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_tx_writeable(struct xdp_sock *xs)
 215 {
 216         if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
 217                 return false;
 218
 219         return true;
 220 }
 221
 222 static bool xsk_is_bound(struct xdp_sock *xs)
 223 {
 224         if (READ_ONCE(xs->state) == XSK_BOUND) {
 225                 /* Matches smp_wmb() in bind(). */
 226                 smp_rmb();
 227                 return true;
 228         }
 229         return false;
 230 }
 231
 232 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp,
 233                    bool explicit_free)
 234 {
 235         u32 len;
 236
 237         if (!xsk_is_bound(xs))
 238                 return -EINVAL;
 239
 240         if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
 241                 return -EINVAL;
 242
 243         len = xdp->data_end - xdp->data;
 244
 245         return xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL ?
 246                 __xsk_rcv_zc(xs, xdp, len) :
 247                 __xsk_rcv(xs, xdp, len, explicit_free);
 248 }
 249
 250 static void xsk_flush(struct xdp_sock *xs)
 251 {
 252         xskq_prod_submit(xs->rx);
 253         __xskq_cons_release(xs->pool->fq);
 254         sock_def_readable(&xs->sk);
 255 }
 256
 257 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 258 {
 259         int err;
 260
 261         spin_lock_bh(&xs->rx_lock);
 262         err = xsk_rcv(xs, xdp, false);
 263         xsk_flush(xs);
 264         spin_unlock_bh(&xs->rx_lock);
 265         return err;
 266 }
 267
 268 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
 269 {
 270         struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
 271         int err;
 272
 273         err = xsk_rcv(xs, xdp, true);
 274         if (err)
 275                 return err;
 276
 277         if (!xs->flush_node.prev)
 278                 list_add(&xs->flush_node, flush_list);
 279
 280         return 0;
 281 }
 282
 283 void __xsk_map_flush(void)
 284 {
 285         struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
 286         struct xdp_sock *xs, *tmp;
 287
 288         list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
 289                 xsk_flush(xs);
 290                 __list_del_clearprev(&xs->flush_node);
 291         }
 292 }
 293
 294 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
 295 {
 296         xskq_prod_submit_n(pool->cq, nb_entries);
 297 }
 298 EXPORT_SYMBOL(xsk_tx_completed);
 299
 300 void xsk_tx_release(struct xsk_buff_pool *pool)
 301 {
 302         struct xdp_sock *xs;
 303
 304         rcu_read_lock();
 305         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
 306                 __xskq_cons_release(xs->tx);
 307                 if (xsk_tx_writeable(xs))
 308                         xs->sk.sk_write_space(&xs->sk);
 309         }
 310         rcu_read_unlock();
 311 }
 312 EXPORT_SYMBOL(xsk_tx_release);
 313
 314 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
 315 {
 316         struct xdp_sock *xs;
 317
 318         rcu_read_lock();
 319         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
 320                 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
 321                         xs->tx->queue_empty_descs++;
 322                         continue;
 323                 }
 324
 325                 /* This is the backpressure mechanism for the Tx path.
 326                  * Reserve space in the completion queue and only proceed
 327                  * if there is space in it. This avoids having to implement
 328                  * any buffering in the Tx path.
 329                  */
 330                 if (xskq_prod_reserve_addr(pool->cq, desc->addr))
 331                         goto out;
 332
 333                 xskq_cons_release(xs->tx);
 334                 rcu_read_unlock();
 335                 return true;
 336         }
 337
 338 out:
 339         rcu_read_unlock();
 340         return false;
 341 }
 342 EXPORT_SYMBOL(xsk_tx_peek_desc);
 343
 344 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
 345 {
 346         struct net_device *dev = xs->dev;
 347         int err;
 348
 349         rcu_read_lock();
 350         err = dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 351         rcu_read_unlock();
 352
 353         return err;
 354 }
 355
 356 static int xsk_zc_xmit(struct xdp_sock *xs)
 357 {
 358         return xsk_wakeup(xs, XDP_WAKEUP_TX);
 359 }
 360
 361 static void xsk_destruct_skb(struct sk_buff *skb)
 362 {
 363         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 364         struct xdp_sock *xs = xdp_sk(skb->sk);
 365         unsigned long flags;
 366
 367         spin_lock_irqsave(&xs->tx_completion_lock, flags);
 368         xskq_prod_submit_addr(xs->pool->cq, addr);
 369         spin_unlock_irqrestore(&xs->tx_completion_lock, flags);
 370
 371         sock_wfree(skb);
 372 }
 373
 374 static int xsk_generic_xmit(struct sock *sk)
 375 {
 376         struct xdp_sock *xs = xdp_sk(sk);
 377         u32 max_batch = TX_BATCH_SIZE;
 378         bool sent_frame = false;
 379         struct xdp_desc desc;
 380         struct sk_buff *skb;
 381         int err = 0;
 382
 383         mutex_lock(&xs->mutex);
 384
 385         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 386                 goto out;
 387
 388         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 389                 char *buffer;
 390                 u64 addr;
 391                 u32 len;
 392
 393                 if (max_batch-- == 0) {
 394                         err = -EAGAIN;
 395                         goto out;
 396                 }
 397
 398                 len = desc.len;
 399                 skb = sock_alloc_send_skb(sk, len, 1, &err);
 400                 if (unlikely(!skb))
 401                         goto out;
 402
 403                 skb_put(skb, len);
 404                 addr = desc.addr;
 405                 buffer = xsk_buff_raw_get_data(xs->pool, addr);
 406                 err = skb_store_bits(skb, 0, buffer, len);
 407                 /* This is the backpressure mechanism for the Tx path.
 408                  * Reserve space in the completion queue and only proceed
 409                  * if there is space in it. This avoids having to implement
 410                  * any buffering in the Tx path.
 411                  */
 412                 if (unlikely(err) || xskq_prod_reserve(xs->pool->cq)) {
 413                         kfree_skb(skb);
 414                         goto out;
 415                 }
 416
 417                 skb->dev = xs->dev;
 418                 skb->priority = sk->sk_priority;
 419                 skb->mark = sk->sk_mark;
 420                 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
 421                 skb->destructor = xsk_destruct_skb;
 422
 423                 err = __dev_direct_xmit(skb, xs->queue_id);
 424                 if  (err == NETDEV_TX_BUSY) {
 425                         /* Tell user-space to retry the send */
 426                         skb->destructor = sock_wfree;
 427                         /* Free skb without triggering the perf drop trace */
 428                         consume_skb(skb);
 429                         err = -EAGAIN;
 430                         goto out;
 431                 }
 432
 433                 xskq_cons_release(xs->tx);
 434                 /* Ignore NET_XMIT_CN as packet might have been sent */
 435                 if (err == NET_XMIT_DROP) {
 436                         /* SKB completed but not sent */
 437                         err = -EBUSY;
 438                         goto out;
 439                 }
 440
 441                 sent_frame = true;
 442         }
 443
 444         xs->tx->queue_empty_descs++;
 445
 446 out:
 447         if (sent_frame)
 448                 if (xsk_tx_writeable(xs))
 449                         sk->sk_write_space(sk);
 450
 451         mutex_unlock(&xs->mutex);
 452         return err;
 453 }
 454
 455 static int __xsk_sendmsg(struct sock *sk)
 456 {
 457         struct xdp_sock *xs = xdp_sk(sk);
 458
 459         if (unlikely(!(xs->dev->flags & IFF_UP)))
 460                 return -ENETDOWN;
 461         if (unlikely(!xs->tx))
 462                 return -ENOBUFS;
 463
 464         return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
 465 }
 466
 467 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 468 {
 469         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 470         struct sock *sk = sock->sk;
 471         struct xdp_sock *xs = xdp_sk(sk);
 472
 473         if (unlikely(!xsk_is_bound(xs)))
 474                 return -ENXIO;
 475         if (unlikely(need_wait))
 476                 return -EOPNOTSUPP;
 477
 478         return __xsk_sendmsg(sk);
 479 }
 480
 481 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 482                              struct poll_table_struct *wait)
 483 {
 484         __poll_t mask = 0;
 485         struct sock *sk = sock->sk;
 486         struct xdp_sock *xs = xdp_sk(sk);
 487         struct xsk_buff_pool *pool;
 488
 489         sock_poll_wait(file, sock, wait);
 490
 491         if (unlikely(!xsk_is_bound(xs)))
 492                 return mask;
 493
 494         pool = xs->pool;
 495
 496         if (pool->cached_need_wakeup) {
 497                 if (xs->zc)
 498                         xsk_wakeup(xs, pool->cached_need_wakeup);
 499                 else
 500                         /* Poll needs to drive Tx also in copy mode */
 501                         __xsk_sendmsg(sk);
 502         }
 503
 504         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 505                 mask |= EPOLLIN | EPOLLRDNORM;
 506         if (xs->tx && xsk_tx_writeable(xs))
 507                 mask |= EPOLLOUT | EPOLLWRNORM;
 508
 509         return mask;
 510 }
 511
 512 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 513                           bool umem_queue)
 514 {
 515         struct xsk_queue *q;
 516
 517         if (entries == 0 || *queue || !is_power_of_2(entries))
 518                 return -EINVAL;
 519
 520         q = xskq_create(entries, umem_queue);
 521         if (!q)
 522                 return -ENOMEM;
 523
 524         /* Make sure queue is ready before it can be seen by others */
 525         smp_wmb();
 526         WRITE_ONCE(*queue, q);
 527         return 0;
 528 }
 529
 530 static void xsk_unbind_dev(struct xdp_sock *xs)
 531 {
 532         struct net_device *dev = xs->dev;
 533
 534         if (xs->state != XSK_BOUND)
 535                 return;
 536         WRITE_ONCE(xs->state, XSK_UNBOUND);
 537
 538         /* Wait for driver to stop using the xdp socket. */
 539         xp_del_xsk(xs->pool, xs);
 540         xs->dev = NULL;
 541         synchronize_net();
 542         dev_put(dev);
 543 }
 544
 545 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 546                                               struct xdp_sock ***map_entry)
 547 {
 548         struct xsk_map *map = NULL;
 549         struct xsk_map_node *node;
 550
 551         *map_entry = NULL;
 552
 553         spin_lock_bh(&xs->map_list_lock);
 554         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 555                                         node);
 556         if (node) {
 557                 WARN_ON(xsk_map_inc(node->map));
 558                 map = node->map;
 559                 *map_entry = node->map_entry;
 560         }
 561         spin_unlock_bh(&xs->map_list_lock);
 562         return map;
 563 }
 564
 565 static void xsk_delete_from_maps(struct xdp_sock *xs)
 566 {
 567         /* This function removes the current XDP socket from all the
 568          * maps it resides in. We need to take extra care here, due to
 569          * the two locks involved. Each map has a lock synchronizing
 570          * updates to the entries, and each socket has a lock that
 571          * synchronizes access to the list of maps (map_list). For
 572          * deadlock avoidance the locks need to be taken in the order
 573          * "map lock"->"socket map list lock". We start off by
 574          * accessing the socket map list, and take a reference to the
 575          * map to guarantee existence between the
 576          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 577          * calls. Then we ask the map to remove the socket, which
 578          * tries to remove the socket from the map. Note that there
 579          * might be updates to the map between
 580          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 581          */
 582         struct xdp_sock **map_entry = NULL;
 583         struct xsk_map *map;
 584
 585         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 586                 xsk_map_try_sock_delete(map, xs, map_entry);
 587                 xsk_map_put(map);
 588         }
 589 }
 590
 591 static int xsk_release(struct socket *sock)
 592 {
 593         struct sock *sk = sock->sk;
 594         struct xdp_sock *xs = xdp_sk(sk);
 595         struct net *net;
 596
 597         if (!sk)
 598                 return 0;
 599
 600         net = sock_net(sk);
 601
 602         mutex_lock(&net->xdp.lock);
 603         sk_del_node_init_rcu(sk);
 604         mutex_unlock(&net->xdp.lock);
 605
 606         local_bh_disable();
 607         sock_prot_inuse_add(net, sk->sk_prot, -1);
 608         local_bh_enable();
 609
 610         xsk_delete_from_maps(xs);
 611         mutex_lock(&xs->mutex);
 612         xsk_unbind_dev(xs);
 613         mutex_unlock(&xs->mutex);
 614
 615         xskq_destroy(xs->rx);
 616         xskq_destroy(xs->tx);
 617         xskq_destroy(xs->fq_tmp);
 618         xskq_destroy(xs->cq_tmp);
 619
 620         sock_orphan(sk);
 621         sock->sk = NULL;
 622
 623         sk_refcnt_debug_release(sk);
 624         sock_put(sk);
 625
 626         return 0;
 627 }
 628
 629 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 630 {
 631         struct socket *sock;
 632         int err;
 633
 634         sock = sockfd_lookup(fd, &err);
 635         if (!sock)
 636                 return ERR_PTR(-ENOTSOCK);
 637
 638         if (sock->sk->sk_family != PF_XDP) {
 639                 sockfd_put(sock);
 640                 return ERR_PTR(-ENOPROTOOPT);
 641         }
 642
 643         return sock;
 644 }
 645
 646 static bool xsk_validate_queues(struct xdp_sock *xs)
 647 {
 648         return xs->fq_tmp && xs->cq_tmp;
 649 }
 650
 651 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 652 {
 653         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 654         struct sock *sk = sock->sk;
 655         struct xdp_sock *xs = xdp_sk(sk);
 656         struct net_device *dev;
 657         u32 flags, qid;
 658         int err = 0;
 659
 660         if (addr_len < sizeof(struct sockaddr_xdp))
 661                 return -EINVAL;
 662         if (sxdp->sxdp_family != AF_XDP)
 663                 return -EINVAL;
 664
 665         flags = sxdp->sxdp_flags;
 666         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 667                       XDP_USE_NEED_WAKEUP))
 668                 return -EINVAL;
 669
 670         rtnl_lock();
 671         mutex_lock(&xs->mutex);
 672         if (xs->state != XSK_READY) {
 673                 err = -EBUSY;
 674                 goto out_release;
 675         }
 676
 677         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 678         if (!dev) {
 679                 err = -ENODEV;
 680                 goto out_release;
 681         }
 682
 683         if (!xs->rx && !xs->tx) {
 684                 err = -EINVAL;
 685                 goto out_unlock;
 686         }
 687
 688         qid = sxdp->sxdp_queue_id;
 689
 690         if (flags & XDP_SHARED_UMEM) {
 691                 struct xdp_sock *umem_xs;
 692                 struct socket *sock;
 693
 694                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 695                     (flags & XDP_USE_NEED_WAKEUP)) {
 696                         /* Cannot specify flags for shared sockets. */
 697                         err = -EINVAL;
 698                         goto out_unlock;
 699                 }
 700
 701                 if (xs->umem) {
 702                         /* We have already our own. */
 703                         err = -EINVAL;
 704                         goto out_unlock;
 705                 }
 706
 707                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 708                 if (IS_ERR(sock)) {
 709                         err = PTR_ERR(sock);
 710                         goto out_unlock;
 711                 }
 712
 713                 umem_xs = xdp_sk(sock->sk);
 714                 if (!xsk_is_bound(umem_xs)) {
 715                         err = -EBADF;
 716                         sockfd_put(sock);
 717                         goto out_unlock;
 718                 }
 719
 720                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 721                         /* Share the umem with another socket on another qid
 722                          * and/or device.
 723                          */
 724                         xs->pool = xp_create_and_assign_umem(xs,
 725                                                              umem_xs->umem);
 726                         if (!xs->pool) {
 727                                 err = -ENOMEM;
 728                                 sockfd_put(sock);
 729                                 goto out_unlock;
 730                         }
 731
 732                         err = xp_assign_dev_shared(xs->pool, umem_xs->umem,
 733                                                    dev, qid);
 734                         if (err) {
 735                                 xp_destroy(xs->pool);
 736                                 xs->pool = NULL;
 737                                 sockfd_put(sock);
 738                                 goto out_unlock;
 739                         }
 740                 } else {
 741                         /* Share the buffer pool with the other socket. */
 742                         if (xs->fq_tmp || xs->cq_tmp) {
 743                                 /* Do not allow setting your own fq or cq. */
 744                                 err = -EINVAL;
 745                                 sockfd_put(sock);
 746                                 goto out_unlock;
 747                         }
 748
 749                         xp_get_pool(umem_xs->pool);
 750                         xs->pool = umem_xs->pool;
 751                 }
 752
 753                 xdp_get_umem(umem_xs->umem);
 754                 WRITE_ONCE(xs->umem, umem_xs->umem);
 755                 sockfd_put(sock);
 756         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 757                 err = -EINVAL;
 758                 goto out_unlock;
 759         } else {
 760                 /* This xsk has its own umem. */
 761                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 762                 if (!xs->pool) {
 763                         err = -ENOMEM;
 764                         goto out_unlock;
 765                 }
 766
 767                 err = xp_assign_dev(xs->pool, dev, qid, flags);
 768                 if (err) {
 769                         xp_destroy(xs->pool);
 770                         xs->pool = NULL;
 771                         goto out_unlock;
 772                 }
 773         }
 774
 775         xs->dev = dev;
 776         xs->zc = xs->umem->zc;
 777         xs->queue_id = qid;
 778         xp_add_xsk(xs->pool, xs);
 779
 780 out_unlock:
 781         if (err) {
 782                 dev_put(dev);
 783         } else {
 784                 /* Matches smp_rmb() in bind() for shared umem
 785                  * sockets, and xsk_is_bound().
 786                  */
 787                 smp_wmb();
 788                 WRITE_ONCE(xs->state, XSK_BOUND);
 789         }
 790 out_release:
 791         mutex_unlock(&xs->mutex);
 792         rtnl_unlock();
 793         return err;
 794 }
 795
 796 struct xdp_umem_reg_v1 {
 797         __u64 addr; /* Start of packet data area */
 798         __u64 len; /* Length of packet data area */
 799         __u32 chunk_size;
 800         __u32 headroom;
 801 };
 802
 803 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 804                           sockptr_t optval, unsigned int optlen)
 805 {
 806         struct sock *sk = sock->sk;
 807         struct xdp_sock *xs = xdp_sk(sk);
 808         int err;
 809
 810         if (level != SOL_XDP)
 811                 return -ENOPROTOOPT;
 812
 813         switch (optname) {
 814         case XDP_RX_RING:
 815         case XDP_TX_RING:
 816         {
 817                 struct xsk_queue **q;
 818                 int entries;
 819
 820                 if (optlen < sizeof(entries))
 821                         return -EINVAL;
 822                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
 823                         return -EFAULT;
 824
 825                 mutex_lock(&xs->mutex);
 826                 if (xs->state != XSK_READY) {
 827                         mutex_unlock(&xs->mutex);
 828                         return -EBUSY;
 829                 }
 830                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
 831                 err = xsk_init_queue(entries, q, false);
 832                 if (!err && optname == XDP_TX_RING)
 833                         /* Tx needs to be explicitly woken up the first time */
 834                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
 835                 mutex_unlock(&xs->mutex);
 836                 return err;
 837         }
 838         case XDP_UMEM_REG:
 839         {
 840                 size_t mr_size = sizeof(struct xdp_umem_reg);
 841                 struct xdp_umem_reg mr = {};
 842                 struct xdp_umem *umem;
 843
 844                 if (optlen < sizeof(struct xdp_umem_reg_v1))
 845                         return -EINVAL;
 846                 else if (optlen < sizeof(mr))
 847                         mr_size = sizeof(struct xdp_umem_reg_v1);
 848
 849                 if (copy_from_sockptr(&mr, optval, mr_size))
 850                         return -EFAULT;
 851
 852                 mutex_lock(&xs->mutex);
 853                 if (xs->state != XSK_READY || xs->umem) {
 854                         mutex_unlock(&xs->mutex);
 855                         return -EBUSY;
 856                 }
 857
 858                 umem = xdp_umem_create(&mr);
 859                 if (IS_ERR(umem)) {
 860                         mutex_unlock(&xs->mutex);
 861                         return PTR_ERR(umem);
 862                 }
 863
 864                 /* Make sure umem is ready before it can be seen by others */
 865                 smp_wmb();
 866                 WRITE_ONCE(xs->umem, umem);
 867                 mutex_unlock(&xs->mutex);
 868                 return 0;
 869         }
 870         case XDP_UMEM_FILL_RING:
 871         case XDP_UMEM_COMPLETION_RING:
 872         {
 873                 struct xsk_queue **q;
 874                 int entries;
 875
 876                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
 877                         return -EFAULT;
 878
 879                 mutex_lock(&xs->mutex);
 880                 if (xs->state != XSK_READY) {
 881                         mutex_unlock(&xs->mutex);
 882                         return -EBUSY;
 883                 }
 884
 885                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
 886                         &xs->cq_tmp;
 887                 err = xsk_init_queue(entries, q, true);
 888                 mutex_unlock(&xs->mutex);
 889                 return err;
 890         }
 891         default:
 892                 break;
 893         }
 894
 895         return -ENOPROTOOPT;
 896 }
 897
 898 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
 899 {
 900         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 901         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 902         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
 903 }
 904
 905 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
 906 {
 907         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 908         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 909         ring->desc = offsetof(struct xdp_umem_ring, desc);
 910 }
 911
 912 struct xdp_statistics_v1 {
 913         __u64 rx_dropped;
 914         __u64 rx_invalid_descs;
 915         __u64 tx_invalid_descs;
 916 };
 917
 918 static int xsk_getsockopt(struct socket *sock, int level, int optname,
 919                           char __user *optval, int __user *optlen)
 920 {
 921         struct sock *sk = sock->sk;
 922         struct xdp_sock *xs = xdp_sk(sk);
 923         int len;
 924
 925         if (level != SOL_XDP)
 926                 return -ENOPROTOOPT;
 927
 928         if (get_user(len, optlen))
 929                 return -EFAULT;
 930         if (len < 0)
 931                 return -EINVAL;
 932
 933         switch (optname) {
 934         case XDP_STATISTICS:
 935         {
 936                 struct xdp_statistics stats = {};
 937                 bool extra_stats = true;
 938                 size_t stats_size;
 939
 940                 if (len < sizeof(struct xdp_statistics_v1)) {
 941                         return -EINVAL;
 942                 } else if (len < sizeof(stats)) {
 943                         extra_stats = false;
 944                         stats_size = sizeof(struct xdp_statistics_v1);
 945                 } else {
 946                         stats_size = sizeof(stats);
 947                 }
 948
 949                 mutex_lock(&xs->mutex);
 950                 stats.rx_dropped = xs->rx_dropped;
 951                 if (extra_stats) {
 952                         stats.rx_ring_full = xs->rx_queue_full;
 953                         stats.rx_fill_ring_empty_descs =
 954                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
 955                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
 956                 } else {
 957                         stats.rx_dropped += xs->rx_queue_full;
 958                 }
 959                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
 960                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
 961                 mutex_unlock(&xs->mutex);
 962
 963                 if (copy_to_user(optval, &stats, stats_size))
 964                         return -EFAULT;
 965                 if (put_user(stats_size, optlen))
 966                         return -EFAULT;
 967
 968                 return 0;
 969         }
 970         case XDP_MMAP_OFFSETS:
 971         {
 972                 struct xdp_mmap_offsets off;
 973                 struct xdp_mmap_offsets_v1 off_v1;
 974                 bool flags_supported = true;
 975                 void *to_copy;
 976
 977                 if (len < sizeof(off_v1))
 978                         return -EINVAL;
 979                 else if (len < sizeof(off))
 980                         flags_supported = false;
 981
 982                 if (flags_supported) {
 983                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
 984                          * except for the flags field added to the end.
 985                          */
 986                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
 987                                                &off.rx);
 988                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
 989                                                &off.tx);
 990                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
 991                                                &off.fr);
 992                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
 993                                                &off.cr);
 994                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
 995                                                 ptrs.flags);
 996                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
 997                                                 ptrs.flags);
 998                         off.fr.flags = offsetof(struct xdp_umem_ring,
 999                                                 ptrs.flags);
1000                         off.cr.flags = offsetof(struct xdp_umem_ring,
1001                                                 ptrs.flags);
1002
1003                         len = sizeof(off);
1004                         to_copy = &off;
1005                 } else {
1006                         xsk_enter_rxtx_offsets(&off_v1.rx);
1007                         xsk_enter_rxtx_offsets(&off_v1.tx);
1008                         xsk_enter_umem_offsets(&off_v1.fr);
1009                         xsk_enter_umem_offsets(&off_v1.cr);
1010
1011                         len = sizeof(off_v1);
1012                         to_copy = &off_v1;
1013                 }
1014
1015                 if (copy_to_user(optval, to_copy, len))
1016                         return -EFAULT;
1017                 if (put_user(len, optlen))
1018                         return -EFAULT;
1019
1020                 return 0;
1021         }
1022         case XDP_OPTIONS:
1023         {
1024                 struct xdp_options opts = {};
1025
1026                 if (len < sizeof(opts))
1027                         return -EINVAL;
1028
1029                 mutex_lock(&xs->mutex);
1030                 if (xs->zc)
1031                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1032                 mutex_unlock(&xs->mutex);
1033
1034                 len = sizeof(opts);
1035                 if (copy_to_user(optval, &opts, len))
1036                         return -EFAULT;
1037                 if (put_user(len, optlen))
1038                         return -EFAULT;
1039
1040                 return 0;
1041         }
1042         default:
1043                 break;
1044         }
1045
1046         return -EOPNOTSUPP;
1047 }
1048
1049 static int xsk_mmap(struct file *file, struct socket *sock,
1050                     struct vm_area_struct *vma)
1051 {
1052         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1053         unsigned long size = vma->vm_end - vma->vm_start;
1054         struct xdp_sock *xs = xdp_sk(sock->sk);
1055         struct xsk_queue *q = NULL;
1056         unsigned long pfn;
1057         struct page *qpg;
1058
1059         if (READ_ONCE(xs->state) != XSK_READY)
1060                 return -EBUSY;
1061
1062         if (offset == XDP_PGOFF_RX_RING) {
1063                 q = READ_ONCE(xs->rx);
1064         } else if (offset == XDP_PGOFF_TX_RING) {
1065                 q = READ_ONCE(xs->tx);
1066         } else {
1067                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1068                 smp_rmb();
1069                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1070                         q = READ_ONCE(xs->fq_tmp);
1071                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1072                         q = READ_ONCE(xs->cq_tmp);
1073         }
1074
1075         if (!q)
1076                 return -EINVAL;
1077
1078         /* Matches the smp_wmb() in xsk_init_queue */
1079         smp_rmb();
1080         qpg = virt_to_head_page(q->ring);
1081         if (size > page_size(qpg))
1082                 return -EINVAL;
1083
1084         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1085         return remap_pfn_range(vma, vma->vm_start, pfn,
1086                                size, vma->vm_page_prot);
1087 }
1088
1089 static int xsk_notifier(struct notifier_block *this,
1090                         unsigned long msg, void *ptr)
1091 {
1092         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1093         struct net *net = dev_net(dev);
1094         struct sock *sk;
1095
1096         switch (msg) {
1097         case NETDEV_UNREGISTER:
1098                 mutex_lock(&net->xdp.lock);
1099                 sk_for_each(sk, &net->xdp.list) {
1100                         struct xdp_sock *xs = xdp_sk(sk);
1101
1102                         mutex_lock(&xs->mutex);
1103                         if (xs->dev == dev) {
1104                                 sk->sk_err = ENETDOWN;
1105                                 if (!sock_flag(sk, SOCK_DEAD))
1106                                         sk->sk_error_report(sk);
1107
1108                                 xsk_unbind_dev(xs);
1109
1110                                 /* Clear device references. */
1111                                 xp_clear_dev(xs->pool);
1112                         }
1113                         mutex_unlock(&xs->mutex);
1114                 }
1115                 mutex_unlock(&net->xdp.lock);
1116                 break;
1117         }
1118         return NOTIFY_DONE;
1119 }
1120
1121 static struct proto xsk_proto = {
1122         .name =         "XDP",
1123         .owner =        THIS_MODULE,
1124         .obj_size =     sizeof(struct xdp_sock),
1125 };
1126
1127 static const struct proto_ops xsk_proto_ops = {
1128         .family         = PF_XDP,
1129         .owner          = THIS_MODULE,
1130         .release        = xsk_release,
1131         .bind           = xsk_bind,
1132         .connect        = sock_no_connect,
1133         .socketpair     = sock_no_socketpair,
1134         .accept         = sock_no_accept,
1135         .getname        = sock_no_getname,
1136         .poll           = xsk_poll,
1137         .ioctl          = sock_no_ioctl,
1138         .listen         = sock_no_listen,
1139         .shutdown       = sock_no_shutdown,
1140         .setsockopt     = xsk_setsockopt,
1141         .getsockopt     = xsk_getsockopt,
1142         .sendmsg        = xsk_sendmsg,
1143         .recvmsg        = sock_no_recvmsg,
1144         .mmap           = xsk_mmap,
1145         .sendpage       = sock_no_sendpage,
1146 };
1147
1148 static void xsk_destruct(struct sock *sk)
1149 {
1150         struct xdp_sock *xs = xdp_sk(sk);
1151
1152         if (!sock_flag(sk, SOCK_DEAD))
1153                 return;
1154
1155         if (!xp_put_pool(xs->pool))
1156                 xdp_put_umem(xs->umem, !xs->pool);
1157
1158         sk_refcnt_debug_dec(sk);
1159 }
1160
1161 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1162                       int kern)
1163 {
1164         struct xdp_sock *xs;
1165         struct sock *sk;
1166
1167         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1168                 return -EPERM;
1169         if (sock->type != SOCK_RAW)
1170                 return -ESOCKTNOSUPPORT;
1171
1172         if (protocol)
1173                 return -EPROTONOSUPPORT;
1174
1175         sock->state = SS_UNCONNECTED;
1176
1177         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1178         if (!sk)
1179                 return -ENOBUFS;
1180
1181         sock->ops = &xsk_proto_ops;
1182
1183         sock_init_data(sock, sk);
1184
1185         sk->sk_family = PF_XDP;
1186
1187         sk->sk_destruct = xsk_destruct;
1188         sk_refcnt_debug_inc(sk);
1189
1190         sock_set_flag(sk, SOCK_RCU_FREE);
1191
1192         xs = xdp_sk(sk);
1193         xs->state = XSK_READY;
1194         mutex_init(&xs->mutex);
1195         spin_lock_init(&xs->rx_lock);
1196         spin_lock_init(&xs->tx_completion_lock);
1197
1198         INIT_LIST_HEAD(&xs->map_list);
1199         spin_lock_init(&xs->map_list_lock);
1200
1201         mutex_lock(&net->xdp.lock);
1202         sk_add_node_rcu(sk, &net->xdp.list);
1203         mutex_unlock(&net->xdp.lock);
1204
1205         local_bh_disable();
1206         sock_prot_inuse_add(net, &xsk_proto, 1);
1207         local_bh_enable();
1208
1209         return 0;
1210 }
1211
1212 static const struct net_proto_family xsk_family_ops = {
1213         .family = PF_XDP,
1214         .create = xsk_create,
1215         .owner  = THIS_MODULE,
1216 };
1217
1218 static struct notifier_block xsk_netdev_notifier = {
1219         .notifier_call  = xsk_notifier,
1220 };
1221
1222 static int __net_init xsk_net_init(struct net *net)
1223 {
1224         mutex_init(&net->xdp.lock);
1225         INIT_HLIST_HEAD(&net->xdp.list);
1226         return 0;
1227 }
1228
1229 static void __net_exit xsk_net_exit(struct net *net)
1230 {
1231         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1232 }
1233
1234 static struct pernet_operations xsk_net_ops = {
1235         .init = xsk_net_init,
1236         .exit = xsk_net_exit,
1237 };
1238
1239 static int __init xsk_init(void)
1240 {
1241         int err, cpu;
1242
1243         err = proto_register(&xsk_proto, 0 /* no slab */);
1244         if (err)
1245                 goto out;
1246
1247         err = sock_register(&xsk_family_ops);
1248         if (err)
1249                 goto out_proto;
1250
1251         err = register_pernet_subsys(&xsk_net_ops);
1252         if (err)
1253                 goto out_sk;
1254
1255         err = register_netdevice_notifier(&xsk_netdev_notifier);
1256         if (err)
1257                 goto out_pernet;
1258
1259         for_each_possible_cpu(cpu)
1260                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1261         return 0;
1262
1263 out_pernet:
1264         unregister_pernet_subsys(&xsk_net_ops);
1265 out_sk:
1266         sock_unregister(PF_XDP);
1267 out_proto:
1268         proto_unregister(&xsk_proto);
1269 out:
1270         return err;
1271 }
1272
1273 fs_initcall(xsk_init);