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/xdp.h>
  27
  28 #include "xsk_queue.h"
  29 #include "xdp_umem.h"
  30 #include "xsk.h"
  31
  32 #define TX_BATCH_SIZE 16
  33
  34 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list);
  35
  36 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool)
  37 {
  38         if (pool->cached_need_wakeup & XDP_WAKEUP_RX)
  39                 return;
  40
  41         pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
  42         pool->cached_need_wakeup |= XDP_WAKEUP_RX;
  43 }
  44 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
  45
  46 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool)
  47 {
  48         struct xdp_sock *xs;
  49
  50         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
  51                 return;
  52
  53         rcu_read_lock();
  54         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
  55                 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
  56         }
  57         rcu_read_unlock();
  58
  59         pool->cached_need_wakeup |= XDP_WAKEUP_TX;
  60 }
  61 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
  62
  63 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool)
  64 {
  65         if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX))
  66                 return;
  67
  68         pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
  69         pool->cached_need_wakeup &= ~XDP_WAKEUP_RX;
  70 }
  71 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
  72
  73 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool)
  74 {
  75         struct xdp_sock *xs;
  76
  77         if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX))
  78                 return;
  79
  80         rcu_read_lock();
  81         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
  82                 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
  83         }
  84         rcu_read_unlock();
  85
  86         pool->cached_need_wakeup &= ~XDP_WAKEUP_TX;
  87 }
  88 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
  89
  90 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool)
  91 {
  92         return pool->uses_need_wakeup;
  93 }
  94 EXPORT_SYMBOL(xsk_uses_need_wakeup);
  95
  96 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev,
  97                                             u16 queue_id)
  98 {
  99         if (queue_id < dev->real_num_rx_queues)
 100                 return dev->_rx[queue_id].pool;
 101         if (queue_id < dev->real_num_tx_queues)
 102                 return dev->_tx[queue_id].pool;
 103
 104         return NULL;
 105 }
 106 EXPORT_SYMBOL(xsk_get_pool_from_qid);
 107
 108 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id)
 109 {
 110         if (queue_id < dev->real_num_rx_queues)
 111                 dev->_rx[queue_id].pool = NULL;
 112         if (queue_id < dev->real_num_tx_queues)
 113                 dev->_tx[queue_id].pool = NULL;
 114 }
 115
 116 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do
 117  * not know if the device has more tx queues than rx, or the opposite.
 118  * This might also change during run time.
 119  */
 120 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool,
 121                         u16 queue_id)
 122 {
 123         if (queue_id >= max_t(unsigned int,
 124                               dev->real_num_rx_queues,
 125                               dev->real_num_tx_queues))
 126                 return -EINVAL;
 127
 128         if (queue_id < dev->real_num_rx_queues)
 129                 dev->_rx[queue_id].pool = pool;
 130         if (queue_id < dev->real_num_tx_queues)
 131                 dev->_tx[queue_id].pool = pool;
 132
 133         return 0;
 134 }
 135
 136 void xp_release(struct xdp_buff_xsk *xskb)
 137 {
 138         xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
 139 }
 140
 141 static u64 xp_get_handle(struct xdp_buff_xsk *xskb)
 142 {
 143         u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
 144
 145         offset += xskb->pool->headroom;
 146         if (!xskb->pool->unaligned)
 147                 return xskb->orig_addr + offset;
 148         return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
 149 }
 150
 151 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
 152 {
 153         struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
 154         u64 addr;
 155         int err;
 156
 157         addr = xp_get_handle(xskb);
 158         err = xskq_prod_reserve_desc(xs->rx, addr, len);
 159         if (err) {
 160                 xs->rx_queue_full++;
 161                 return err;
 162         }
 163
 164         xp_release(xskb);
 165         return 0;
 166 }
 167
 168 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len)
 169 {
 170         void *from_buf, *to_buf;
 171         u32 metalen;
 172
 173         if (unlikely(xdp_data_meta_unsupported(from))) {
 174                 from_buf = from->data;
 175                 to_buf = to->data;
 176                 metalen = 0;
 177         } else {
 178                 from_buf = from->data_meta;
 179                 metalen = from->data - from->data_meta;
 180                 to_buf = to->data - metalen;
 181         }
 182
 183         memcpy(to_buf, from_buf, len + metalen);
 184 }
 185
 186 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len,
 187                      bool explicit_free)
 188 {
 189         struct xdp_buff *xsk_xdp;
 190         int err;
 191
 192         if (len > xsk_pool_get_rx_frame_size(xs->pool)) {
 193                 xs->rx_dropped++;
 194                 return -ENOSPC;
 195         }
 196
 197         xsk_xdp = xsk_buff_alloc(xs->pool);
 198         if (!xsk_xdp) {
 199                 xs->rx_dropped++;
 200                 return -ENOSPC;
 201         }
 202
 203         xsk_copy_xdp(xsk_xdp, xdp, len);
 204         err = __xsk_rcv_zc(xs, xsk_xdp, len);
 205         if (err) {
 206                 xsk_buff_free(xsk_xdp);
 207                 return err;
 208         }
 209         if (explicit_free)
 210                 xdp_return_buff(xdp);
 211         return 0;
 212 }
 213
 214 static bool xsk_is_bound(struct xdp_sock *xs)
 215 {
 216         if (READ_ONCE(xs->state) == XSK_BOUND) {
 217                 /* Matches smp_wmb() in bind(). */
 218                 smp_rmb();
 219                 return true;
 220         }
 221         return false;
 222 }
 223
 224 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp,
 225                    bool explicit_free)
 226 {
 227         u32 len;
 228
 229         if (!xsk_is_bound(xs))
 230                 return -EINVAL;
 231
 232         if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
 233                 return -EINVAL;
 234
 235         len = xdp->data_end - xdp->data;
 236
 237         return xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL ?
 238                 __xsk_rcv_zc(xs, xdp, len) :
 239                 __xsk_rcv(xs, xdp, len, explicit_free);
 240 }
 241
 242 static void xsk_flush(struct xdp_sock *xs)
 243 {
 244         xskq_prod_submit(xs->rx);
 245         __xskq_cons_release(xs->pool->fq);
 246         sock_def_readable(&xs->sk);
 247 }
 248
 249 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 250 {
 251         int err;
 252
 253         spin_lock_bh(&xs->rx_lock);
 254         err = xsk_rcv(xs, xdp, false);
 255         xsk_flush(xs);
 256         spin_unlock_bh(&xs->rx_lock);
 257         return err;
 258 }
 259
 260 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
 261 {
 262         struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
 263         int err;
 264
 265         err = xsk_rcv(xs, xdp, true);
 266         if (err)
 267                 return err;
 268
 269         if (!xs->flush_node.prev)
 270                 list_add(&xs->flush_node, flush_list);
 271
 272         return 0;
 273 }
 274
 275 void __xsk_map_flush(void)
 276 {
 277         struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
 278         struct xdp_sock *xs, *tmp;
 279
 280         list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
 281                 xsk_flush(xs);
 282                 __list_del_clearprev(&xs->flush_node);
 283         }
 284 }
 285
 286 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
 287 {
 288         xskq_prod_submit_n(pool->cq, nb_entries);
 289 }
 290 EXPORT_SYMBOL(xsk_tx_completed);
 291
 292 void xsk_tx_release(struct xsk_buff_pool *pool)
 293 {
 294         struct xdp_sock *xs;
 295
 296         rcu_read_lock();
 297         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
 298                 __xskq_cons_release(xs->tx);
 299                 xs->sk.sk_write_space(&xs->sk);
 300         }
 301         rcu_read_unlock();
 302 }
 303 EXPORT_SYMBOL(xsk_tx_release);
 304
 305 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
 306 {
 307         struct xdp_sock *xs;
 308
 309         rcu_read_lock();
 310         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
 311                 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
 312                         xs->tx->queue_empty_descs++;
 313                         continue;
 314                 }
 315
 316                 /* This is the backpressure mechanism for the Tx path.
 317                  * Reserve space in the completion queue and only proceed
 318                  * if there is space in it. This avoids having to implement
 319                  * any buffering in the Tx path.
 320                  */
 321                 if (xskq_prod_reserve_addr(pool->cq, desc->addr))
 322                         goto out;
 323
 324                 xskq_cons_release(xs->tx);
 325                 rcu_read_unlock();
 326                 return true;
 327         }
 328
 329 out:
 330         rcu_read_unlock();
 331         return false;
 332 }
 333 EXPORT_SYMBOL(xsk_tx_peek_desc);
 334
 335 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
 336 {
 337         struct net_device *dev = xs->dev;
 338         int err;
 339
 340         rcu_read_lock();
 341         err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 342         rcu_read_unlock();
 343
 344         return err;
 345 }
 346
 347 static int xsk_zc_xmit(struct xdp_sock *xs)
 348 {
 349         return xsk_wakeup(xs, XDP_WAKEUP_TX);
 350 }
 351
 352 static void xsk_destruct_skb(struct sk_buff *skb)
 353 {
 354         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 355         struct xdp_sock *xs = xdp_sk(skb->sk);
 356         unsigned long flags;
 357
 358         spin_lock_irqsave(&xs->tx_completion_lock, flags);
 359         xskq_prod_submit_addr(xs->pool->cq, addr);
 360         spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
 361
 362         sock_wfree(skb);
 363 }
 364
 365 static int xsk_generic_xmit(struct sock *sk)
 366 {
 367         struct xdp_sock *xs = xdp_sk(sk);
 368         u32 max_batch = TX_BATCH_SIZE;
 369         bool sent_frame = false;
 370         struct xdp_desc desc;
 371         struct sk_buff *skb;
 372         int err = 0;
 373
 374         mutex_lock(&xs->mutex);
 375
 376         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 377                 goto out;
 378
 379         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 380                 char *buffer;
 381                 u64 addr;
 382                 u32 len;
 383
 384                 if (max_batch-- == 0) {
 385                         err = -EAGAIN;
 386                         goto out;
 387                 }
 388
 389                 len = desc.len;
 390                 skb = sock_alloc_send_skb(sk, len, 1, &err);
 391                 if (unlikely(!skb))
 392                         goto out;
 393
 394                 skb_put(skb, len);
 395                 addr = desc.addr;
 396                 buffer = xsk_buff_raw_get_data(xs->pool, addr);
 397                 err = skb_store_bits(skb, 0, buffer, len);
 398                 /* This is the backpressure mechanism for the Tx path.
 399                  * Reserve space in the completion queue and only proceed
 400                  * if there is space in it. This avoids having to implement
 401                  * any buffering in the Tx path.
 402                  */
 403                 if (unlikely(err) || xskq_prod_reserve(xs->pool->cq)) {
 404                         kfree_skb(skb);
 405                         goto out;
 406                 }
 407
 408                 skb->dev = xs->dev;
 409                 skb->priority = sk->sk_priority;
 410                 skb->mark = sk->sk_mark;
 411                 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
 412                 skb->destructor = xsk_destruct_skb;
 413
 414                 /* Hinder dev_direct_xmit from freeing the packet and
 415                  * therefore completing it in the destructor
 416                  */
 417                 refcount_inc(&skb->users);
 418                 err = dev_direct_xmit(skb, xs->queue_id);
 419                 if  (err == NETDEV_TX_BUSY) {
 420                         /* Tell user-space to retry the send */
 421                         skb->destructor = sock_wfree;
 422                         /* Free skb without triggering the perf drop trace */
 423                         consume_skb(skb);
 424                         err = -EAGAIN;
 425                         goto out;
 426                 }
 427
 428                 xskq_cons_release(xs->tx);
 429                 /* Ignore NET_XMIT_CN as packet might have been sent */
 430                 if (err == NET_XMIT_DROP) {
 431                         /* SKB completed but not sent */
 432                         kfree_skb(skb);
 433                         err = -EBUSY;
 434                         goto out;
 435                 }
 436
 437                 consume_skb(skb);
 438                 sent_frame = true;
 439         }
 440
 441         xs->tx->queue_empty_descs++;
 442
 443 out:
 444         if (sent_frame)
 445                 sk->sk_write_space(sk);
 446
 447         mutex_unlock(&xs->mutex);
 448         return err;
 449 }
 450
 451 static int __xsk_sendmsg(struct sock *sk)
 452 {
 453         struct xdp_sock *xs = xdp_sk(sk);
 454
 455         if (unlikely(!(xs->dev->flags & IFF_UP)))
 456                 return -ENETDOWN;
 457         if (unlikely(!xs->tx))
 458                 return -ENOBUFS;
 459
 460         return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
 461 }
 462
 463 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 464 {
 465         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 466         struct sock *sk = sock->sk;
 467         struct xdp_sock *xs = xdp_sk(sk);
 468
 469         if (unlikely(!xsk_is_bound(xs)))
 470                 return -ENXIO;
 471         if (unlikely(need_wait))
 472                 return -EOPNOTSUPP;
 473
 474         return __xsk_sendmsg(sk);
 475 }
 476
 477 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 478                              struct poll_table_struct *wait)
 479 {
 480         __poll_t mask = datagram_poll(file, sock, wait);
 481         struct sock *sk = sock->sk;
 482         struct xdp_sock *xs = xdp_sk(sk);
 483         struct xsk_buff_pool *pool;
 484
 485         if (unlikely(!xsk_is_bound(xs)))
 486                 return mask;
 487
 488         pool = xs->pool;
 489
 490         if (pool->cached_need_wakeup) {
 491                 if (xs->zc)
 492                         xsk_wakeup(xs, pool->cached_need_wakeup);
 493                 else
 494                         /* Poll needs to drive Tx also in copy mode */
 495                         __xsk_sendmsg(sk);
 496         }
 497
 498         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 499                 mask |= EPOLLIN | EPOLLRDNORM;
 500         if (xs->tx && !xskq_cons_is_full(xs->tx))
 501                 mask |= EPOLLOUT | EPOLLWRNORM;
 502
 503         return mask;
 504 }
 505
 506 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 507                           bool umem_queue)
 508 {
 509         struct xsk_queue *q;
 510
 511         if (entries == 0 || *queue || !is_power_of_2(entries))
 512                 return -EINVAL;
 513
 514         q = xskq_create(entries, umem_queue);
 515         if (!q)
 516                 return -ENOMEM;
 517
 518         /* Make sure queue is ready before it can be seen by others */
 519         smp_wmb();
 520         WRITE_ONCE(*queue, q);
 521         return 0;
 522 }
 523
 524 static void xsk_unbind_dev(struct xdp_sock *xs)
 525 {
 526         struct net_device *dev = xs->dev;
 527
 528         if (xs->state != XSK_BOUND)
 529                 return;
 530         WRITE_ONCE(xs->state, XSK_UNBOUND);
 531
 532         /* Wait for driver to stop using the xdp socket. */
 533         xp_del_xsk(xs->pool, xs);
 534         xs->dev = NULL;
 535         synchronize_net();
 536         dev_put(dev);
 537 }
 538
 539 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 540                                               struct xdp_sock ***map_entry)
 541 {
 542         struct xsk_map *map = NULL;
 543         struct xsk_map_node *node;
 544
 545         *map_entry = NULL;
 546
 547         spin_lock_bh(&xs->map_list_lock);
 548         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 549                                         node);
 550         if (node) {
 551                 WARN_ON(xsk_map_inc(node->map));
 552                 map = node->map;
 553                 *map_entry = node->map_entry;
 554         }
 555         spin_unlock_bh(&xs->map_list_lock);
 556         return map;
 557 }
 558
 559 static void xsk_delete_from_maps(struct xdp_sock *xs)
 560 {
 561         /* This function removes the current XDP socket from all the
 562          * maps it resides in. We need to take extra care here, due to
 563          * the two locks involved. Each map has a lock synchronizing
 564          * updates to the entries, and each socket has a lock that
 565          * synchronizes access to the list of maps (map_list). For
 566          * deadlock avoidance the locks need to be taken in the order
 567          * "map lock"->"socket map list lock". We start off by
 568          * accessing the socket map list, and take a reference to the
 569          * map to guarantee existence between the
 570          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 571          * calls. Then we ask the map to remove the socket, which
 572          * tries to remove the socket from the map. Note that there
 573          * might be updates to the map between
 574          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 575          */
 576         struct xdp_sock **map_entry = NULL;
 577         struct xsk_map *map;
 578
 579         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 580                 xsk_map_try_sock_delete(map, xs, map_entry);
 581                 xsk_map_put(map);
 582         }
 583 }
 584
 585 static int xsk_release(struct socket *sock)
 586 {
 587         struct sock *sk = sock->sk;
 588         struct xdp_sock *xs = xdp_sk(sk);
 589         struct net *net;
 590
 591         if (!sk)
 592                 return 0;
 593
 594         net = sock_net(sk);
 595
 596         mutex_lock(&net->xdp.lock);
 597         sk_del_node_init_rcu(sk);
 598         mutex_unlock(&net->xdp.lock);
 599
 600         local_bh_disable();
 601         sock_prot_inuse_add(net, sk->sk_prot, -1);
 602         local_bh_enable();
 603
 604         xsk_delete_from_maps(xs);
 605         mutex_lock(&xs->mutex);
 606         xsk_unbind_dev(xs);
 607         mutex_unlock(&xs->mutex);
 608
 609         xskq_destroy(xs->rx);
 610         xskq_destroy(xs->tx);
 611         xskq_destroy(xs->fq_tmp);
 612         xskq_destroy(xs->cq_tmp);
 613
 614         sock_orphan(sk);
 615         sock->sk = NULL;
 616
 617         sk_refcnt_debug_release(sk);
 618         sock_put(sk);
 619
 620         return 0;
 621 }
 622
 623 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 624 {
 625         struct socket *sock;
 626         int err;
 627
 628         sock = sockfd_lookup(fd, &err);
 629         if (!sock)
 630                 return ERR_PTR(-ENOTSOCK);
 631
 632         if (sock->sk->sk_family != PF_XDP) {
 633                 sockfd_put(sock);
 634                 return ERR_PTR(-ENOPROTOOPT);
 635         }
 636
 637         return sock;
 638 }
 639
 640 static bool xsk_validate_queues(struct xdp_sock *xs)
 641 {
 642         return xs->fq_tmp && xs->cq_tmp;
 643 }
 644
 645 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 646 {
 647         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 648         struct sock *sk = sock->sk;
 649         struct xdp_sock *xs = xdp_sk(sk);
 650         struct net_device *dev;
 651         u32 flags, qid;
 652         int err = 0;
 653
 654         if (addr_len < sizeof(struct sockaddr_xdp))
 655                 return -EINVAL;
 656         if (sxdp->sxdp_family != AF_XDP)
 657                 return -EINVAL;
 658
 659         flags = sxdp->sxdp_flags;
 660         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 661                       XDP_USE_NEED_WAKEUP))
 662                 return -EINVAL;
 663
 664         rtnl_lock();
 665         mutex_lock(&xs->mutex);
 666         if (xs->state != XSK_READY) {
 667                 err = -EBUSY;
 668                 goto out_release;
 669         }
 670
 671         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 672         if (!dev) {
 673                 err = -ENODEV;
 674                 goto out_release;
 675         }
 676
 677         if (!xs->rx && !xs->tx) {
 678                 err = -EINVAL;
 679                 goto out_unlock;
 680         }
 681
 682         qid = sxdp->sxdp_queue_id;
 683
 684         if (flags & XDP_SHARED_UMEM) {
 685                 struct xdp_sock *umem_xs;
 686                 struct socket *sock;
 687
 688                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 689                     (flags & XDP_USE_NEED_WAKEUP)) {
 690                         /* Cannot specify flags for shared sockets. */
 691                         err = -EINVAL;
 692                         goto out_unlock;
 693                 }
 694
 695                 if (xs->umem) {
 696                         /* We have already our own. */
 697                         err = -EINVAL;
 698                         goto out_unlock;
 699                 }
 700
 701                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 702                 if (IS_ERR(sock)) {
 703                         err = PTR_ERR(sock);
 704                         goto out_unlock;
 705                 }
 706
 707                 umem_xs = xdp_sk(sock->sk);
 708                 if (!xsk_is_bound(umem_xs)) {
 709                         err = -EBADF;
 710                         sockfd_put(sock);
 711                         goto out_unlock;
 712                 }
 713
 714                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 715                         /* Share the umem with another socket on another qid
 716                          * and/or device.
 717                          */
 718                         xs->pool = xp_create_and_assign_umem(xs,
 719                                                              umem_xs->umem);
 720                         if (!xs->pool) {
 721                                 err = -ENOMEM;
 722                                 sockfd_put(sock);
 723                                 goto out_unlock;
 724                         }
 725
 726                         err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
 727                                                    dev, qid);
 728                         if (err) {
 729                                 xp_destroy(xs->pool);
 730                                 xs->pool = NULL;
 731                                 sockfd_put(sock);
 732                                 goto out_unlock;
 733                         }
 734                 } else {
 735                         /* Share the buffer pool with the other socket. */
 736                         if (xs->fq_tmp || xs->cq_tmp) {
 737                                 /* Do not allow setting your own fq or cq. */
 738                                 err = -EINVAL;
 739                                 sockfd_put(sock);
 740                                 goto out_unlock;
 741                         }
 742
 743                         xp_get_pool(umem_xs->pool);
 744                         xs->pool = umem_xs->pool;
 745                 }
 746
 747                 xdp_get_umem(umem_xs->umem);
 748                 WRITE_ONCE(xs->umem, umem_xs->umem);
 749                 sockfd_put(sock);
 750         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 751                 err = -EINVAL;
 752                 goto out_unlock;
 753         } else {
 754                 /* This xsk has its own umem. */
 755                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 756                 if (!xs->pool) {
 757                         err = -ENOMEM;
 758                         goto out_unlock;
 759                 }
 760
 761                 err = xp_assign_dev(xs->pool, dev, qid, flags);
 762                 if (err) {
 763                         xp_destroy(xs->pool);
 764                         xs->pool = NULL;
 765                         goto out_unlock;
 766                 }
 767         }
 768
 769         xs->dev = dev;
 770         xs->zc = xs->umem->zc;
 771         xs->queue_id = qid;
 772         xp_add_xsk(xs->pool, xs);
 773
 774 out_unlock:
 775         if (err) {
 776                 dev_put(dev);
 777         } else {
 778                 /* Matches smp_rmb() in bind() for shared umem
 779                  * sockets, and xsk_is_bound().
 780                  */
 781                 smp_wmb();
 782                 WRITE_ONCE(xs->state, XSK_BOUND);
 783         }
 784 out_release:
 785         mutex_unlock(&xs->mutex);
 786         rtnl_unlock();
 787         return err;
 788 }
 789
 790 struct xdp_umem_reg_v1 {
 791         __u64 addr; /* Start of packet data area */
 792         __u64 len; /* Length of packet data area */
 793         __u32 chunk_size;
 794         __u32 headroom;
 795 };
 796
 797 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 798                           sockptr_t optval, unsigned int optlen)
 799 {
 800         struct sock *sk = sock->sk;
 801         struct xdp_sock *xs = xdp_sk(sk);
 802         int err;
 803
 804         if (level != SOL_XDP)
 805                 return -ENOPROTOOPT;
 806
 807         switch (optname) {
 808         case XDP_RX_RING:
 809         case XDP_TX_RING:
 810         {
 811                 struct xsk_queue **q;
 812                 int entries;
 813
 814                 if (optlen < sizeof(entries))
 815                         return -EINVAL;
 816                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
 817                         return -EFAULT;
 818
 819                 mutex_lock(&xs->mutex);
 820                 if (xs->state != XSK_READY) {
 821                         mutex_unlock(&xs->mutex);
 822                         return -EBUSY;
 823                 }
 824                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
 825                 err = xsk_init_queue(entries, q, false);
 826                 if (!err && optname == XDP_TX_RING)
 827                         /* Tx needs to be explicitly woken up the first time */
 828                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
 829                 mutex_unlock(&xs->mutex);
 830                 return err;
 831         }
 832         case XDP_UMEM_REG:
 833         {
 834                 size_t mr_size = sizeof(struct xdp_umem_reg);
 835                 struct xdp_umem_reg mr = {};
 836                 struct xdp_umem *umem;
 837
 838                 if (optlen < sizeof(struct xdp_umem_reg_v1))
 839                         return -EINVAL;
 840                 else if (optlen < sizeof(mr))
 841                         mr_size = sizeof(struct xdp_umem_reg_v1);
 842
 843                 if (copy_from_sockptr(&mr, optval, mr_size))
 844                         return -EFAULT;
 845
 846                 mutex_lock(&xs->mutex);
 847                 if (xs->state != XSK_READY || xs->umem) {
 848                         mutex_unlock(&xs->mutex);
 849                         return -EBUSY;
 850                 }
 851
 852                 umem = xdp_umem_create(&mr);
 853                 if (IS_ERR(umem)) {
 854                         mutex_unlock(&xs->mutex);
 855                         return PTR_ERR(umem);
 856                 }
 857
 858                 /* Make sure umem is ready before it can be seen by others */
 859                 smp_wmb();
 860                 WRITE_ONCE(xs->umem, umem);
 861                 mutex_unlock(&xs->mutex);
 862                 return 0;
 863         }
 864         case XDP_UMEM_FILL_RING:
 865         case XDP_UMEM_COMPLETION_RING:
 866         {
 867                 struct xsk_queue **q;
 868                 int entries;
 869
 870                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
 871                         return -EFAULT;
 872
 873                 mutex_lock(&xs->mutex);
 874                 if (xs->state != XSK_READY) {
 875                         mutex_unlock(&xs->mutex);
 876                         return -EBUSY;
 877                 }
 878
 879                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
 880                         &xs->cq_tmp;
 881                 err = xsk_init_queue(entries, q, true);
 882                 mutex_unlock(&xs->mutex);
 883                 return err;
 884         }
 885         default:
 886                 break;
 887         }
 888
 889         return -ENOPROTOOPT;
 890 }
 891
 892 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
 893 {
 894         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 895         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 896         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
 897 }
 898
 899 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
 900 {
 901         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 902         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 903         ring->desc = offsetof(struct xdp_umem_ring, desc);
 904 }
 905
 906 struct xdp_statistics_v1 {
 907         __u64 rx_dropped;
 908         __u64 rx_invalid_descs;
 909         __u64 tx_invalid_descs;
 910 };
 911
 912 static int xsk_getsockopt(struct socket *sock, int level, int optname,
 913                           char __user *optval, int __user *optlen)
 914 {
 915         struct sock *sk = sock->sk;
 916         struct xdp_sock *xs = xdp_sk(sk);
 917         int len;
 918
 919         if (level != SOL_XDP)
 920                 return -ENOPROTOOPT;
 921
 922         if (get_user(len, optlen))
 923                 return -EFAULT;
 924         if (len < 0)
 925                 return -EINVAL;
 926
 927         switch (optname) {
 928         case XDP_STATISTICS:
 929         {
 930                 struct xdp_statistics stats = {};
 931                 bool extra_stats = true;
 932                 size_t stats_size;
 933
 934                 if (len < sizeof(struct xdp_statistics_v1)) {
 935                         return -EINVAL;
 936                 } else if (len < sizeof(stats)) {
 937                         extra_stats = false;
 938                         stats_size = sizeof(struct xdp_statistics_v1);
 939                 } else {
 940                         stats_size = sizeof(stats);
 941                 }
 942
 943                 mutex_lock(&xs->mutex);
 944                 stats.rx_dropped = xs->rx_dropped;
 945                 if (extra_stats) {
 946                         stats.rx_ring_full = xs->rx_queue_full;
 947                         stats.rx_fill_ring_empty_descs =
 948                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
 949                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
 950                 } else {
 951                         stats.rx_dropped += xs->rx_queue_full;
 952                 }
 953                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
 954                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
 955                 mutex_unlock(&xs->mutex);
 956
 957                 if (copy_to_user(optval, &stats, stats_size))
 958                         return -EFAULT;
 959                 if (put_user(stats_size, optlen))
 960                         return -EFAULT;
 961
 962                 return 0;
 963         }
 964         case XDP_MMAP_OFFSETS:
 965         {
 966                 struct xdp_mmap_offsets off;
 967                 struct xdp_mmap_offsets_v1 off_v1;
 968                 bool flags_supported = true;
 969                 void *to_copy;
 970
 971                 if (len < sizeof(off_v1))
 972                         return -EINVAL;
 973                 else if (len < sizeof(off))
 974                         flags_supported = false;
 975
 976                 if (flags_supported) {
 977                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
 978                          * except for the flags field added to the end.
 979                          */
 980                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
 981                                                &off.rx);
 982                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
 983                                                &off.tx);
 984                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
 985                                                &off.fr);
 986                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
 987                                                &off.cr);
 988                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
 989                                                 ptrs.flags);
 990                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
 991                                                 ptrs.flags);
 992                         off.fr.flags = offsetof(struct xdp_umem_ring,
 993                                                 ptrs.flags);
 994                         off.cr.flags = offsetof(struct xdp_umem_ring,
 995                                                 ptrs.flags);
 996
 997                         len = sizeof(off);
 998                         to_copy = &off;
 999                 } else {
1000                         xsk_enter_rxtx_offsets(&off_v1.rx);
1001                         xsk_enter_rxtx_offsets(&off_v1.tx);
1002                         xsk_enter_umem_offsets(&off_v1.fr);
1003                         xsk_enter_umem_offsets(&off_v1.cr);
1004
1005                         len = sizeof(off_v1);
1006                         to_copy = &off_v1;
1007                 }
1008
1009                 if (copy_to_user(optval, to_copy, len))
1010                         return -EFAULT;
1011                 if (put_user(len, optlen))
1012                         return -EFAULT;
1013
1014                 return 0;
1015         }
1016         case XDP_OPTIONS:
1017         {
1018                 struct xdp_options opts = {};
1019
1020                 if (len < sizeof(opts))
1021                         return -EINVAL;
1022
1023                 mutex_lock(&xs->mutex);
1024                 if (xs->zc)
1025                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1026                 mutex_unlock(&xs->mutex);
1027
1028                 len = sizeof(opts);
1029                 if (copy_to_user(optval, &opts, len))
1030                         return -EFAULT;
1031                 if (put_user(len, optlen))
1032                         return -EFAULT;
1033
1034                 return 0;
1035         }
1036         default:
1037                 break;
1038         }
1039
1040         return -EOPNOTSUPP;
1041 }
1042
1043 static int xsk_mmap(struct file *file, struct socket *sock,
1044                     struct vm_area_struct *vma)
1045 {
1046         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1047         unsigned long size = vma->vm_end - vma->vm_start;
1048         struct xdp_sock *xs = xdp_sk(sock->sk);
1049         struct xsk_queue *q = NULL;
1050         unsigned long pfn;
1051         struct page *qpg;
1052
1053         if (READ_ONCE(xs->state) != XSK_READY)
1054                 return -EBUSY;
1055
1056         if (offset == XDP_PGOFF_RX_RING) {
1057                 q = READ_ONCE(xs->rx);
1058         } else if (offset == XDP_PGOFF_TX_RING) {
1059                 q = READ_ONCE(xs->tx);
1060         } else {
1061                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1062                 smp_rmb();
1063                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1064                         q = READ_ONCE(xs->fq_tmp);
1065                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1066                         q = READ_ONCE(xs->cq_tmp);
1067         }
1068
1069         if (!q)
1070                 return -EINVAL;
1071
1072         /* Matches the smp_wmb() in xsk_init_queue */
1073         smp_rmb();
1074         qpg = virt_to_head_page(q->ring);
1075         if (size > page_size(qpg))
1076                 return -EINVAL;
1077
1078         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1079         return remap_pfn_range(vma, vma->vm_start, pfn,
1080                                size, vma->vm_page_prot);
1081 }
1082
1083 static int xsk_notifier(struct notifier_block *this,
1084                         unsigned long msg, void *ptr)
1085 {
1086         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1087         struct net *net = dev_net(dev);
1088         struct sock *sk;
1089
1090         switch (msg) {
1091         case NETDEV_UNREGISTER:
1092                 mutex_lock(&net->xdp.lock);
1093                 sk_for_each(sk, &net->xdp.list) {
1094                         struct xdp_sock *xs = xdp_sk(sk);
1095
1096                         mutex_lock(&xs->mutex);
1097                         if (xs->dev == dev) {
1098                                 sk->sk_err = ENETDOWN;
1099                                 if (!sock_flag(sk, SOCK_DEAD))
1100                                         sk->sk_error_report(sk);
1101
1102                                 xsk_unbind_dev(xs);
1103
1104                                 /* Clear device references. */
1105                                 xp_clear_dev(xs->pool);
1106                         }
1107                         mutex_unlock(&xs->mutex);
1108                 }
1109                 mutex_unlock(&net->xdp.lock);
1110                 break;
1111         }
1112         return NOTIFY_DONE;
1113 }
1114
1115 static struct proto xsk_proto = {
1116         .name =         "XDP",
1117         .owner =        THIS_MODULE,
1118         .obj_size =     sizeof(struct xdp_sock),
1119 };
1120
1121 static const struct proto_ops xsk_proto_ops = {
1122         .family         = PF_XDP,
1123         .owner          = THIS_MODULE,
1124         .release        = xsk_release,
1125         .bind           = xsk_bind,
1126         .connect        = sock_no_connect,
1127         .socketpair     = sock_no_socketpair,
1128         .accept         = sock_no_accept,
1129         .getname        = sock_no_getname,
1130         .poll           = xsk_poll,
1131         .ioctl          = sock_no_ioctl,
1132         .listen         = sock_no_listen,
1133         .shutdown       = sock_no_shutdown,
1134         .setsockopt     = xsk_setsockopt,
1135         .getsockopt     = xsk_getsockopt,
1136         .sendmsg        = xsk_sendmsg,
1137         .recvmsg        = sock_no_recvmsg,
1138         .mmap           = xsk_mmap,
1139         .sendpage       = sock_no_sendpage,
1140 };
1141
1142 static void xsk_destruct(struct sock *sk)
1143 {
1144         struct xdp_sock *xs = xdp_sk(sk);
1145
1146         if (!sock_flag(sk, SOCK_DEAD))
1147                 return;
1148
1149         if (!xp_put_pool(xs->pool))
1150                 xdp_put_umem(xs->umem);
1151
1152         sk_refcnt_debug_dec(sk);
1153 }
1154
1155 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1156                       int kern)
1157 {
1158         struct xdp_sock *xs;
1159         struct sock *sk;
1160
1161         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1162                 return -EPERM;
1163         if (sock->type != SOCK_RAW)
1164                 return -ESOCKTNOSUPPORT;
1165
1166         if (protocol)
1167                 return -EPROTONOSUPPORT;
1168
1169         sock->state = SS_UNCONNECTED;
1170
1171         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1172         if (!sk)
1173                 return -ENOBUFS;
1174
1175         sock->ops = &xsk_proto_ops;
1176
1177         sock_init_data(sock, sk);
1178
1179         sk->sk_family = PF_XDP;
1180
1181         sk->sk_destruct = xsk_destruct;
1182         sk_refcnt_debug_inc(sk);
1183
1184         sock_set_flag(sk, SOCK_RCU_FREE);
1185
1186         xs = xdp_sk(sk);
1187         xs->state = XSK_READY;
1188         mutex_init(&xs->mutex);
1189         spin_lock_init(&xs->rx_lock);
1190         spin_lock_init(&xs->tx_completion_lock);
1191
1192         INIT_LIST_HEAD(&xs->map_list);
1193         spin_lock_init(&xs->map_list_lock);
1194
1195         mutex_lock(&net->xdp.lock);
1196         sk_add_node_rcu(sk, &net->xdp.list);
1197         mutex_unlock(&net->xdp.lock);
1198
1199         local_bh_disable();
1200         sock_prot_inuse_add(net, &xsk_proto, 1);
1201         local_bh_enable();
1202
1203         return 0;
1204 }
1205
1206 static const struct net_proto_family xsk_family_ops = {
1207         .family = PF_XDP,
1208         .create = xsk_create,
1209         .owner  = THIS_MODULE,
1210 };
1211
1212 static struct notifier_block xsk_netdev_notifier = {
1213         .notifier_call  = xsk_notifier,
1214 };
1215
1216 static int __net_init xsk_net_init(struct net *net)
1217 {
1218         mutex_init(&net->xdp.lock);
1219         INIT_HLIST_HEAD(&net->xdp.list);
1220         return 0;
1221 }
1222
1223 static void __net_exit xsk_net_exit(struct net *net)
1224 {
1225         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1226 }
1227
1228 static struct pernet_operations xsk_net_ops = {
1229         .init = xsk_net_init,
1230         .exit = xsk_net_exit,
1231 };
1232
1233 static int __init xsk_init(void)
1234 {
1235         int err, cpu;
1236
1237         err = proto_register(&xsk_proto, 0 /* no slab */);
1238         if (err)
1239                 goto out;
1240
1241         err = sock_register(&xsk_family_ops);
1242         if (err)
1243                 goto out_proto;
1244
1245         err = register_pernet_subsys(&xsk_net_ops);
1246         if (err)
1247                 goto out_sk;
1248
1249         err = register_netdevice_notifier(&xsk_netdev_notifier);
1250         if (err)
1251                 goto out_pernet;
1252
1253         for_each_possible_cpu(cpu)
1254                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1255         return 0;
1256
1257 out_pernet:
1258         unregister_pernet_subsys(&xsk_net_ops);
1259 out_sk:
1260         sock_unregister(PF_XDP);
1261 out_proto:
1262         proto_unregister(&xsk_proto);
1263 out:
1264         return err;
1265 }
1266
1267 fs_initcall(xsk_init);