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->num_rx_queues)
 111                 dev->_rx[queue_id].pool = NULL;
 112         if (queue_id < dev->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->pool->cq_lock, flags);
 368         xskq_prod_submit_addr(xs->pool->cq, addr);
 369         spin_unlock_irqrestore(&xs->pool->cq_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         unsigned long flags;
 382         int err = 0;
 383         u32 hr, tr;
 384
 385         mutex_lock(&xs->mutex);
 386
 387         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 388                 goto out;
 389
 390         hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
 391         tr = xs->dev->needed_tailroom;
 392
 393         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 394                 char *buffer;
 395                 u64 addr;
 396                 u32 len;
 397
 398                 if (max_batch-- == 0) {
 399                         err = -EAGAIN;
 400                         goto out;
 401                 }
 402
 403                 len = desc.len;
 404                 skb = sock_alloc_send_skb(sk, hr + len + tr, 1, &err);
 405                 if (unlikely(!skb))
 406                         goto out;
 407
 408                 skb_reserve(skb, hr);
 409                 skb_put(skb, len);
 410
 411                 addr = desc.addr;
 412                 buffer = xsk_buff_raw_get_data(xs->pool, addr);
 413                 err = skb_store_bits(skb, 0, buffer, len);
 414                 /* This is the backpressure mechanism for the Tx path.
 415                  * Reserve space in the completion queue and only proceed
 416                  * if there is space in it. This avoids having to implement
 417                  * any buffering in the Tx path.
 418                  */
 419                 spin_lock_irqsave(&xs->pool->cq_lock, flags);
 420                 if (unlikely(err) || xskq_prod_reserve(xs->pool->cq)) {
 421                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 422                         kfree_skb(skb);
 423                         goto out;
 424                 }
 425                 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 426
 427                 skb->dev = xs->dev;
 428                 skb->priority = sk->sk_priority;
 429                 skb->mark = sk->sk_mark;
 430                 skb_shinfo(skb)->destructor_arg = (void *)(long)desc.addr;
 431                 skb->destructor = xsk_destruct_skb;
 432
 433                 err = __dev_direct_xmit(skb, xs->queue_id);
 434                 if  (err == NETDEV_TX_BUSY) {
 435                         /* Tell user-space to retry the send */
 436                         skb->destructor = sock_wfree;
 437                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 438                         xskq_prod_cancel(xs->pool->cq);
 439                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 440                         /* Free skb without triggering the perf drop trace */
 441                         consume_skb(skb);
 442                         err = -EAGAIN;
 443                         goto out;
 444                 }
 445
 446                 xskq_cons_release(xs->tx);
 447                 /* Ignore NET_XMIT_CN as packet might have been sent */
 448                 if (err == NET_XMIT_DROP) {
 449                         /* SKB completed but not sent */
 450                         err = -EBUSY;
 451                         goto out;
 452                 }
 453
 454                 sent_frame = true;
 455         }
 456
 457         xs->tx->queue_empty_descs++;
 458
 459 out:
 460         if (sent_frame)
 461                 if (xsk_tx_writeable(xs))
 462                         sk->sk_write_space(sk);
 463
 464         mutex_unlock(&xs->mutex);
 465         return err;
 466 }
 467
 468 static int __xsk_sendmsg(struct sock *sk)
 469 {
 470         struct xdp_sock *xs = xdp_sk(sk);
 471
 472         if (unlikely(!(xs->dev->flags & IFF_UP)))
 473                 return -ENETDOWN;
 474         if (unlikely(!xs->tx))
 475                 return -ENOBUFS;
 476
 477         return xs->zc ? xsk_zc_xmit(xs) : xsk_generic_xmit(sk);
 478 }
 479
 480 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 481 {
 482         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 483         struct sock *sk = sock->sk;
 484         struct xdp_sock *xs = xdp_sk(sk);
 485
 486         if (unlikely(!xsk_is_bound(xs)))
 487                 return -ENXIO;
 488         if (unlikely(need_wait))
 489                 return -EOPNOTSUPP;
 490
 491         return __xsk_sendmsg(sk);
 492 }
 493
 494 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 495                              struct poll_table_struct *wait)
 496 {
 497         __poll_t mask = 0;
 498         struct sock *sk = sock->sk;
 499         struct xdp_sock *xs = xdp_sk(sk);
 500         struct xsk_buff_pool *pool;
 501
 502         sock_poll_wait(file, sock, wait);
 503
 504         if (unlikely(!xsk_is_bound(xs)))
 505                 return mask;
 506
 507         pool = xs->pool;
 508
 509         if (pool->cached_need_wakeup) {
 510                 if (xs->zc)
 511                         xsk_wakeup(xs, pool->cached_need_wakeup);
 512                 else
 513                         /* Poll needs to drive Tx also in copy mode */
 514                         __xsk_sendmsg(sk);
 515         }
 516
 517         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 518                 mask |= EPOLLIN | EPOLLRDNORM;
 519         if (xs->tx && xsk_tx_writeable(xs))
 520                 mask |= EPOLLOUT | EPOLLWRNORM;
 521
 522         return mask;
 523 }
 524
 525 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 526                           bool umem_queue)
 527 {
 528         struct xsk_queue *q;
 529
 530         if (entries == 0 || *queue || !is_power_of_2(entries))
 531                 return -EINVAL;
 532
 533         q = xskq_create(entries, umem_queue);
 534         if (!q)
 535                 return -ENOMEM;
 536
 537         /* Make sure queue is ready before it can be seen by others */
 538         smp_wmb();
 539         WRITE_ONCE(*queue, q);
 540         return 0;
 541 }
 542
 543 static void xsk_unbind_dev(struct xdp_sock *xs)
 544 {
 545         struct net_device *dev = xs->dev;
 546
 547         if (xs->state != XSK_BOUND)
 548                 return;
 549         WRITE_ONCE(xs->state, XSK_UNBOUND);
 550
 551         /* Wait for driver to stop using the xdp socket. */
 552         xp_del_xsk(xs->pool, xs);
 553         xs->dev = NULL;
 554         synchronize_net();
 555         dev_put(dev);
 556 }
 557
 558 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 559                                               struct xdp_sock ***map_entry)
 560 {
 561         struct xsk_map *map = NULL;
 562         struct xsk_map_node *node;
 563
 564         *map_entry = NULL;
 565
 566         spin_lock_bh(&xs->map_list_lock);
 567         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 568                                         node);
 569         if (node) {
 570                 WARN_ON(xsk_map_inc(node->map));
 571                 map = node->map;
 572                 *map_entry = node->map_entry;
 573         }
 574         spin_unlock_bh(&xs->map_list_lock);
 575         return map;
 576 }
 577
 578 static void xsk_delete_from_maps(struct xdp_sock *xs)
 579 {
 580         /* This function removes the current XDP socket from all the
 581          * maps it resides in. We need to take extra care here, due to
 582          * the two locks involved. Each map has a lock synchronizing
 583          * updates to the entries, and each socket has a lock that
 584          * synchronizes access to the list of maps (map_list). For
 585          * deadlock avoidance the locks need to be taken in the order
 586          * "map lock"->"socket map list lock". We start off by
 587          * accessing the socket map list, and take a reference to the
 588          * map to guarantee existence between the
 589          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 590          * calls. Then we ask the map to remove the socket, which
 591          * tries to remove the socket from the map. Note that there
 592          * might be updates to the map between
 593          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 594          */
 595         struct xdp_sock **map_entry = NULL;
 596         struct xsk_map *map;
 597
 598         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 599                 xsk_map_try_sock_delete(map, xs, map_entry);
 600                 xsk_map_put(map);
 601         }
 602 }
 603
 604 static int xsk_release(struct socket *sock)
 605 {
 606         struct sock *sk = sock->sk;
 607         struct xdp_sock *xs = xdp_sk(sk);
 608         struct net *net;
 609
 610         if (!sk)
 611                 return 0;
 612
 613         net = sock_net(sk);
 614
 615         mutex_lock(&net->xdp.lock);
 616         sk_del_node_init_rcu(sk);
 617         mutex_unlock(&net->xdp.lock);
 618
 619         local_bh_disable();
 620         sock_prot_inuse_add(net, sk->sk_prot, -1);
 621         local_bh_enable();
 622
 623         xsk_delete_from_maps(xs);
 624         mutex_lock(&xs->mutex);
 625         xsk_unbind_dev(xs);
 626         mutex_unlock(&xs->mutex);
 627
 628         xskq_destroy(xs->rx);
 629         xskq_destroy(xs->tx);
 630         xskq_destroy(xs->fq_tmp);
 631         xskq_destroy(xs->cq_tmp);
 632
 633         sock_orphan(sk);
 634         sock->sk = NULL;
 635
 636         sk_refcnt_debug_release(sk);
 637         sock_put(sk);
 638
 639         return 0;
 640 }
 641
 642 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 643 {
 644         struct socket *sock;
 645         int err;
 646
 647         sock = sockfd_lookup(fd, &err);
 648         if (!sock)
 649                 return ERR_PTR(-ENOTSOCK);
 650
 651         if (sock->sk->sk_family != PF_XDP) {
 652                 sockfd_put(sock);
 653                 return ERR_PTR(-ENOPROTOOPT);
 654         }
 655
 656         return sock;
 657 }
 658
 659 static bool xsk_validate_queues(struct xdp_sock *xs)
 660 {
 661         return xs->fq_tmp && xs->cq_tmp;
 662 }
 663
 664 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 665 {
 666         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 667         struct sock *sk = sock->sk;
 668         struct xdp_sock *xs = xdp_sk(sk);
 669         struct net_device *dev;
 670         int bound_dev_if;
 671         u32 flags, qid;
 672         int err = 0;
 673
 674         if (addr_len < sizeof(struct sockaddr_xdp))
 675                 return -EINVAL;
 676         if (sxdp->sxdp_family != AF_XDP)
 677                 return -EINVAL;
 678
 679         flags = sxdp->sxdp_flags;
 680         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 681                       XDP_USE_NEED_WAKEUP))
 682                 return -EINVAL;
 683
 684         bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
 685         if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex)
 686                 return -EINVAL;
 687
 688         rtnl_lock();
 689         mutex_lock(&xs->mutex);
 690         if (xs->state != XSK_READY) {
 691                 err = -EBUSY;
 692                 goto out_release;
 693         }
 694
 695         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 696         if (!dev) {
 697                 err = -ENODEV;
 698                 goto out_release;
 699         }
 700
 701         if (!xs->rx && !xs->tx) {
 702                 err = -EINVAL;
 703                 goto out_unlock;
 704         }
 705
 706         qid = sxdp->sxdp_queue_id;
 707
 708         if (flags & XDP_SHARED_UMEM) {
 709                 struct xdp_sock *umem_xs;
 710                 struct socket *sock;
 711
 712                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 713                     (flags & XDP_USE_NEED_WAKEUP)) {
 714                         /* Cannot specify flags for shared sockets. */
 715                         err = -EINVAL;
 716                         goto out_unlock;
 717                 }
 718
 719                 if (xs->umem) {
 720                         /* We have already our own. */
 721                         err = -EINVAL;
 722                         goto out_unlock;
 723                 }
 724
 725                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 726                 if (IS_ERR(sock)) {
 727                         err = PTR_ERR(sock);
 728                         goto out_unlock;
 729                 }
 730
 731                 umem_xs = xdp_sk(sock->sk);
 732                 if (!xsk_is_bound(umem_xs)) {
 733                         err = -EBADF;
 734                         sockfd_put(sock);
 735                         goto out_unlock;
 736                 }
 737
 738                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 739                         /* Share the umem with another socket on another qid
 740                          * and/or device.
 741                          */
 742                         xs->pool = xp_create_and_assign_umem(xs,
 743                                                              umem_xs->umem);
 744                         if (!xs->pool) {
 745                                 err = -ENOMEM;
 746                                 sockfd_put(sock);
 747                                 goto out_unlock;
 748                         }
 749
 750                         err = xp_assign_dev_shared(xs->pool, umem_xs, dev,
 751                                                    qid);
 752                         if (err) {
 753                                 xp_destroy(xs->pool);
 754                                 xs->pool = NULL;
 755                                 sockfd_put(sock);
 756                                 goto out_unlock;
 757                         }
 758                 } else {
 759                         /* Share the buffer pool with the other socket. */
 760                         if (xs->fq_tmp || xs->cq_tmp) {
 761                                 /* Do not allow setting your own fq or cq. */
 762                                 err = -EINVAL;
 763                                 sockfd_put(sock);
 764                                 goto out_unlock;
 765                         }
 766
 767                         xp_get_pool(umem_xs->pool);
 768                         xs->pool = umem_xs->pool;
 769                 }
 770
 771                 xdp_get_umem(umem_xs->umem);
 772                 WRITE_ONCE(xs->umem, umem_xs->umem);
 773                 sockfd_put(sock);
 774         } else if (!xs->umem || !xsk_validate_queues(xs)) {
 775                 err = -EINVAL;
 776                 goto out_unlock;
 777         } else {
 778                 /* This xsk has its own umem. */
 779                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
 780                 if (!xs->pool) {
 781                         err = -ENOMEM;
 782                         goto out_unlock;
 783                 }
 784
 785                 err = xp_assign_dev(xs->pool, dev, qid, flags);
 786                 if (err) {
 787                         xp_destroy(xs->pool);
 788                         xs->pool = NULL;
 789                         goto out_unlock;
 790                 }
 791         }
 792
 793         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
 794         xs->fq_tmp = NULL;
 795         xs->cq_tmp = NULL;
 796
 797         xs->dev = dev;
 798         xs->zc = xs->umem->zc;
 799         xs->queue_id = qid;
 800         xp_add_xsk(xs->pool, xs);
 801
 802 out_unlock:
 803         if (err) {
 804                 dev_put(dev);
 805         } else {
 806                 /* Matches smp_rmb() in bind() for shared umem
 807                  * sockets, and xsk_is_bound().
 808                  */
 809                 smp_wmb();
 810                 WRITE_ONCE(xs->state, XSK_BOUND);
 811         }
 812 out_release:
 813         mutex_unlock(&xs->mutex);
 814         rtnl_unlock();
 815         return err;
 816 }
 817
 818 struct xdp_umem_reg_v1 {
 819         __u64 addr; /* Start of packet data area */
 820         __u64 len; /* Length of packet data area */
 821         __u32 chunk_size;
 822         __u32 headroom;
 823 };
 824
 825 static int xsk_setsockopt(struct socket *sock, int level, int optname,
 826                           sockptr_t optval, unsigned int optlen)
 827 {
 828         struct sock *sk = sock->sk;
 829         struct xdp_sock *xs = xdp_sk(sk);
 830         int err;
 831
 832         if (level != SOL_XDP)
 833                 return -ENOPROTOOPT;
 834
 835         switch (optname) {
 836         case XDP_RX_RING:
 837         case XDP_TX_RING:
 838         {
 839                 struct xsk_queue **q;
 840                 int entries;
 841
 842                 if (optlen < sizeof(entries))
 843                         return -EINVAL;
 844                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
 845                         return -EFAULT;
 846
 847                 mutex_lock(&xs->mutex);
 848                 if (xs->state != XSK_READY) {
 849                         mutex_unlock(&xs->mutex);
 850                         return -EBUSY;
 851                 }
 852                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
 853                 err = xsk_init_queue(entries, q, false);
 854                 if (!err && optname == XDP_TX_RING)
 855                         /* Tx needs to be explicitly woken up the first time */
 856                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
 857                 mutex_unlock(&xs->mutex);
 858                 return err;
 859         }
 860         case XDP_UMEM_REG:
 861         {
 862                 size_t mr_size = sizeof(struct xdp_umem_reg);
 863                 struct xdp_umem_reg mr = {};
 864                 struct xdp_umem *umem;
 865
 866                 if (optlen < sizeof(struct xdp_umem_reg_v1))
 867                         return -EINVAL;
 868                 else if (optlen < sizeof(mr))
 869                         mr_size = sizeof(struct xdp_umem_reg_v1);
 870
 871                 if (copy_from_sockptr(&mr, optval, mr_size))
 872                         return -EFAULT;
 873
 874                 mutex_lock(&xs->mutex);
 875                 if (xs->state != XSK_READY || xs->umem) {
 876                         mutex_unlock(&xs->mutex);
 877                         return -EBUSY;
 878                 }
 879
 880                 umem = xdp_umem_create(&mr);
 881                 if (IS_ERR(umem)) {
 882                         mutex_unlock(&xs->mutex);
 883                         return PTR_ERR(umem);
 884                 }
 885
 886                 /* Make sure umem is ready before it can be seen by others */
 887                 smp_wmb();
 888                 WRITE_ONCE(xs->umem, umem);
 889                 mutex_unlock(&xs->mutex);
 890                 return 0;
 891         }
 892         case XDP_UMEM_FILL_RING:
 893         case XDP_UMEM_COMPLETION_RING:
 894         {
 895                 struct xsk_queue **q;
 896                 int entries;
 897
 898                 if (optlen < sizeof(entries))
 899                         return -EINVAL;
 900                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
 901                         return -EFAULT;
 902
 903                 mutex_lock(&xs->mutex);
 904                 if (xs->state != XSK_READY) {
 905                         mutex_unlock(&xs->mutex);
 906                         return -EBUSY;
 907                 }
 908
 909                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
 910                         &xs->cq_tmp;
 911                 err = xsk_init_queue(entries, q, true);
 912                 mutex_unlock(&xs->mutex);
 913                 return err;
 914         }
 915         default:
 916                 break;
 917         }
 918
 919         return -ENOPROTOOPT;
 920 }
 921
 922 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
 923 {
 924         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
 925         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
 926         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
 927 }
 928
 929 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
 930 {
 931         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
 932         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
 933         ring->desc = offsetof(struct xdp_umem_ring, desc);
 934 }
 935
 936 struct xdp_statistics_v1 {
 937         __u64 rx_dropped;
 938         __u64 rx_invalid_descs;
 939         __u64 tx_invalid_descs;
 940 };
 941
 942 static int xsk_getsockopt(struct socket *sock, int level, int optname,
 943                           char __user *optval, int __user *optlen)
 944 {
 945         struct sock *sk = sock->sk;
 946         struct xdp_sock *xs = xdp_sk(sk);
 947         int len;
 948
 949         if (level != SOL_XDP)
 950                 return -ENOPROTOOPT;
 951
 952         if (get_user(len, optlen))
 953                 return -EFAULT;
 954         if (len < 0)
 955                 return -EINVAL;
 956
 957         switch (optname) {
 958         case XDP_STATISTICS:
 959         {
 960                 struct xdp_statistics stats = {};
 961                 bool extra_stats = true;
 962                 size_t stats_size;
 963
 964                 if (len < sizeof(struct xdp_statistics_v1)) {
 965                         return -EINVAL;
 966                 } else if (len < sizeof(stats)) {
 967                         extra_stats = false;
 968                         stats_size = sizeof(struct xdp_statistics_v1);
 969                 } else {
 970                         stats_size = sizeof(stats);
 971                 }
 972
 973                 mutex_lock(&xs->mutex);
 974                 stats.rx_dropped = xs->rx_dropped;
 975                 if (extra_stats) {
 976                         stats.rx_ring_full = xs->rx_queue_full;
 977                         stats.rx_fill_ring_empty_descs =
 978                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
 979                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
 980                 } else {
 981                         stats.rx_dropped += xs->rx_queue_full;
 982                 }
 983                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
 984                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
 985                 mutex_unlock(&xs->mutex);
 986
 987                 if (copy_to_user(optval, &stats, stats_size))
 988                         return -EFAULT;
 989                 if (put_user(stats_size, optlen))
 990                         return -EFAULT;
 991
 992                 return 0;
 993         }
 994         case XDP_MMAP_OFFSETS:
 995         {
 996                 struct xdp_mmap_offsets off;
 997                 struct xdp_mmap_offsets_v1 off_v1;
 998                 bool flags_supported = true;
 999                 void *to_copy;
1000
1001                 if (len < sizeof(off_v1))
1002                         return -EINVAL;
1003                 else if (len < sizeof(off))
1004                         flags_supported = false;
1005
1006                 if (flags_supported) {
1007                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1008                          * except for the flags field added to the end.
1009                          */
1010                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1011                                                &off.rx);
1012                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1013                                                &off.tx);
1014                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1015                                                &off.fr);
1016                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1017                                                &off.cr);
1018                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1019                                                 ptrs.flags);
1020                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1021                                                 ptrs.flags);
1022                         off.fr.flags = offsetof(struct xdp_umem_ring,
1023                                                 ptrs.flags);
1024                         off.cr.flags = offsetof(struct xdp_umem_ring,
1025                                                 ptrs.flags);
1026
1027                         len = sizeof(off);
1028                         to_copy = &off;
1029                 } else {
1030                         xsk_enter_rxtx_offsets(&off_v1.rx);
1031                         xsk_enter_rxtx_offsets(&off_v1.tx);
1032                         xsk_enter_umem_offsets(&off_v1.fr);
1033                         xsk_enter_umem_offsets(&off_v1.cr);
1034
1035                         len = sizeof(off_v1);
1036                         to_copy = &off_v1;
1037                 }
1038
1039                 if (copy_to_user(optval, to_copy, len))
1040                         return -EFAULT;
1041                 if (put_user(len, optlen))
1042                         return -EFAULT;
1043
1044                 return 0;
1045         }
1046         case XDP_OPTIONS:
1047         {
1048                 struct xdp_options opts = {};
1049
1050                 if (len < sizeof(opts))
1051                         return -EINVAL;
1052
1053                 mutex_lock(&xs->mutex);
1054                 if (xs->zc)
1055                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1056                 mutex_unlock(&xs->mutex);
1057
1058                 len = sizeof(opts);
1059                 if (copy_to_user(optval, &opts, len))
1060                         return -EFAULT;
1061                 if (put_user(len, optlen))
1062                         return -EFAULT;
1063
1064                 return 0;
1065         }
1066         default:
1067                 break;
1068         }
1069
1070         return -EOPNOTSUPP;
1071 }
1072
1073 static int xsk_mmap(struct file *file, struct socket *sock,
1074                     struct vm_area_struct *vma)
1075 {
1076         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1077         unsigned long size = vma->vm_end - vma->vm_start;
1078         struct xdp_sock *xs = xdp_sk(sock->sk);
1079         struct xsk_queue *q = NULL;
1080         unsigned long pfn;
1081         struct page *qpg;
1082
1083         if (READ_ONCE(xs->state) != XSK_READY)
1084                 return -EBUSY;
1085
1086         if (offset == XDP_PGOFF_RX_RING) {
1087                 q = READ_ONCE(xs->rx);
1088         } else if (offset == XDP_PGOFF_TX_RING) {
1089                 q = READ_ONCE(xs->tx);
1090         } else {
1091                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1092                 smp_rmb();
1093                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1094                         q = READ_ONCE(xs->fq_tmp);
1095                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1096                         q = READ_ONCE(xs->cq_tmp);
1097         }
1098
1099         if (!q)
1100                 return -EINVAL;
1101
1102         /* Matches the smp_wmb() in xsk_init_queue */
1103         smp_rmb();
1104         qpg = virt_to_head_page(q->ring);
1105         if (size > page_size(qpg))
1106                 return -EINVAL;
1107
1108         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1109         return remap_pfn_range(vma, vma->vm_start, pfn,
1110                                size, vma->vm_page_prot);
1111 }
1112
1113 static int xsk_notifier(struct notifier_block *this,
1114                         unsigned long msg, void *ptr)
1115 {
1116         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1117         struct net *net = dev_net(dev);
1118         struct sock *sk;
1119
1120         switch (msg) {
1121         case NETDEV_UNREGISTER:
1122                 mutex_lock(&net->xdp.lock);
1123                 sk_for_each(sk, &net->xdp.list) {
1124                         struct xdp_sock *xs = xdp_sk(sk);
1125
1126                         mutex_lock(&xs->mutex);
1127                         if (xs->dev == dev) {
1128                                 sk->sk_err = ENETDOWN;
1129                                 if (!sock_flag(sk, SOCK_DEAD))
1130                                         sk->sk_error_report(sk);
1131
1132                                 xsk_unbind_dev(xs);
1133
1134                                 /* Clear device references. */
1135                                 xp_clear_dev(xs->pool);
1136                         }
1137                         mutex_unlock(&xs->mutex);
1138                 }
1139                 mutex_unlock(&net->xdp.lock);
1140                 break;
1141         }
1142         return NOTIFY_DONE;
1143 }
1144
1145 static struct proto xsk_proto = {
1146         .name =         "XDP",
1147         .owner =        THIS_MODULE,
1148         .obj_size =     sizeof(struct xdp_sock),
1149 };
1150
1151 static const struct proto_ops xsk_proto_ops = {
1152         .family         = PF_XDP,
1153         .owner          = THIS_MODULE,
1154         .release        = xsk_release,
1155         .bind           = xsk_bind,
1156         .connect        = sock_no_connect,
1157         .socketpair     = sock_no_socketpair,
1158         .accept         = sock_no_accept,
1159         .getname        = sock_no_getname,
1160         .poll           = xsk_poll,
1161         .ioctl          = sock_no_ioctl,
1162         .listen         = sock_no_listen,
1163         .shutdown       = sock_no_shutdown,
1164         .setsockopt     = xsk_setsockopt,
1165         .getsockopt     = xsk_getsockopt,
1166         .sendmsg        = xsk_sendmsg,
1167         .recvmsg        = sock_no_recvmsg,
1168         .mmap           = xsk_mmap,
1169         .sendpage       = sock_no_sendpage,
1170 };
1171
1172 static void xsk_destruct(struct sock *sk)
1173 {
1174         struct xdp_sock *xs = xdp_sk(sk);
1175
1176         if (!sock_flag(sk, SOCK_DEAD))
1177                 return;
1178
1179         if (!xp_put_pool(xs->pool))
1180                 xdp_put_umem(xs->umem, !xs->pool);
1181
1182         sk_refcnt_debug_dec(sk);
1183 }
1184
1185 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1186                       int kern)
1187 {
1188         struct xdp_sock *xs;
1189         struct sock *sk;
1190
1191         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1192                 return -EPERM;
1193         if (sock->type != SOCK_RAW)
1194                 return -ESOCKTNOSUPPORT;
1195
1196         if (protocol)
1197                 return -EPROTONOSUPPORT;
1198
1199         sock->state = SS_UNCONNECTED;
1200
1201         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1202         if (!sk)
1203                 return -ENOBUFS;
1204
1205         sock->ops = &xsk_proto_ops;
1206
1207         sock_init_data(sock, sk);
1208
1209         sk->sk_family = PF_XDP;
1210
1211         sk->sk_destruct = xsk_destruct;
1212         sk_refcnt_debug_inc(sk);
1213
1214         sock_set_flag(sk, SOCK_RCU_FREE);
1215
1216         xs = xdp_sk(sk);
1217         xs->state = XSK_READY;
1218         mutex_init(&xs->mutex);
1219         spin_lock_init(&xs->rx_lock);
1220
1221         INIT_LIST_HEAD(&xs->map_list);
1222         spin_lock_init(&xs->map_list_lock);
1223
1224         mutex_lock(&net->xdp.lock);
1225         sk_add_node_rcu(sk, &net->xdp.list);
1226         mutex_unlock(&net->xdp.lock);
1227
1228         local_bh_disable();
1229         sock_prot_inuse_add(net, &xsk_proto, 1);
1230         local_bh_enable();
1231
1232         return 0;
1233 }
1234
1235 static const struct net_proto_family xsk_family_ops = {
1236         .family = PF_XDP,
1237         .create = xsk_create,
1238         .owner  = THIS_MODULE,
1239 };
1240
1241 static struct notifier_block xsk_netdev_notifier = {
1242         .notifier_call  = xsk_notifier,
1243 };
1244
1245 static int __net_init xsk_net_init(struct net *net)
1246 {
1247         mutex_init(&net->xdp.lock);
1248         INIT_HLIST_HEAD(&net->xdp.list);
1249         return 0;
1250 }
1251
1252 static void __net_exit xsk_net_exit(struct net *net)
1253 {
1254         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1255 }
1256
1257 static struct pernet_operations xsk_net_ops = {
1258         .init = xsk_net_init,
1259         .exit = xsk_net_exit,
1260 };
1261
1262 static int __init xsk_init(void)
1263 {
1264         int err, cpu;
1265
1266         err = proto_register(&xsk_proto, 0 /* no slab */);
1267         if (err)
1268                 goto out;
1269
1270         err = sock_register(&xsk_family_ops);
1271         if (err)
1272                 goto out_proto;
1273
1274         err = register_pernet_subsys(&xsk_net_ops);
1275         if (err)
1276                 goto out_sk;
1277
1278         err = register_netdevice_notifier(&xsk_netdev_notifier);
1279         if (err)
1280                 goto out_pernet;
1281
1282         for_each_possible_cpu(cpu)
1283                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1284         return 0;
1285
1286 out_pernet:
1287         unregister_pernet_subsys(&xsk_net_ops);
1288 out_sk:
1289         sock_unregister(PF_XDP);
1290 out_proto:
1291         proto_unregister(&xsk_proto);
1292 out:
1293         return err;
1294 }
1295
1296 fs_initcall(xsk_init);