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 nb_pkts)
 359 {
 360         struct xdp_sock *xs;
 361
 362         rcu_read_lock();
 363         if (!list_is_singular(&pool->xsk_tx_list)) {
 364                 /* Fallback to the non-batched version */
 365                 rcu_read_unlock();
 366                 return xsk_tx_peek_release_fallback(pool, nb_pkts);
 367         }
 368
 369         xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
 370         if (!xs) {
 371                 nb_pkts = 0;
 372                 goto out;
 373         }
 374
 375         nb_pkts = xskq_cons_nb_entries(xs->tx, nb_pkts);
 376
 377         /* This is the backpressure mechanism for the Tx path. Try to
 378          * reserve space in the completion queue for all packets, but
 379          * if there are fewer slots available, just process that many
 380          * packets. This avoids having to implement any buffering in
 381          * the Tx path.
 382          */
 383         nb_pkts = xskq_prod_nb_free(pool->cq, nb_pkts);
 384         if (!nb_pkts)
 385                 goto out;
 386
 387         nb_pkts = xskq_cons_read_desc_batch(xs->tx, pool, nb_pkts);
 388         if (!nb_pkts) {
 389                 xs->tx->queue_empty_descs++;
 390                 goto out;
 391         }
 392
 393         __xskq_cons_release(xs->tx);
 394         xskq_prod_write_addr_batch(pool->cq, pool->tx_descs, nb_pkts);
 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_generic_xmit(struct sock *sk)
 598 {
 599         int ret;
 600
 601         /* Drop the RCU lock since the SKB path might sleep. */
 602         rcu_read_unlock();
 603         ret = __xsk_generic_xmit(sk);
 604         /* Reaquire RCU lock before going into common code. */
 605         rcu_read_lock();
 606
 607         return ret;
 608 }
 609
 610 static bool xsk_no_wakeup(struct sock *sk)
 611 {
 612 #ifdef CONFIG_NET_RX_BUSY_POLL
 613         /* Prefer busy-polling, skip the wakeup. */
 614         return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
 615                 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
 616 #else
 617         return false;
 618 #endif
 619 }
 620
 621 static int xsk_check_common(struct xdp_sock *xs)
 622 {
 623         if (unlikely(!xsk_is_bound(xs)))
 624                 return -ENXIO;
 625         if (unlikely(!(xs->dev->flags & IFF_UP)))
 626                 return -ENETDOWN;
 627
 628         return 0;
 629 }
 630
 631 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 632 {
 633         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 634         struct sock *sk = sock->sk;
 635         struct xdp_sock *xs = xdp_sk(sk);
 636         struct xsk_buff_pool *pool;
 637         int err;
 638
 639         err = xsk_check_common(xs);
 640         if (err)
 641                 return err;
 642         if (unlikely(need_wait))
 643                 return -EOPNOTSUPP;
 644         if (unlikely(!xs->tx))
 645                 return -ENOBUFS;
 646
 647         if (sk_can_busy_loop(sk)) {
 648                 if (xs->zc)
 649                         __sk_mark_napi_id_once(sk, xsk_pool_get_napi_id(xs->pool));
 650                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 651         }
 652
 653         if (xs->zc && xsk_no_wakeup(sk))
 654                 return 0;
 655
 656         pool = xs->pool;
 657         if (pool->cached_need_wakeup & XDP_WAKEUP_TX) {
 658                 if (xs->zc)
 659                         return xsk_wakeup(xs, XDP_WAKEUP_TX);
 660                 return xsk_generic_xmit(sk);
 661         }
 662         return 0;
 663 }
 664
 665 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 666 {
 667         int ret;
 668
 669         rcu_read_lock();
 670         ret = __xsk_sendmsg(sock, m, total_len);
 671         rcu_read_unlock();
 672
 673         return ret;
 674 }
 675
 676 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 677 {
 678         bool need_wait = !(flags & MSG_DONTWAIT);
 679         struct sock *sk = sock->sk;
 680         struct xdp_sock *xs = xdp_sk(sk);
 681         int err;
 682
 683         err = xsk_check_common(xs);
 684         if (err)
 685                 return err;
 686         if (unlikely(!xs->rx))
 687                 return -ENOBUFS;
 688         if (unlikely(need_wait))
 689                 return -EOPNOTSUPP;
 690
 691         if (sk_can_busy_loop(sk))
 692                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 693
 694         if (xsk_no_wakeup(sk))
 695                 return 0;
 696
 697         if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
 698                 return xsk_wakeup(xs, XDP_WAKEUP_RX);
 699         return 0;
 700 }
 701
 702 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 703 {
 704         int ret;
 705
 706         rcu_read_lock();
 707         ret = __xsk_recvmsg(sock, m, len, flags);
 708         rcu_read_unlock();
 709
 710         return ret;
 711 }
 712
 713 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 714                              struct poll_table_struct *wait)
 715 {
 716         __poll_t mask = 0;
 717         struct sock *sk = sock->sk;
 718         struct xdp_sock *xs = xdp_sk(sk);
 719         struct xsk_buff_pool *pool;
 720
 721         sock_poll_wait(file, sock, wait);
 722
 723         rcu_read_lock();
 724         if (xsk_check_common(xs))
 725                 goto skip_tx;
 726
 727         pool = xs->pool;
 728
 729         if (pool->cached_need_wakeup) {
 730                 if (xs->zc)
 731                         xsk_wakeup(xs, pool->cached_need_wakeup);
 732                 else if (xs->tx)
 733                         /* Poll needs to drive Tx also in copy mode */
 734                         xsk_generic_xmit(sk);
 735         }
 736
 737 skip_tx:
 738         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 739                 mask |= EPOLLIN | EPOLLRDNORM;
 740         if (xs->tx && xsk_tx_writeable(xs))
 741                 mask |= EPOLLOUT | EPOLLWRNORM;
 742
 743         rcu_read_unlock();
 744         return mask;
 745 }
 746
 747 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 748                           bool umem_queue)
 749 {
 750         struct xsk_queue *q;
 751
 752         if (entries == 0 || *queue || !is_power_of_2(entries))
 753                 return -EINVAL;
 754
 755         q = xskq_create(entries, umem_queue);
 756         if (!q)
 757                 return -ENOMEM;
 758
 759         /* Make sure queue is ready before it can be seen by others */
 760         smp_wmb();
 761         WRITE_ONCE(*queue, q);
 762         return 0;
 763 }
 764
 765 static void xsk_unbind_dev(struct xdp_sock *xs)
 766 {
 767         struct net_device *dev = xs->dev;
 768
 769         if (xs->state != XSK_BOUND)
 770                 return;
 771         WRITE_ONCE(xs->state, XSK_UNBOUND);
 772
 773         /* Wait for driver to stop using the xdp socket. */
 774         xp_del_xsk(xs->pool, xs);
 775         synchronize_net();
 776         dev_put(dev);
 777 }
 778
 779 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 780                                               struct xdp_sock __rcu ***map_entry)
 781 {
 782         struct xsk_map *map = NULL;
 783         struct xsk_map_node *node;
 784
 785         *map_entry = NULL;
 786
 787         spin_lock_bh(&xs->map_list_lock);
 788         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 789                                         node);
 790         if (node) {
 791                 bpf_map_inc(&node->map->map);
 792                 map = node->map;
 793                 *map_entry = node->map_entry;
 794         }
 795         spin_unlock_bh(&xs->map_list_lock);
 796         return map;
 797 }
 798
 799 static void xsk_delete_from_maps(struct xdp_sock *xs)
 800 {
 801         /* This function removes the current XDP socket from all the
 802          * maps it resides in. We need to take extra care here, due to
 803          * the two locks involved. Each map has a lock synchronizing
 804          * updates to the entries, and each socket has a lock that
 805          * synchronizes access to the list of maps (map_list). For
 806          * deadlock avoidance the locks need to be taken in the order
 807          * "map lock"->"socket map list lock". We start off by
 808          * accessing the socket map list, and take a reference to the
 809          * map to guarantee existence between the
 810          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 811          * calls. Then we ask the map to remove the socket, which
 812          * tries to remove the socket from the map. Note that there
 813          * might be updates to the map between
 814          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 815          */
 816         struct xdp_sock __rcu **map_entry = NULL;
 817         struct xsk_map *map;
 818
 819         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 820                 xsk_map_try_sock_delete(map, xs, map_entry);
 821                 bpf_map_put(&map->map);
 822         }
 823 }
 824
 825 static int xsk_release(struct socket *sock)
 826 {
 827         struct sock *sk = sock->sk;
 828         struct xdp_sock *xs = xdp_sk(sk);
 829         struct net *net;
 830
 831         if (!sk)
 832                 return 0;
 833
 834         net = sock_net(sk);
 835
 836         mutex_lock(&net->xdp.lock);
 837         sk_del_node_init_rcu(sk);
 838         mutex_unlock(&net->xdp.lock);
 839
 840         sock_prot_inuse_add(net, sk->sk_prot, -1);
 841
 842         xsk_delete_from_maps(xs);
 843         mutex_lock(&xs->mutex);
 844         xsk_unbind_dev(xs);
 845         mutex_unlock(&xs->mutex);
 846
 847         xskq_destroy(xs->rx);
 848         xskq_destroy(xs->tx);
 849         xskq_destroy(xs->fq_tmp);
 850         xskq_destroy(xs->cq_tmp);
 851
 852         sock_orphan(sk);
 853         sock->sk = NULL;
 854
 855         sk_refcnt_debug_release(sk);
 856         sock_put(sk);
 857
 858         return 0;
 859 }
 860
 861 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 862 {
 863         struct socket *sock;
 864         int err;
 865
 866         sock = sockfd_lookup(fd, &err);
 867         if (!sock)
 868                 return ERR_PTR(-ENOTSOCK);
 869
 870         if (sock->sk->sk_family != PF_XDP) {
 871                 sockfd_put(sock);
 872                 return ERR_PTR(-ENOPROTOOPT);
 873         }
 874
 875         return sock;
 876 }
 877
 878 static bool xsk_validate_queues(struct xdp_sock *xs)
 879 {
 880         return xs->fq_tmp && xs->cq_tmp;
 881 }
 882
 883 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 884 {
 885         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 886         struct sock *sk = sock->sk;
 887         struct xdp_sock *xs = xdp_sk(sk);
 888         struct net_device *dev;
 889         u32 flags, qid;
 890         int err = 0;
 891
 892         if (addr_len < sizeof(struct sockaddr_xdp))
 893                 return -EINVAL;
 894         if (sxdp->sxdp_family != AF_XDP)
 895                 return -EINVAL;
 896
 897         flags = sxdp->sxdp_flags;
 898         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 899                       XDP_USE_NEED_WAKEUP))
 900                 return -EINVAL;
 901
 902         rtnl_lock();
 903         mutex_lock(&xs->mutex);
 904         if (xs->state != XSK_READY) {
 905                 err = -EBUSY;
 906                 goto out_release;
 907         }
 908
 909         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 910         if (!dev) {
 911                 err = -ENODEV;
 912                 goto out_release;
 913         }
 914
 915         if (!xs->rx && !xs->tx) {
 916                 err = -EINVAL;
 917                 goto out_unlock;
 918         }
 919
 920         qid = sxdp->sxdp_queue_id;
 921
 922         if (flags & XDP_SHARED_UMEM) {
 923                 struct xdp_sock *umem_xs;
 924                 struct socket *sock;
 925
 926                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 927                     (flags & XDP_USE_NEED_WAKEUP)) {
 928                         /* Cannot specify flags for shared sockets. */
 929                         err = -EINVAL;
 930                         goto out_unlock;
 931                 }
 932
 933                 if (xs->umem) {
 934                         /* We have already our own. */
 935                         err = -EINVAL;
 936                         goto out_unlock;
 937                 }
 938
 939                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 940                 if (IS_ERR(sock)) {
 941                         err = PTR_ERR(sock);
 942                         goto out_unlock;
 943                 }
 944
 945                 umem_xs = xdp_sk(sock->sk);
 946                 if (!xsk_is_bound(umem_xs)) {
 947                         err = -EBADF;
 948                         sockfd_put(sock);
 949                         goto out_unlock;
 950                 }
 951
 952                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 953                         /* Share the umem with another socket on another qid
 954                          * and/or device.
 955                          */
 956                         xs->pool = xp_create_and_assign_umem(xs,
 957                                                              umem_xs->umem);
 958                         if (!xs->pool) {
 959                                 err = -ENOMEM;
 960                                 sockfd_put(sock);
 961                                 goto out_unlock;
 962                         }
 963
 964                         err = xp_assign_dev_shared(xs->pool, umem_xs, dev,
 965                                                    qid);
 966                         if (err) {
 967                                 xp_destroy(xs->pool);
 968                                 xs->pool = NULL;
 969                                 sockfd_put(sock);
 970                                 goto out_unlock;
 971                         }
 972                 } else {
 973                         /* Share the buffer pool with the other socket. */
 974                         if (xs->fq_tmp || xs->cq_tmp) {
 975                                 /* Do not allow setting your own fq or cq. */
 976                                 err = -EINVAL;
 977                                 sockfd_put(sock);
 978                                 goto out_unlock;
 979                         }
 980
 981                         xp_get_pool(umem_xs->pool);
 982                         xs->pool = umem_xs->pool;
 983
 984                         /* If underlying shared umem was created without Tx
 985                          * ring, allocate Tx descs array that Tx batching API
 986                          * utilizes
 987                          */
 988                         if (xs->tx && !xs->pool->tx_descs) {
 989                                 err = xp_alloc_tx_descs(xs->pool, xs);
 990                                 if (err) {
 991                                         xp_put_pool(xs->pool);
 992                                         sockfd_put(sock);
 993                                         goto out_unlock;
 994                                 }
 995                         }
 996                 }
 997
 998                 xdp_get_umem(umem_xs->umem);
 999                 WRITE_ONCE(xs->umem, umem_xs->umem);
1000                 sockfd_put(sock);
1001         } else if (!xs->umem || !xsk_validate_queues(xs)) {
1002                 err = -EINVAL;
1003                 goto out_unlock;
1004         } else {
1005                 /* This xsk has its own umem. */
1006                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
1007                 if (!xs->pool) {
1008                         err = -ENOMEM;
1009                         goto out_unlock;
1010                 }
1011
1012                 err = xp_assign_dev(xs->pool, dev, qid, flags);
1013                 if (err) {
1014                         xp_destroy(xs->pool);
1015                         xs->pool = NULL;
1016                         goto out_unlock;
1017                 }
1018         }
1019
1020         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1021         xs->fq_tmp = NULL;
1022         xs->cq_tmp = NULL;
1023
1024         xs->dev = dev;
1025         xs->zc = xs->umem->zc;
1026         xs->queue_id = qid;
1027         xp_add_xsk(xs->pool, xs);
1028
1029 out_unlock:
1030         if (err) {
1031                 dev_put(dev);
1032         } else {
1033                 /* Matches smp_rmb() in bind() for shared umem
1034                  * sockets, and xsk_is_bound().
1035                  */
1036                 smp_wmb();
1037                 WRITE_ONCE(xs->state, XSK_BOUND);
1038         }
1039 out_release:
1040         mutex_unlock(&xs->mutex);
1041         rtnl_unlock();
1042         return err;
1043 }
1044
1045 struct xdp_umem_reg_v1 {
1046         __u64 addr; /* Start of packet data area */
1047         __u64 len; /* Length of packet data area */
1048         __u32 chunk_size;
1049         __u32 headroom;
1050 };
1051
1052 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1053                           sockptr_t optval, unsigned int optlen)
1054 {
1055         struct sock *sk = sock->sk;
1056         struct xdp_sock *xs = xdp_sk(sk);
1057         int err;
1058
1059         if (level != SOL_XDP)
1060                 return -ENOPROTOOPT;
1061
1062         switch (optname) {
1063         case XDP_RX_RING:
1064         case XDP_TX_RING:
1065         {
1066                 struct xsk_queue **q;
1067                 int entries;
1068
1069                 if (optlen < sizeof(entries))
1070                         return -EINVAL;
1071                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1072                         return -EFAULT;
1073
1074                 mutex_lock(&xs->mutex);
1075                 if (xs->state != XSK_READY) {
1076                         mutex_unlock(&xs->mutex);
1077                         return -EBUSY;
1078                 }
1079                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1080                 err = xsk_init_queue(entries, q, false);
1081                 if (!err && optname == XDP_TX_RING)
1082                         /* Tx needs to be explicitly woken up the first time */
1083                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1084                 mutex_unlock(&xs->mutex);
1085                 return err;
1086         }
1087         case XDP_UMEM_REG:
1088         {
1089                 size_t mr_size = sizeof(struct xdp_umem_reg);
1090                 struct xdp_umem_reg mr = {};
1091                 struct xdp_umem *umem;
1092
1093                 if (optlen < sizeof(struct xdp_umem_reg_v1))
1094                         return -EINVAL;
1095                 else if (optlen < sizeof(mr))
1096                         mr_size = sizeof(struct xdp_umem_reg_v1);
1097
1098                 if (copy_from_sockptr(&mr, optval, mr_size))
1099                         return -EFAULT;
1100
1101                 mutex_lock(&xs->mutex);
1102                 if (xs->state != XSK_READY || xs->umem) {
1103                         mutex_unlock(&xs->mutex);
1104                         return -EBUSY;
1105                 }
1106
1107                 umem = xdp_umem_create(&mr);
1108                 if (IS_ERR(umem)) {
1109                         mutex_unlock(&xs->mutex);
1110                         return PTR_ERR(umem);
1111                 }
1112
1113                 /* Make sure umem is ready before it can be seen by others */
1114                 smp_wmb();
1115                 WRITE_ONCE(xs->umem, umem);
1116                 mutex_unlock(&xs->mutex);
1117                 return 0;
1118         }
1119         case XDP_UMEM_FILL_RING:
1120         case XDP_UMEM_COMPLETION_RING:
1121         {
1122                 struct xsk_queue **q;
1123                 int entries;
1124
1125                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1126                         return -EFAULT;
1127
1128                 mutex_lock(&xs->mutex);
1129                 if (xs->state != XSK_READY) {
1130                         mutex_unlock(&xs->mutex);
1131                         return -EBUSY;
1132                 }
1133
1134                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1135                         &xs->cq_tmp;
1136                 err = xsk_init_queue(entries, q, true);
1137                 mutex_unlock(&xs->mutex);
1138                 return err;
1139         }
1140         default:
1141                 break;
1142         }
1143
1144         return -ENOPROTOOPT;
1145 }
1146
1147 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1148 {
1149         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1150         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1151         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1152 }
1153
1154 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1155 {
1156         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1157         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1158         ring->desc = offsetof(struct xdp_umem_ring, desc);
1159 }
1160
1161 struct xdp_statistics_v1 {
1162         __u64 rx_dropped;
1163         __u64 rx_invalid_descs;
1164         __u64 tx_invalid_descs;
1165 };
1166
1167 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1168                           char __user *optval, int __user *optlen)
1169 {
1170         struct sock *sk = sock->sk;
1171         struct xdp_sock *xs = xdp_sk(sk);
1172         int len;
1173
1174         if (level != SOL_XDP)
1175                 return -ENOPROTOOPT;
1176
1177         if (get_user(len, optlen))
1178                 return -EFAULT;
1179         if (len < 0)
1180                 return -EINVAL;
1181
1182         switch (optname) {
1183         case XDP_STATISTICS:
1184         {
1185                 struct xdp_statistics stats = {};
1186                 bool extra_stats = true;
1187                 size_t stats_size;
1188
1189                 if (len < sizeof(struct xdp_statistics_v1)) {
1190                         return -EINVAL;
1191                 } else if (len < sizeof(stats)) {
1192                         extra_stats = false;
1193                         stats_size = sizeof(struct xdp_statistics_v1);
1194                 } else {
1195                         stats_size = sizeof(stats);
1196                 }
1197
1198                 mutex_lock(&xs->mutex);
1199                 stats.rx_dropped = xs->rx_dropped;
1200                 if (extra_stats) {
1201                         stats.rx_ring_full = xs->rx_queue_full;
1202                         stats.rx_fill_ring_empty_descs =
1203                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1204                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1205                 } else {
1206                         stats.rx_dropped += xs->rx_queue_full;
1207                 }
1208                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1209                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1210                 mutex_unlock(&xs->mutex);
1211
1212                 if (copy_to_user(optval, &stats, stats_size))
1213                         return -EFAULT;
1214                 if (put_user(stats_size, optlen))
1215                         return -EFAULT;
1216
1217                 return 0;
1218         }
1219         case XDP_MMAP_OFFSETS:
1220         {
1221                 struct xdp_mmap_offsets off;
1222                 struct xdp_mmap_offsets_v1 off_v1;
1223                 bool flags_supported = true;
1224                 void *to_copy;
1225
1226                 if (len < sizeof(off_v1))
1227                         return -EINVAL;
1228                 else if (len < sizeof(off))
1229                         flags_supported = false;
1230
1231                 if (flags_supported) {
1232                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1233                          * except for the flags field added to the end.
1234                          */
1235                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1236                                                &off.rx);
1237                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1238                                                &off.tx);
1239                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1240                                                &off.fr);
1241                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1242                                                &off.cr);
1243                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1244                                                 ptrs.flags);
1245                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1246                                                 ptrs.flags);
1247                         off.fr.flags = offsetof(struct xdp_umem_ring,
1248                                                 ptrs.flags);
1249                         off.cr.flags = offsetof(struct xdp_umem_ring,
1250                                                 ptrs.flags);
1251
1252                         len = sizeof(off);
1253                         to_copy = &off;
1254                 } else {
1255                         xsk_enter_rxtx_offsets(&off_v1.rx);
1256                         xsk_enter_rxtx_offsets(&off_v1.tx);
1257                         xsk_enter_umem_offsets(&off_v1.fr);
1258                         xsk_enter_umem_offsets(&off_v1.cr);
1259
1260                         len = sizeof(off_v1);
1261                         to_copy = &off_v1;
1262                 }
1263
1264                 if (copy_to_user(optval, to_copy, len))
1265                         return -EFAULT;
1266                 if (put_user(len, optlen))
1267                         return -EFAULT;
1268
1269                 return 0;
1270         }
1271         case XDP_OPTIONS:
1272         {
1273                 struct xdp_options opts = {};
1274
1275                 if (len < sizeof(opts))
1276                         return -EINVAL;
1277
1278                 mutex_lock(&xs->mutex);
1279                 if (xs->zc)
1280                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1281                 mutex_unlock(&xs->mutex);
1282
1283                 len = sizeof(opts);
1284                 if (copy_to_user(optval, &opts, len))
1285                         return -EFAULT;
1286                 if (put_user(len, optlen))
1287                         return -EFAULT;
1288
1289                 return 0;
1290         }
1291         default:
1292                 break;
1293         }
1294
1295         return -EOPNOTSUPP;
1296 }
1297
1298 static int xsk_mmap(struct file *file, struct socket *sock,
1299                     struct vm_area_struct *vma)
1300 {
1301         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1302         unsigned long size = vma->vm_end - vma->vm_start;
1303         struct xdp_sock *xs = xdp_sk(sock->sk);
1304         struct xsk_queue *q = NULL;
1305         unsigned long pfn;
1306         struct page *qpg;
1307
1308         if (READ_ONCE(xs->state) != XSK_READY)
1309                 return -EBUSY;
1310
1311         if (offset == XDP_PGOFF_RX_RING) {
1312                 q = READ_ONCE(xs->rx);
1313         } else if (offset == XDP_PGOFF_TX_RING) {
1314                 q = READ_ONCE(xs->tx);
1315         } else {
1316                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1317                 smp_rmb();
1318                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1319                         q = READ_ONCE(xs->fq_tmp);
1320                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1321                         q = READ_ONCE(xs->cq_tmp);
1322         }
1323
1324         if (!q)
1325                 return -EINVAL;
1326
1327         /* Matches the smp_wmb() in xsk_init_queue */
1328         smp_rmb();
1329         qpg = virt_to_head_page(q->ring);
1330         if (size > page_size(qpg))
1331                 return -EINVAL;
1332
1333         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1334         return remap_pfn_range(vma, vma->vm_start, pfn,
1335                                size, vma->vm_page_prot);
1336 }
1337
1338 static int xsk_notifier(struct notifier_block *this,
1339                         unsigned long msg, void *ptr)
1340 {
1341         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1342         struct net *net = dev_net(dev);
1343         struct sock *sk;
1344
1345         switch (msg) {
1346         case NETDEV_UNREGISTER:
1347                 mutex_lock(&net->xdp.lock);
1348                 sk_for_each(sk, &net->xdp.list) {
1349                         struct xdp_sock *xs = xdp_sk(sk);
1350
1351                         mutex_lock(&xs->mutex);
1352                         if (xs->dev == dev) {
1353                                 sk->sk_err = ENETDOWN;
1354                                 if (!sock_flag(sk, SOCK_DEAD))
1355                                         sk_error_report(sk);
1356
1357                                 xsk_unbind_dev(xs);
1358
1359                                 /* Clear device references. */
1360                                 xp_clear_dev(xs->pool);
1361                         }
1362                         mutex_unlock(&xs->mutex);
1363                 }
1364                 mutex_unlock(&net->xdp.lock);
1365                 break;
1366         }
1367         return NOTIFY_DONE;
1368 }
1369
1370 static struct proto xsk_proto = {
1371         .name =         "XDP",
1372         .owner =        THIS_MODULE,
1373         .obj_size =     sizeof(struct xdp_sock),
1374 };
1375
1376 static const struct proto_ops xsk_proto_ops = {
1377         .family         = PF_XDP,
1378         .owner          = THIS_MODULE,
1379         .release        = xsk_release,
1380         .bind           = xsk_bind,
1381         .connect        = sock_no_connect,
1382         .socketpair     = sock_no_socketpair,
1383         .accept         = sock_no_accept,
1384         .getname        = sock_no_getname,
1385         .poll           = xsk_poll,
1386         .ioctl          = sock_no_ioctl,
1387         .listen         = sock_no_listen,
1388         .shutdown       = sock_no_shutdown,
1389         .setsockopt     = xsk_setsockopt,
1390         .getsockopt     = xsk_getsockopt,
1391         .sendmsg        = xsk_sendmsg,
1392         .recvmsg        = xsk_recvmsg,
1393         .mmap           = xsk_mmap,
1394         .sendpage       = sock_no_sendpage,
1395 };
1396
1397 static void xsk_destruct(struct sock *sk)
1398 {
1399         struct xdp_sock *xs = xdp_sk(sk);
1400
1401         if (!sock_flag(sk, SOCK_DEAD))
1402                 return;
1403
1404         if (!xp_put_pool(xs->pool))
1405                 xdp_put_umem(xs->umem, !xs->pool);
1406
1407         sk_refcnt_debug_dec(sk);
1408 }
1409
1410 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1411                       int kern)
1412 {
1413         struct xdp_sock *xs;
1414         struct sock *sk;
1415
1416         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1417                 return -EPERM;
1418         if (sock->type != SOCK_RAW)
1419                 return -ESOCKTNOSUPPORT;
1420
1421         if (protocol)
1422                 return -EPROTONOSUPPORT;
1423
1424         sock->state = SS_UNCONNECTED;
1425
1426         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1427         if (!sk)
1428                 return -ENOBUFS;
1429
1430         sock->ops = &xsk_proto_ops;
1431
1432         sock_init_data(sock, sk);
1433
1434         sk->sk_family = PF_XDP;
1435
1436         sk->sk_destruct = xsk_destruct;
1437         sk_refcnt_debug_inc(sk);
1438
1439         sock_set_flag(sk, SOCK_RCU_FREE);
1440
1441         xs = xdp_sk(sk);
1442         xs->state = XSK_READY;
1443         mutex_init(&xs->mutex);
1444         spin_lock_init(&xs->rx_lock);
1445
1446         INIT_LIST_HEAD(&xs->map_list);
1447         spin_lock_init(&xs->map_list_lock);
1448
1449         mutex_lock(&net->xdp.lock);
1450         sk_add_node_rcu(sk, &net->xdp.list);
1451         mutex_unlock(&net->xdp.lock);
1452
1453         sock_prot_inuse_add(net, &xsk_proto, 1);
1454
1455         return 0;
1456 }
1457
1458 static const struct net_proto_family xsk_family_ops = {
1459         .family = PF_XDP,
1460         .create = xsk_create,
1461         .owner  = THIS_MODULE,
1462 };
1463
1464 static struct notifier_block xsk_netdev_notifier = {
1465         .notifier_call  = xsk_notifier,
1466 };
1467
1468 static int __net_init xsk_net_init(struct net *net)
1469 {
1470         mutex_init(&net->xdp.lock);
1471         INIT_HLIST_HEAD(&net->xdp.list);
1472         return 0;
1473 }
1474
1475 static void __net_exit xsk_net_exit(struct net *net)
1476 {
1477         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1478 }
1479
1480 static struct pernet_operations xsk_net_ops = {
1481         .init = xsk_net_init,
1482         .exit = xsk_net_exit,
1483 };
1484
1485 static int __init xsk_init(void)
1486 {
1487         int err, cpu;
1488
1489         err = proto_register(&xsk_proto, 0 /* no slab */);
1490         if (err)
1491                 goto out;
1492
1493         err = sock_register(&xsk_family_ops);
1494         if (err)
1495                 goto out_proto;
1496
1497         err = register_pernet_subsys(&xsk_net_ops);
1498         if (err)
1499                 goto out_sk;
1500
1501         err = register_netdevice_notifier(&xsk_netdev_notifier);
1502         if (err)
1503                 goto out_pernet;
1504
1505         for_each_possible_cpu(cpu)
1506                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1507         return 0;
1508
1509 out_pernet:
1510         unregister_pernet_subsys(&xsk_net_ops);
1511 out_sk:
1512         sock_unregister(PF_XDP);
1513 out_proto:
1514         proto_unregister(&xsk_proto);
1515 out:
1516         return err;
1517 }
1518
1519 fs_initcall(xsk_init);