();
 659         ret = __xsk_sendmsg(sock, m, total_len);
 660         rcu_read_unlock();
 661
 662         return ret;
 663 }
 664
 665 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 666 {
 667         bool need_wait = !(flags & MSG_DONTWAIT);
 668         struct sock *sk = sock->sk;
 669         struct xdp_sock *xs = xdp_sk(sk);
 670
 671         if (unlikely(!xsk_is_bound(xs)))
 672                 return -ENXIO;
 673         if (unlikely(!(xs->dev->flags & IFF_UP)))
 674                 return -ENETDOWN;
 675         if (unlikely(!xs->rx))
 676                 return -ENOBUFS;
 677         if (unlikely(need_wait))
 678                 return -EOPNOTSUPP;
 679
 680         if (sk_can_busy_loop(sk))
 681                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 682
 683         if (xsk_no_wakeup(sk))
 684                 return 0;
 685
 686         if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
 687                 return xsk_wakeup(xs, XDP_WAKEUP_RX);
 688         return 0;
 689 }
 690
 691 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 692 {
 693         int ret;
 694
 695         rcu_read_lock();
 696         ret = __xsk_recvmsg(sock, m, len, flags);
 697         rcu_read_unlock();
 698
 699         return ret;
 700 }
 701
 702 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 703                              struct poll_table_struct *wait)
 704 {
 705         __poll_t mask = 0;
 706         struct sock *sk = sock->sk;
 707         struct xdp_sock *xs = xdp_sk(sk);
 708         struct xsk_buff_pool *pool;
 709
 710         sock_poll_wait(file, sock, wait);
 711
 712         rcu_read_lock();
 713         if (unlikely(!xsk_is_bound(xs))) {
 714                 rcu_read_unlock();
 715                 return mask;
 716         }
 717
 718         pool = xs->pool;
 719
 720         if (pool->cached_need_wakeup) {
 721                 if (xs->zc)
 722                         xsk_wakeup(xs, pool->cached_need_wakeup);
 723                 else
 724                         /* Poll needs to drive Tx also in copy mode */
 725                         xsk_xmit(sk);
 726         }
 727
 728         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 729                 mask |= EPOLLIN | EPOLLRDNORM;
 730         if (xs->tx && xsk_tx_writeable(xs))
 731                 mask |= EPOLLOUT | EPOLLWRNORM;
 732
 733         rcu_read_unlock();
 734         return mask;
 735 }
 736
 737 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 738                           bool umem_queue)
 739 {
 740         struct xsk_queue *q;
 741
 742         if (entries == 0 || *queue || !is_power_of_2(entries))
 743                 return -EINVAL;
 744
 745         q = xskq_create(entries, umem_queue);
 746         if (!q)
 747                 return -ENOMEM;
 748
 749         /* Make sure queue is ready before it can be seen by others */
 750         smp_wmb();
 751         WRITE_ONCE(*queue, q);
 752         return 0;
 753 }
 754
 755 static void xsk_unbind_dev(struct xdp_sock *xs)
 756 {
 757         struct net_device *dev = xs->dev;
 758
 759         if (xs->state != XSK_BOUND)
 760                 return;
 761         WRITE_ONCE(xs->state, XSK_UNBOUND);
 762
 763         /* Wait for driver to stop using the xdp socket. */
 764         xp_del_xsk(xs->pool, xs);
 765         synchronize_net();
 766         dev_put(dev);
 767 }
 768
 769 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 770                                               struct xdp_sock __rcu ***map_entry)
 771 {
 772         struct xsk_map *map = NULL;
 773         struct xsk_map_node *node;
 774
 775         *map_entry = NULL;
 776
 777         spin_lock_bh(&xs->map_list_lock);
 778         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 779                                         node);
 780         if (node) {
 781                 bpf_map_inc(&node->map->map);
 782                 map = node->map;
 783                 *map_entry = node->map_entry;
 784         }
 785         spin_unlock_bh(&xs->map_list_lock);
 786         return map;
 787 }
 788
 789 static void xsk_delete_from_maps(struct xdp_sock *xs)
 790 {
 791         /* This function removes the current XDP socket from all the
 792          * maps it resides in. We need to take extra care here, due to
 793          * the two locks involved. Each map has a lock synchronizing
 794          * updates to the entries, and each socket has a lock that
 795          * synchronizes access to the list of maps (map_list). For
 796          * deadlock avoidance the locks need to be taken in the order
 797          * "map lock"->"socket map list lock". We start off by
 798          * accessing the socket map list, and take a reference to the
 799          * map to guarantee existence between the
 800          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 801          * calls. Then we ask the map to remove the socket, which
 802          * tries to remove the socket from the map. Note that there
 803          * might be updates to the map between
 804          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 805          */
 806         struct xdp_sock __rcu **map_entry = NULL;
 807         struct xsk_map *map;
 808
 809         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 810                 xsk_map_try_sock_delete(map, xs, map_entry);
 811                 bpf_map_put(&map->map);
 812         }
 813 }
 814
 815 static int xsk_release(struct socket *sock)
 816 {
 817         struct sock *sk = sock->sk;
 818         struct xdp_sock *xs = xdp_sk(sk);
 819         struct net *net;
 820
 821         if (!sk)
 822                 return 0;
 823
 824         net = sock_net(sk);
 825
 826         mutex_lock(&net->xdp.lock);
 827         sk_del_node_init_rcu(sk);
 828         mutex_unlock(&net->xdp.lock);
 829
 830         sock_prot_inuse_add(net, sk->sk_prot, -1);
 831
 832         xsk_delete_from_maps(xs);
 833         mutex_lock(&xs->mutex);
 834         xsk_unbind_dev(xs);
 835         mutex_unlock(&xs->mutex);
 836
 837         xskq_destroy(xs->rx);
 838         xskq_destroy(xs->tx);
 839         xskq_destroy(xs->fq_tmp);
 840         xskq_destroy(xs->cq_tmp);
 841
 842         sock_orphan(sk);
 843         sock->sk = NULL;
 844
 845         sk_refcnt_debug_release(sk);
 846         sock_put(sk);
 847
 848         return 0;
 849 }
 850
 851 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 852 {
 853         struct socket *sock;
 854         int err;
 855
 856         sock = sockfd_lookup(fd, &err);
 857         if (!sock)
 858                 return ERR_PTR(-ENOTSOCK);
 859
 860         if (sock->sk->sk_family != PF_XDP) {
 861                 sockfd_put(sock);
 862                 return ERR_PTR(-ENOPROTOOPT);
 863         }
 864
 865         return sock;
 866 }
 867
 868 static bool xsk_validate_queues(struct xdp_sock *xs)
 869 {
 870         return xs->fq_tmp && xs->cq_tmp;
 871 }
 872
 873 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 874 {
 875         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 876         struct sock *sk = sock->sk;
 877         struct xdp_sock *xs = xdp_sk(sk);
 878         struct net_device *dev;
 879         u32 flags, qid;
 880         int err = 0;
 881
 882         if (addr_len < sizeof(struct sockaddr_xdp))
 883                 return -EINVAL;
 884         if (sxdp->sxdp_family != AF_XDP)
 885                 return -EINVAL;
 886
 887         flags = sxdp->sxdp_flags;
 888         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 889                       XDP_USE_NEED_WAKEUP))
 890                 return -EINVAL;
 891
 892         rtnl_lock();
 893         mutex_lock(&xs->mutex);
 894         if (xs->state != XSK_READY) {
 895                 err = -EBUSY;
 896                 goto out_release;
 897         }
 898
 899         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 900         if (!dev) {
 901                 err = -ENODEV;
 902                 goto out_release;
 903         }
 904
 905         if (!xs->rx && !xs->tx) {
 906                 err = -EINVAL;
 907                 goto out_unlock;
 908         }
 909
 910         qid = sxdp->sxdp_queue_id;
 911
 912         if (flags & XDP_SHARED_UMEM) {
 913                 struct xdp_sock *umem_xs;
 914                 struct socket *sock;
 915
 916                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 917                     (flags & XDP_USE_NEED_WAKEUP)) {
 918                         /* Cannot specify flags for shared sockets. */
 919                         err = -EINVAL;
 920                         goto out_unlock;
 921                 }
 922
 923                 if (xs->umem) {
 924                         /* We have already our own. */
 925                         err = -EINVAL;
 926                         goto out_unlock;
 927                 }
 928
 929                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 930                 if (IS_ERR(sock)) {
 931                         err = PTR_ERR(sock);
 932                         goto out_unlock;
 933                 }
 934
 935                 umem_xs = xdp_sk(sock->sk);
 936                 if (!xsk_is_bound(umem_xs)) {
 937                         err = -EBADF;
 938                         sockfd_put(sock);
 939                         goto out_unlock;
 940                 }
 941
 942                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 943                         /* Share the umem with another socket on another qid
 944                          * and/or device.
 945                          */
 946                         xs->pool = xp_create_and_assign_umem(xs,
 947                                                              umem_xs->umem);
 948                         if (!xs->pool) {
 949                                 err = -ENOMEM;
 950                                 sockfd_put(sock);
 951                                 goto out_unlock;
 952                         }
 953
 954                         err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
 955                                                    dev, qid);
 956                         if (err) {
 957                                 xp_destroy(xs->pool);
 958                                 xs->pool = NULL;
 959                                 sockfd_put(sock);
 960                                 goto out_unlock;
 961                         }
 962                 } else {
 963                         /* Share the buffer pool with the other socket. */
 964                         if (xs->fq_tmp || xs->cq_tmp) {
 965                                 /* Do not allow setting your own fq or cq. */
 966                                 err = -EINVAL;
 967                                 sockfd_put(sock);
 968                                 goto out_unlock;
 969                         }
 970
 971                         xp_get_pool(umem_xs->pool);
 972                         xs->pool = umem_xs->pool;
 973
 974                         /* If underlying shared umem was created without Tx
 975                          * ring, allocate Tx descs array that Tx batching API
 976                          * utilizes
 977                          */
 978                         if (xs->tx && !xs->pool->tx_descs) {
 979                                 err = xp_alloc_tx_descs(xs->pool, xs);
 980                                 if (err) {
 981                                         xp_put_pool(xs->pool);
 982                                         sockfd_put(sock);
 983                                         goto out_unlock;
 984                                 }
 985                         }
 986                 }
 987
 988                 xdp_get_umem(umem_xs->umem);
 989                 WRITE_ONCE(xs->umem, umem_xs->umem);
 990                 sockfd_put(sock);
 991         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 992                 err = -EINVAL;
 993                 goto out_unlock;
 994         } else {
 995                 /* This xsk has its own umem. */
 996                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 997                 if (!xs->pool) {
 998                         err = -ENOMEM;
 999                         goto out_unlock;
1000                 }
1001
1002                 err = xp_assign_dev(xs->pool, dev, qid, flags);
1003                 if (err) {
1004                         xp_destroy(xs->pool);
1005                         xs->pool = NULL;
1006                         goto out_unlock;
1007                 }
1008         }
1009
1010         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1011         xs->fq_tmp = NULL;
1012         xs->cq_tmp = NULL;
1013
1014         xs->dev = dev;
1015         xs->zc = xs->umem->zc;
1016         xs->queue_id = qid;
1017         xp_add_xsk(xs->pool, xs);
1018
1019 out_unlock:
1020         if (err) {
1021                 dev_put(dev);
1022         } else {
1023                 /* Matches smp_rmb() in bind() for shared umem
1024                  * sockets, and xsk_is_bound().
1025                  */
1026                 smp_wmb();
1027                 WRITE_ONCE(xs->state, XSK_BOUND);
1028         }
1029 out_release:
1030         mutex_unlock(&xs->mutex);
1031         rtnl_unlock();
1032         return err;
1033 }
1034
1035 struct xdp_umem_reg_v1 {
1036         __u64 addr; /* Start of packet data area */
1037         __u64 len; /* Length of packet data area */
1038         __u32 chunk_size;
1039         __u32 headroom;
1040 };
1041
1042 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1043                           sockptr_t optval, unsigned int optlen)
1044 {
1045         struct sock *sk = sock->sk;
1046         struct xdp_sock *xs = xdp_sk(sk);
1047         int err;
1048
1049         if (level != SOL_XDP)
1050                 return -ENOPROTOOPT;
1051
1052         switch (optname) {
1053         case XDP_RX_RING:
1054         case XDP_TX_RING:
1055         {
1056                 struct xsk_queue **q;
1057                 int entries;
1058
1059                 if (optlen < sizeof(entries))
1060                         return -EINVAL;
1061                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1062                         return -EFAULT;
1063
1064                 mutex_lock(&xs->mutex);
1065                 if (xs->state != XSK_READY) {
1066                         mutex_unlock(&xs->mutex);
1067                         return -EBUSY;
1068                 }
1069                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1070                 err = xsk_init_queue(entries, q, false);
1071                 if (!err && optname == XDP_TX_RING)
1072                         /* Tx needs to be explicitly woken up the first time */
1073                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1074                 mutex_unlock(&xs->mutex);
1075                 return err;
1076         }
1077         case XDP_UMEM_REG:
1078         {
1079                 size_t mr_size = sizeof(struct xdp_umem_reg);
1080                 struct xdp_umem_reg mr = {};
1081                 struct xdp_umem *umem;
1082
1083                 if (optlen < sizeof(struct xdp_umem_reg_v1))
1084                         return -EINVAL;
1085                 else if (optlen < sizeof(mr))
1086                         mr_size = sizeof(struct xdp_umem_reg_v1);
1087
1088                 if (copy_from_sockptr(&mr, optval, mr_size))
1089                         return -EFAULT;
1090
1091                 mutex_lock(&xs->mutex);
1092                 if (xs->state != XSK_READY || xs->umem) {
1093                         mutex_unlock(&xs->mutex);
1094                         return -EBUSY;
1095                 }
1096
1097                 umem = xdp_umem_create(&mr);
1098                 if (IS_ERR(umem)) {
1099                         mutex_unlock(&xs->mutex);
1100                         return PTR_ERR(umem);
1101                 }
1102
1103                 /* Make sure umem is ready before it can be seen by others */
1104                 smp_wmb();
1105                 WRITE_ONCE(xs->umem, umem);
1106                 mutex_unlock(&xs->mutex);
1107                 return 0;
1108         }
1109         case XDP_UMEM_FILL_RING:
1110         case XDP_UMEM_COMPLETION_RING:
1111         {
1112                 struct xsk_queue **q;
1113                 int entries;
1114
1115                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1116                         return -EFAULT;
1117
1118                 mutex_lock(&xs->mutex);
1119                 if (xs->state != XSK_READY) {
1120                         mutex_unlock(&xs->mutex);
1121                         return -EBUSY;
1122                 }
1123
1124                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1125                         &xs->cq_tmp;
1126                 err = xsk_init_queue(entries, q, true);
1127                 mutex_unlock(&xs->mutex);
1128                 return err;
1129         }
1130         default:
1131                 break;
1132         }
1133
1134         return -ENOPROTOOPT;
1135 }
1136
1137 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1138 {
1139         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1140         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1141         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1142 }
1143
1144 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1145 {
1146         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1147         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1148         ring->desc = offsetof(struct xdp_umem_ring, desc);
1149 }
1150
1151 struct xdp_statistics_v1 {
1152         __u64 rx_dropped;
1153         __u64 rx_invalid_descs;
1154         __u64 tx_invalid_descs;
1155 };
1156
1157 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1158                           char __user *optval, int __user *optlen)
1159 {
1160         struct sock *sk = sock->sk;
1161         struct xdp_sock *xs = xdp_sk(sk);
1162         int len;
1163
1164         if (level != SOL_XDP)
1165                 return -ENOPROTOOPT;
1166
1167         if (get_user(len, optlen))
1168                 return -EFAULT;
1169         if (len < 0)
1170                 return -EINVAL;
1171
1172         switch (optname) {
1173         case XDP_STATISTICS:
1174         {
1175                 struct xdp_statistics stats = {};
1176                 bool extra_stats = true;
1177                 size_t stats_size;
1178
1179                 if (len < sizeof(struct xdp_statistics_v1)) {
1180                         return -EINVAL;
1181                 } else if (len < sizeof(stats)) {
1182                         extra_stats = false;
1183                         stats_size = sizeof(struct xdp_statistics_v1);
1184                 } else {
1185                         stats_size = sizeof(stats);
1186                 }
1187
1188                 mutex_lock(&xs->mutex);
1189                 stats.rx_dropped = xs->rx_dropped;
1190                 if (extra_stats) {
1191                         stats.rx_ring_full = xs->rx_queue_full;
1192                         stats.rx_fill_ring_empty_descs =
1193                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1194                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1195                 } else {
1196                         stats.rx_dropped += xs->rx_queue_full;
1197                 }
1198                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1199                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1200                 mutex_unlock(&xs->mutex);
1201
1202                 if (copy_to_user(optval, &stats, stats_size))
1203                         return -EFAULT;
1204                 if (put_user(stats_size, optlen))
1205                         return -EFAULT;
1206
1207                 return 0;
1208         }
1209         case XDP_MMAP_OFFSETS:
1210         {
1211                 struct xdp_mmap_offsets off;
1212                 struct xdp_mmap_offsets_v1 off_v1;
1213                 bool flags_supported = true;
1214                 void *to_copy;
1215
1216                 if (len < sizeof(off_v1))
1217                         return -EINVAL;
1218                 else if (len < sizeof(off))
1219                         flags_supported = false;
1220
1221                 if (flags_supported) {
1222                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1223                          * except for the flags field added to the end.
1224                          */
1225                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1226                                                &off.rx);
1227                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1228                                                &off.tx);
1229                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1230                                                &off.fr);
1231                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1232                                                &off.cr);
1233                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1234                                                 ptrs.flags);
1235                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1236                                                 ptrs.flags);
1237                         off.fr.flags = offsetof(struct xdp_umem_ring,
1238                                                 ptrs.flags);
1239                         off.cr.flags = offsetof(struct xdp_umem_ring,
1240                                                 ptrs.flags);
1241
1242                         len = sizeof(off);
1243                         to_copy = &off;
1244                 } else {
1245                         xsk_enter_rxtx_offsets(&off_v1.rx);
1246                         xsk_enter_rxtx_offsets(&off_v1.tx);
1247                         xsk_enter_umem_offsets(&off_v1.fr);
1248                         xsk_enter_umem_offsets(&off_v1.cr);
1249
1250                         len = sizeof(off_v1);
1251                         to_copy = &off_v1;
1252                 }
1253
1254                 if (copy_to_user(optval, to_copy, len))
1255                         return -EFAULT;
1256                 if (put_user(len, optlen))
1257                         return -EFAULT;
1258
1259                 return 0;
1260         }
1261         case XDP_OPTIONS:
1262         {
1263                 struct xdp_options opts = {};
1264
1265                 if (len < sizeof(opts))
1266                         return -EINVAL;
1267
1268                 mutex_lock(&xs->mutex);
1269                 if (xs->zc)
1270                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1271                 mutex_unlock(&xs->mutex);
1272
1273                 len = sizeof(opts);
1274                 if (copy_to_user(optval, &opts, len))
1275                         return -EFAULT;
1276                 if (put_user(len, optlen))
1277                         return -EFAULT;
1278
1279                 return 0;
1280         }
1281         default:
1282                 break;
1283         }
1284
1285         return -EOPNOTSUPP;
1286 }
1287
1288 static int xsk_mmap(struct file *file, struct socket *sock,
1289                     struct vm_area_struct *vma)
1290 {
1291         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1292         unsigned long size = vma->vm_end - vma->vm_start;
1293         struct xdp_sock *xs = xdp_sk(sock->sk);
1294         struct xsk_queue *q = NULL;
1295         unsigned long pfn;
1296         struct page *qpg;
1297
1298         if (READ_ONCE(xs->state) != XSK_READY)
1299                 return -EBUSY;
1300
1301         if (offset == XDP_PGOFF_RX_RING) {
1302                 q = READ_ONCE(xs->rx);
1303         } else if (offset == XDP_PGOFF_TX_RING) {
1304                 q = READ_ONCE(xs->tx);
1305         } else {
1306                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1307                 smp_rmb();
1308                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1309                         q = READ_ONCE(xs->fq_tmp);
1310                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1311                         q = READ_ONCE(xs->cq_tmp);
1312         }
1313
1314         if (!q)
1315                 return -EINVAL;
1316
1317         /* Matches the smp_wmb() in xsk_init_queue */
1318         smp_rmb();
1319         qpg = virt_to_head_page(q->ring);
1320         if (size > page_size(qpg))
1321                 return -EINVAL;
1322
1323         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1324         return remap_pfn_range(vma, vma->vm_start, pfn,
1325                                size, vma->vm_page_prot);
1326 }
1327
1328 static int xsk_notifier(struct notifier_block *this,
1329                         unsigned long msg, void *ptr)
1330 {
1331         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1332         struct net *net = dev_net(dev);
1333         struct sock *sk;
1334
1335         switch (msg) {
1336         case NETDEV_UNREGISTER:
1337                 mutex_lock(&net->xdp.lock);
1338                 sk_for_each(sk, &net->xdp.list) {
1339                         struct xdp_sock *xs = xdp_sk(sk);
1340
1341                         mutex_lock(&xs->mutex);
1342                         if (xs->dev == dev) {
1343                                 sk->sk_err = ENETDOWN;
1344                                 if (!sock_flag(sk, SOCK_DEAD))
1345                                         sk_error_report(sk);
1346
1347                                 xsk_unbind_dev(xs);
1348
1349                                 /* Clear device references. */
1350                                 xp_clear_dev(xs->pool);
1351                         }
1352                         mutex_unlock(&xs->mutex);
1353                 }
1354                 mutex_unlock(&net->xdp.lock);
1355                 break;
1356         }
1357         return NOTIFY_DONE;
1358 }
1359
1360 static struct proto xsk_proto = {
1361         .name =         "XDP",
1362         .owner =        THIS_MODULE,
1363         .obj_size =     sizeof(struct xdp_sock),
1364 };
1365
1366 static const struct proto_ops xsk_proto_ops = {
1367         .family         = PF_XDP,
1368         .owner          = THIS_MODULE,
1369         .release        = xsk_release,
1370         .bind           = xsk_bind,
1371         .connect        = sock_no_connect,
1372         .socketpair     = sock_no_socketpair,
1373         .accept         = sock_no_accept,
1374         .getname        = sock_no_getname,
1375         .poll           = xsk_poll,
1376         .ioctl          = sock_no_ioctl,
1377         .listen         = sock_no_listen,
1378         .shutdown       = sock_no_shutdown,
1379         .setsockopt     = xsk_setsockopt,
1380         .getsockopt     = xsk_getsockopt,
1381         .sendmsg        = xsk_sendmsg,
1382         .recvmsg        = xsk_recvmsg,
1383         .mmap           = xsk_mmap,
1384         .sendpage       = sock_no_sendpage,
1385 };
1386
1387 static void xsk_destruct(struct sock *sk)
1388 {
1389         struct xdp_sock *xs = xdp_sk(sk);
1390
1391         if (!sock_flag(sk, SOCK_DEAD))
1392                 return;
1393
1394         if (!xp_put_pool(xs->pool))
1395                 xdp_put_umem(xs->umem, !xs->pool);
1396
1397         sk_refcnt_debug_dec(sk);
1398 }
1399
1400 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1401                       int kern)
1402 {
1403         struct xdp_sock *xs;
1404         struct sock *sk;
1405
1406         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1407                 return -EPERM;
1408         if (sock->type != SOCK_RAW)
1409                 return -ESOCKTNOSUPPORT;
1410
1411         if (protocol)
1412                 return -EPROTONOSUPPORT;
1413
1414         sock->state = SS_UNCONNECTED;
1415
1416         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1417         if (!sk)
1418                 return -ENOBUFS;
1419
1420         sock->ops = &xsk_proto_ops;
1421
1422         sock_init_data(sock, sk);
1423
1424         sk->sk_family = PF_XDP;
1425
1426         sk->sk_destruct = xsk_destruct;
1427         sk_refcnt_debug_inc(sk);
1428
1429         sock_set_flag(sk, SOCK_RCU_FREE);
1430
1431         xs = xdp_sk(sk);
1432         xs->state = XSK_READY;
1433         mutex_init(&xs->mutex);
1434         spin_lock_init(&xs->rx_lock);
1435
1436         INIT_LIST_HEAD(&xs->map_list);
1437         spin_lock_init(&xs->map_list_lock);
1438
1439         mutex_lock(&net->xdp.lock);
1440         sk_add_node_rcu(sk, &net->xdp.list);
1441         mutex_unlock(&net->xdp.lock);
1442
1443         sock_prot_inuse_add(net, &xsk_proto, 1);
1444
1445         return 0;
1446 }
1447
1448 static const struct net_proto_family xsk_family_ops = {
1449         .family = PF_XDP,
1450         .create = xsk_create,
1451         .owner  = THIS_MODULE,
1452 };
1453
1454 static struct notifier_block xsk_netdev_notifier = {
1455         .notifier_call  = xsk_notifier,
1456 };
1457
1458 static int __net_init xsk_net_init(struct net *net)
1459 {
1460         mutex_init(&net->xdp.lock);
1461         INIT_HLIST_HEAD(&net->xdp.list);
1462         return 0;
1463 }
1464
1465 static void __net_exit xsk_net_exit(struct net *net)
1466 {
1467         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1468 }
1469
1470 static struct pernet_operations xsk_net_ops = {
1471         .init = xsk_net_init,
1472         .exit = xsk_net_exit,
1473 };
1474
1475 static int __init xsk_init(void)
1476 {
1477         int err, cpu;
1478
1479         err = proto_register(&xsk_proto, 0 /* no slab */);
1480         if (err)
1481                 goto out;
1482
1483         err = sock_register(&xsk_family_ops);
1484         if (err)
1485                 goto out_proto;
1486
1487         err = register_pernet_subsys(&xsk_net_ops);
1488         if (err)
1489                 goto out_sk;
1490
1491         err = register_netdevice_notifier(&xsk_netdev_notifier);
1492         if (err)
1493                 goto out_pernet;
1494
1495         for_each_possible_cpu(cpu)
1496                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1497         return 0;
1498
1499 out_pernet:
1500         unregister_pernet_subsys(&xsk_net_ops);
1501 out_sk:
1502         sock_unregister(PF_XDP);
1503 out_proto:
1504         proto_unregister(&xsk_proto);
1505 out:
1506         return err;
1507 }
1508
1509 fs_initcall(xsk_init);