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