net/xdp/xsk.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* XDP sockets
   3  *
   4  * AF_XDP sockets allows a channel between XDP programs and userspace
   5  * applications.
   6  * Copyright(c) 2018 Intel Corporation.
   7  *
   8  * Author(s): Björn Töpel <[email protected]>
   9  *            Magnus Karlsson <[email protected]>
  10  */
  11
  12 #define pr_fmt(fmt) "AF_XDP: %s: " fmt, __func__
  13
  14 #include <linux/if_xdp.h>
  15 #include <linux/init.h>
  16 #include <linux/sched/mm.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/task.h>
  19 #include <linux/socket.h>
  20 #include <linux/file.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/net.h>
  23 #include <linux/netdevice.h>
  24 #include <linux/rculist.h>
  25 #include <net/xdp_sock_drv.h>
  26 #include <net/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                 err = __dev_direct_xmit(skb, xs->queue_id);
 415                 if  (err == NETDEV_TX_BUSY) {
 416                         /* Tell user-space to retry the send */
 417                         skb->destructor = sock_wfree;
 418                         /* Free skb without triggering the perf drop trace */
 419                         consume_skb(skb);
 420                         err = -EAGAIN;
 421                         goto out;
 422                 }
 423
 424                 xskq_cons_release(xs->tx);
 425                 /* Ignore NET_XMIT_CN as packet might have been sent */
 426                 if (err == NET_XMIT_DROP) {
 427                         /* SKB completed but not sent */
 428                         err = -EBUSY;
 429                         goto out;
 430                 }
 431
 432                 sent_frame = true;
 433         }
 434
 435         xs->tx->queue_empty_descs++;
 436
 437 out:
 438         if (sent_frame)
 439                 sk->sk_write_space(sk);
 440
 441         mutex_unlock(&xs->mutex);
 442         return err;
 443 }
 444
 445 static int __xsk_sendmsg(struct sock *sk)
 446 {
 447         struct xdp_sock *xs = xdp_sk(sk);
 448
 449         if (unlikely(!(xs->dev->flags & IFF_UP)))
 450                 return -ENETDOWN;
 451         if (unlikely(!xs->tx))
 452                 return -ENOBUFS;
 453
 454         return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
 455 }
 456
 457 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 458 {
 459         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 460         struct sock *sk = sock->sk;
 461         struct xdp_sock *xs = xdp_sk(sk);
 462
 463         if (unlikely(!xsk_is_bound(xs)))
 464                 return -ENXIO;
 465         if (unlikely(need_wait))
 466                 return -EOPNOTSUPP;
 467
 468         return __xsk_sendmsg(sk);
 469 }
 470
 471 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 472                              struct poll_table_struct *wait)
 473 {
 474         __poll_t mask = datagram_poll(file, sock, wait);
 475         struct sock *sk = sock->sk;
 476         struct xdp_sock *xs = xdp_sk(sk);
 477         struct xsk_buff_pool *pool;
 478
 479         if (unlikely(!xsk_is_bound(xs)))
 480                 return mask;
 481
 482         pool = xs->pool;
 483
 484         if (pool->cached_need_wakeup) {
 485                 if (xs->zc)
 486                         xsk_wakeup(xs, pool->cached_need_wakeup);
 487                 else
 488                         /* Poll needs to drive Tx also in copy mode */
 489                         __xsk_sendmsg(sk);
 490         }
 491
 492         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 493                 mask |= EPOLLIN | EPOLLRDNORM;
 494         if (xs->tx && !xskq_cons_is_full(xs->tx))
 495                 mask |= EPOLLOUT | EPOLLWRNORM;
 496
 497         return mask;
 498 }
 499
 500 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 501                           bool umem_queue)
 502 {
 503         struct xsk_queue *q;
 504
 505         if (entries == 0 || *queue || !is_power_of_2(entries))
 506                 return -EINVAL;
 507
 508         q = xskq_create(entries, umem_queue);
 509         if (!q)
 510                 return -ENOMEM;
 511
 512         /* Make sure queue is ready before it can be seen by others */
 513         smp_wmb();
 514         WRITE_ONCE(*queue, q);
 515         return 0;
 516 }
 517
 518 static void xsk_unbind_dev(struct xdp_sock *xs)
 519 {
 520         struct net_device *dev = xs->dev;
 521
 522         if (xs->state != XSK_BOUND)
 523                 return;
 524         WRITE_ONCE(xs->state, XSK_UNBOUND);
 525
 526         /* Wait for driver to stop using the xdp socket. */
 527         xp_del_xsk(xs->pool, xs);
 528         xs->dev = NULL;
 529         synchronize_net();
 530         dev_put(dev);
 531 }
 532
 533 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 534                                               struct xdp_sock ***map_entry)
 535 {
 536         struct xsk_map *map = NULL;
 537         struct xsk_map_node *node;
 538
 539         *map_entry = NULL;
 540
 541         spin_lock_bh(&xs->map_list_lock);
 542         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 543                                         node);
 544         if (node) {
 545                 WARN_ON(xsk_map_inc(node->map));
 546                 map = node->map;
 547                 *map_entry = node->map_entry;
 548         }
 549         spin_unlock_bh(&xs->map_list_lock);
 550         return map;
 551 }
 552
 553 static void xsk_delete_from_maps(struct xdp_sock *xs)
 554 {
 555         /* This function removes the current XDP socket from all the
 556          * maps it resides in. We need to take extra care here, due to
 557          * the two locks involved. Each map has a lock synchronizing
 558          * updates to the entries, and each socket has a lock that
 559          * synchronizes access to the list of maps (map_list). For
 560          * deadlock avoidance the locks need to be taken in the order
 561          * "map lock"->"socket map list lock". We start off by
 562          * accessing the socket map list, and take a reference to the
 563          * map to guarantee existence between the
 564          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 565          * calls. Then we ask the map to remove the socket, which
 566          * tries to remove the socket from the map. Note that there
 567          * might be updates to the map between
 568          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 569          */
 570         struct xdp_sock **map_entry = NULL;
 571         struct xsk_map *map;
 572
 573         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 574                 xsk_map_try_sock_delete(map, xs, map_entry);
 575                 xsk_map_put(map);
 576         }
 577 }
 578
 579 static int xsk_release(struct socket *sock)
 580 {
 581         struct sock *sk = sock->sk;
 582         struct xdp_sock *xs = xdp_sk(sk);
 583         struct net *net;
 584
 585         if (!sk)
 586                 return 0;
 587
 588         net = sock_net(sk);
 589
 590         mutex_lock(&net->xdp.lock);
 591         sk_del_node_init_rcu(sk);
 592         mutex_unlock(&net->xdp.lock);
 593
 594         local_bh_disable();
 595         sock_prot_inuse_add(net, sk->sk_prot, -1);
 596         local_bh_enable();
 597
 598         xsk_delete_from_maps(xs);
 599         mutex_lock(&xs->mutex);
 600         xsk_unbind_dev(xs);
 601         mutex_unlock(&xs->mutex);
 602
 603         xskq_destroy(xs->rx);
 604         xskq_destroy(xs->tx);
 605         xskq_destroy(xs->fq_tmp);
 606         xskq_destroy(xs->cq_tmp);
 607
 608         sock_orphan(sk);
 609         sock->sk = NULL;
 610
 611         sk_refcnt_debug_release(sk);
 612         sock_put(sk);
 613
 614         return 0;
 615 }
 616
 617 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 618 {
 619         struct socket *sock;
 620         int err;
 621
 622         sock = sockfd_lookup(fd, &err);
 623         if (!sock)
 624                 return ERR_PTR(-ENOTSOCK);
 625
 626         if (sock->sk->sk_family != PF_XDP) {
 627                 sockfd_put(sock);
 628                 return ERR_PTR(-ENOPROTOOPT);
 629         }
 630
 631         return sock;
 632 }
 633
 634 static bool xsk_validate_queues(struct xdp_sock *xs)
 635 {
 636         return xs->fq_tmp && xs->cq_tmp;
 637 }
 638
 639 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 640 {
 641         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 642         struct sock *sk = sock->sk;
 643         struct xdp_sock *xs = xdp_sk(sk);
 644         struct net_device *dev;
 645         u32 flags, qid;
 646         int err = 0;
 647
 648         if (addr_len < sizeof(struct sockaddr_xdp))
 649                 return -EINVAL;
 650         if (sxdp->sxdp_family != AF_XDP)
 651                 return -EINVAL;
 652
 653         flags = sxdp->sxdp_flags;
 654         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 655                       XDP_USE_NEED_WAKEUP))
 656                 return -EINVAL;
 657
 658         rtnl_lock();
 659         mutex_lock(&xs->mutex);
 660         if (xs->state != XSK_READY) {
 661                 err = -EBUSY;
 662                 goto out_release;
 663         }
 664
 665         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 666         if (!dev) {
 667                 err = -ENODEV;
 668                 goto out_release;
 669         }
 670
 671         if (!xs->rx && !xs->tx) {
 672                 err = -EINVAL;
 673                 goto out_unlock;
 674         }
 675
 676         qid = sxdp->sxdp_queue_id;
 677
 678         if (flags & XDP_SHARED_UMEM) {
 679                 struct xdp_sock *umem_xs;
 680                 struct socket *sock;
 681
 682                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 683                     (flags & XDP_USE_NEED_WAKEUP)) {
 684                         /* Cannot specify flags for shared sockets. */
 685                         err = -EINVAL;
 686                         goto out_unlock;
 687                 }
 688
 689                 if (xs->umem) {
 690                         /* We have already our own. */
 691                         err = -EINVAL;
 692                         goto out_unlock;
 693                 }
 694
 695                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 696                 if (IS_ERR(sock)) {
 697                         err = PTR_ERR(sock);
 698                         goto out_unlock;
 699                 }
 700
 701                 umem_xs = xdp_sk(sock->sk);
 702                 if (!xsk_is_bound(umem_xs)) {
 703                         err = -EBADF;
 704                         sockfd_put(sock);
 705                         goto out_unlock;
 706                 }
 707
 708                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 709                         /* Share the umem with another socket on another qid
 710                          * and/or device.
 711                          */
 712                         xs->pool = xp_create_and_assign_umem(xs,
 713                                                              umem_xs->umem);
 714                         if (!xs->pool) {
 715                                 err = -ENOMEM;
 716                                 sockfd_put(sock);
 717                                 goto out_unlock;
 718                         }
 719
 720                         err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
 721                                                    dev, qid);
 722                         if (err) {
 723                                 xp_destroy(xs->pool);
 724                                 xs->pool = NULL;
 725                                 sockfd_put(sock);
 726                                 goto out_unlock;
 727                         }
 728                 } else {
 729                         /* Share the buffer pool with the other socket. */
 730                         if (xs->fq_tmp || xs->cq_tmp) {
 731                                 /* Do not allow setting your own fq or cq. */
 732                                 err = -EINVAL;
 733                                 sockfd_put(sock);
 734                                 goto out_unlock;
 735                         }
 736
 737                         xp_get_pool(umem_xs->pool);
 738                         xs->pool = umem_xs->pool;
 739                 }
 740
 741                 xdp_get_umem(umem_xs->umem);
 742                 WRITE_ONCE(xs->umem, umem_xs->umem);
 743                 sockfd_put(sock);
 744         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 745                 err = -EINVAL;
 746                 goto out_unlock;
 747         } else {
 748                 /* This xsk has its own umem. */
 749                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 750                 if (!xs->pool) {
 751                         err = -ENOMEM;
 752                         goto out_unlock;
 753                 }
 754
 755                 err = xp_assign_dev(xs->pool, dev, qid, flags);
 756                 if (err) {
 757                         xp_destroy(xs->pool);
 758                         xs->pool = NULL;
 759                         goto out_unlock;
 760                 }
 761         }
 762
 763         xs->dev = dev;
 764         xs->zc = xs->umem->zc;
 765         xs->queue_id = qid;
 766         xp_add_xsk(xs->pool, xs);
 767
 768 out_unlock:
 769         if (err) {
 770                 dev_put(dev);
 771         } else {
 772                 /* Matches smp_rmb() in bind() for shared umem
 773                  * sockets, and xsk_is_bound().
 774                  */
 775                 smp_wmb();
 776                 WRITE_ONCE(xs->state, XSK_BOUND);
 777         }
 778 out_release:
 779         mutex_unlock(&xs->mutex);
 780         rtnl_unlock();
 781         return err;
 782 }
 783
 784 struct xdp_umem_reg_v1 {
 785         __u64 addr; /* Start of packet data area */
 786         __u64 len; /* Length of packet data area */
 787         __u32 chunk_size;
 788         __u32 headroom;
 789 };
 790
 791 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 792                           sockptr_t optval, unsigned int optlen)
 793 {
 794         struct sock *sk = sock->sk;
 795         struct xdp_sock *xs = xdp_sk(sk);
 796         int err;
 797
 798         if (level != SOL_XDP)
 799                 return -ENOPROTOOPT;
 800
 801         switch (optname) {
 802         case XDP_RX_RING:
 803         case XDP_TX_RING:
 804         {
 805                 struct xsk_queue **q;
 806                 int entries;
 807
 808                 if (optlen < sizeof(entries))
 809                         return -EINVAL;
 810                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
 811                         return -EFAULT;
 812
 813                 mutex_lock(&xs->mutex);
 814                 if (xs->state != XSK_READY) {
 815                         mutex_unlock(&xs->mutex);
 816                         return -EBUSY;
 817                 }
 818                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
 819                 err = xsk_init_queue(entries, q, false);
 820                 if (!err && optname == XDP_TX_RING)
 821                         /* Tx needs to be explicitly woken up the first time */
 822                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
 823                 mutex_unlock(&xs->mutex);
 824                 return err;
 825         }
 826         case XDP_UMEM_REG:
 827         {
 828                 size_t mr_size = sizeof(struct xdp_umem_reg);
 829                 struct xdp_umem_reg mr = {};
 830                 struct xdp_umem *umem;
 831
 832                 if (optlen < sizeof(struct xdp_umem_reg_v1))
 833                         return -EINVAL;
 834                 else if (optlen < sizeof(mr))
 835                         mr_size = sizeof(struct xdp_umem_reg_v1);
 836
 837                 if (copy_from_sockptr(&mr, optval, mr_size))
 838                         return -EFAULT;
 839
 840                 mutex_lock(&xs->mutex);
 841                 if (xs->state != XSK_READY || xs->umem) {
 842                         mutex_unlock(&xs->mutex);
 843                         return -EBUSY;
 844                 }
 845
 846                 umem = xdp_umem_create(&mr);
 847                 if (IS_ERR(umem)) {
 848                         mutex_unlock(&xs->mutex);
 849                         return PTR_ERR(umem);
 850                 }
 851
 852                 /* Make sure umem is ready before it can be seen by others */
 853                 smp_wmb();
 854                 WRITE_ONCE(xs->umem, umem);
 855                 mutex_unlock(&xs->mutex);
 856                 return 0;
 857         }
 858         case XDP_UMEM_FILL_RING:
 859         case XDP_UMEM_COMPLETION_RING:
 860         {
 861                 struct xsk_queue **q;
 862                 int entries;
 863
 864                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
 865                         return -EFAULT;
 866
 867                 mutex_lock(&xs->mutex);
 868                 if (xs->state != XSK_READY) {
 869                         mutex_unlock(&xs->mutex);
 870                         return -EBUSY;
 871                 }
 872
 873                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
 874                         &xs->cq_tmp;
 875                 err = xsk_init_queue(entries, q, true);
 876                 mutex_unlock(&xs->mutex);
 877                 return err;
 878         }
 879         default:
 880                 break;
 881         }
 882
 883         return -ENOPROTOOPT;
 884 }
 885
 886 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
 887 {
 888         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 889         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 890         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
 891 }
 892
 893 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
 894 {
 895         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 896         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 897         ring->desc = offsetof(struct xdp_umem_ring, desc);
 898 }
 899
 900 struct xdp_statistics_v1 {
 901         __u64 rx_dropped;
 902         __u64 rx_invalid_descs;
 903         __u64 tx_invalid_descs;
 904 };
 905
 906 static int xsk_getsockopt(struct socket *sock, int level, int optname,
 907                           char __user *optval, int __user *optlen)
 908 {
 909         struct sock *sk = sock->sk;
 910         struct xdp_sock *xs = xdp_sk(sk);
 911         int len;
 912
 913         if (level != SOL_XDP)
 914                 return -ENOPROTOOPT;
 915
 916         if (get_user(len, optlen))
 917                 return -EFAULT;
 918         if (len < 0)
 919                 return -EINVAL;
 920
 921         switch (optname) {
 922         case XDP_STATISTICS:
 923         {
 924                 struct xdp_statistics stats = {};
 925                 bool extra_stats = true;
 926                 size_t stats_size;
 927
 928                 if (len < sizeof(struct xdp_statistics_v1)) {
 929                         return -EINVAL;
 930                 } else if (len < sizeof(stats)) {
 931                         extra_stats = false;
 932                         stats_size = sizeof(struct xdp_statistics_v1);
 933                 } else {
 934                         stats_size = sizeof(stats);
 935                 }
 936
 937                 mutex_lock(&xs->mutex);
 938                 stats.rx_dropped = xs->rx_dropped;
 939                 if (extra_stats) {
 940                         stats.rx_ring_full = xs->rx_queue_full;
 941                         stats.rx_fill_ring_empty_descs =
 942                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
 943                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
 944                 } else {
 945                         stats.rx_dropped += xs->rx_queue_full;
 946                 }
 947                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
 948                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
 949                 mutex_unlock(&xs->mutex);
 950
 951                 if (copy_to_user(optval, &stats, stats_size))
 952                         return -EFAULT;
 953                 if (put_user(stats_size, optlen))
 954                         return -EFAULT;
 955
 956                 return 0;
 957         }
 958         case XDP_MMAP_OFFSETS:
 959         {
 960                 struct xdp_mmap_offsets off;
 961                 struct xdp_mmap_offsets_v1 off_v1;
 962                 bool flags_supported = true;
 963                 void *to_copy;
 964
 965                 if (len < sizeof(off_v1))
 966                         return -EINVAL;
 967                 else if (len < sizeof(off))
 968                         flags_supported = false;
 969
 970                 if (flags_supported) {
 971                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
 972                          * except for the flags field added to the end.
 973                          */
 974                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
 975                                                &off.rx);
 976                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
 977                                                &off.tx);
 978                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
 979                                                &off.fr);
 980                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
 981                                                &off.cr);
 982                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
 983                                                 ptrs.flags);
 984                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
 985                                                 ptrs.flags);
 986                         off.fr.flags = offsetof(struct xdp_umem_ring,
 987                                                 ptrs.flags);
 988                         off.cr.flags = offsetof(struct xdp_umem_ring,
 989                                                 ptrs.flags);
 990
 991                         len = sizeof(off);
 992                         to_copy = &off;
 993                 } else {
 994                         xsk_enter_rxtx_offsets(&off_v1.rx);
 995                         xsk_enter_rxtx_offsets(&off_v1.tx);
 996                         xsk_enter_umem_offsets(&off_v1.fr);
 997                         xsk_enter_umem_offsets(&off_v1.cr);
 998
 999                         len = sizeof(off_v1);
1000                         to_copy = &off_v1;
1001                 }
1002
1003                 if (copy_to_user(optval, to_copy, len))
1004                         return -EFAULT;
1005                 if (put_user(len, optlen))
1006                         return -EFAULT;
1007
1008                 return 0;
1009         }
1010         case XDP_OPTIONS:
1011         {
1012                 struct xdp_options opts = {};
1013
1014                 if (len < sizeof(opts))
1015                         return -EINVAL;
1016
1017                 mutex_lock(&xs->mutex);
1018                 if (xs->zc)
1019                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1020                 mutex_unlock(&xs->mutex);
1021
1022                 len = sizeof(opts);
1023                 if (copy_to_user(optval, &opts, len))
1024                         return -EFAULT;
1025                 if (put_user(len, optlen))
1026                         return -EFAULT;
1027
1028                 return 0;
1029         }
1030         default:
1031                 break;
1032         }
1033
1034         return -EOPNOTSUPP;
1035 }
1036
1037 static int xsk_mmap(struct file *file, struct socket *sock,
1038                     struct vm_area_struct *vma)
1039 {
1040         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1041         unsigned long size = vma->vm_end - vma->vm_start;
1042         struct xdp_sock *xs = xdp_sk(sock->sk);
1043         struct xsk_queue *q = NULL;
1044         unsigned long pfn;
1045         struct page *qpg;
1046
1047         if (READ_ONCE(xs->state) != XSK_READY)
1048                 return -EBUSY;
1049
1050         if (offset == XDP_PGOFF_RX_RING) {
1051                 q = READ_ONCE(xs->rx);
1052         } else if (offset == XDP_PGOFF_TX_RING) {
1053                 q = READ_ONCE(xs->tx);
1054         } else {
1055                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1056                 smp_rmb();
1057                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1058                         q = READ_ONCE(xs->fq_tmp);
1059                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1060                         q = READ_ONCE(xs->cq_tmp);
1061         }
1062
1063         if (!q)
1064                 return -EINVAL;
1065
1066         /* Matches the smp_wmb() in xsk_init_queue */
1067         smp_rmb();
1068         qpg = virt_to_head_page(q->ring);
1069         if (size > page_size(qpg))
1070                 return -EINVAL;
1071
1072         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1073         return remap_pfn_range(vma, vma->vm_start, pfn,
1074                                size, vma->vm_page_prot);
1075 }
1076
1077 static int xsk_notifier(struct notifier_block *this,
1078                         unsigned long msg, void *ptr)
1079 {
1080         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1081         struct net *net = dev_net(dev);
1082         struct sock *sk;
1083
1084         switch (msg) {
1085         case NETDEV_UNREGISTER:
1086                 mutex_lock(&net->xdp.lock);
1087                 sk_for_each(sk, &net->xdp.list) {
1088                         struct xdp_sock *xs = xdp_sk(sk);
1089
1090                         mutex_lock(&xs->mutex);
1091                         if (xs->dev == dev) {
1092                                 sk->sk_err = ENETDOWN;
1093                                 if (!sock_flag(sk, SOCK_DEAD))
1094                                         sk->sk_error_report(sk);
1095
1096                                 xsk_unbind_dev(xs);
1097
1098                                 /* Clear device references. */
1099                                 xp_clear_dev(xs->pool);
1100                         }
1101                         mutex_unlock(&xs->mutex);
1102                 }
1103                 mutex_unlock(&net->xdp.lock);
1104                 break;
1105         }
1106         return NOTIFY_DONE;
1107 }
1108
1109 static struct proto xsk_proto = {
1110         .name =         "XDP",
1111         .owner =        THIS_MODULE,
1112         .obj_size =     sizeof(struct xdp_sock),
1113 };
1114
1115 static const struct proto_ops xsk_proto_ops = {
1116         .family         = PF_XDP,
1117         .owner          = THIS_MODULE,
1118         .release        = xsk_release,
1119         .bind           = xsk_bind,
1120         .connect        = sock_no_connect,
1121         .socketpair     = sock_no_socketpair,
1122         .accept         = sock_no_accept,
1123         .getname        = sock_no_getname,
1124         .poll           = xsk_poll,
1125         .ioctl          = sock_no_ioctl,
1126         .listen         = sock_no_listen,
1127         .shutdown       = sock_no_shutdown,
1128         .setsockopt     = xsk_setsockopt,
1129         .getsockopt     = xsk_getsockopt,
1130         .sendmsg        = xsk_sendmsg,
1131         .recvmsg        = sock_no_recvmsg,
1132         .mmap           = xsk_mmap,
1133         .sendpage       = sock_no_sendpage,
1134 };
1135
1136 static void xsk_destruct(struct sock *sk)
1137 {
1138         struct xdp_sock *xs = xdp_sk(sk);
1139
1140         if (!sock_flag(sk, SOCK_DEAD))
1141                 return;
1142
1143         if (!xp_put_pool(xs->pool))
1144                 xdp_put_umem(xs->umem, !xs->pool);
1145
1146         sk_refcnt_debug_dec(sk);
1147 }
1148
1149 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1150                       int kern)
1151 {
1152         struct xdp_sock *xs;
1153         struct sock *sk;
1154
1155         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1156                 return -EPERM;
1157         if (sock->type != SOCK_RAW)
1158                 return -ESOCKTNOSUPPORT;
1159
1160         if (protocol)
1161                 return -EPROTONOSUPPORT;
1162
1163         sock->state = SS_UNCONNECTED;
1164
1165         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1166         if (!sk)
1167                 return -ENOBUFS;
1168
1169         sock->ops = &xsk_proto_ops;
1170
1171         sock_init_data(sock, sk);
1172
1173         sk->sk_family = PF_XDP;
1174
1175         sk->sk_destruct = xsk_destruct;
1176         sk_refcnt_debug_inc(sk);
1177
1178         sock_set_flag(sk, SOCK_RCU_FREE);
1179
1180         xs = xdp_sk(sk);
1181         xs->state = XSK_READY;
1182         mutex_init(&xs->mutex);
1183         spin_lock_init(&xs->rx_lock);
1184         spin_lock_init(&xs->tx_completion_lock);
1185
1186         INIT_LIST_HEAD(&xs->map_list);
1187         spin_lock_init(&xs->map_list_lock);
1188
1189         mutex_lock(&net->xdp.lock);
1190         sk_add_node_rcu(sk, &net->xdp.list);
1191         mutex_unlock(&net->xdp.lock);
1192
1193         local_bh_disable();
1194         sock_prot_inuse_add(net, &xsk_proto, 1);
1195         local_bh_enable();
1196
1197         return 0;
1198 }
1199
1200 static const struct net_proto_family xsk_family_ops = {
1201         .family = PF_XDP,
1202         .create = xsk_create,
1203         .owner  = THIS_MODULE,
1204 };
1205
1206 static struct notifier_block xsk_netdev_notifier = {
1207         .notifier_call  = xsk_notifier,
1208 };
1209
1210 static int __net_init xsk_net_init(struct net *net)
1211 {
1212         mutex_init(&net->xdp.lock);
1213         INIT_HLIST_HEAD(&net->xdp.list);
1214         return 0;
1215 }
1216
1217 static void __net_exit xsk_net_exit(struct net *net)
1218 {
1219         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1220 }
1221
1222 static struct pernet_operations xsk_net_ops = {
1223         .init = xsk_net_init,
1224         .exit = xsk_net_exit,
1225 };
1226
1227 static int __init xsk_init(void)
1228 {
1229         int err, cpu;
1230
1231         err = proto_register(&xsk_proto, 0 /* no slab */);
1232         if (err)
1233                 goto out;
1234
1235         err = sock_register(&xsk_family_ops);
1236         if (err)
1237                 goto out_proto;
1238
1239         err = register_pernet_subsys(&xsk_net_ops);
1240         if (err)
1241                 goto out_sk;
1242
1243         err = register_netdevice_notifier(&xsk_netdev_notifier);
1244         if (err)
1245                 goto out_pernet;
1246
1247         for_each_possible_cpu(cpu)
1248                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1249         return 0;
1250
1251 out_pernet:
1252         unregister_pernet_subsys(&xsk_net_ops);
1253 out_sk:
1254         sock_unregister(PF_XDP);
1255 out_proto:
1256         proto_unregister(&xsk_proto);
1257 out:
1258         return err;
1259 }
1260
1261 fs_initcall(xsk_init);