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/busy_poll.h>
  27 #include <net/xdp.h>
  28
  29 #include "xsk_queue.h"
  30 #include "xdp_umem.h"
  31 #include "xsk.h"
  32
  33 #define TX_BATCH_SIZE 32
  34
  35 static DEFINE_PER_CPU(struct list_head, xskmap_flush_list);
  36
  37 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool)
  38 {
  39         if (pool->cached_need_wakeup & XDP_WAKEUP_RX)
  40                 return;
  41
  42         pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
  43         pool->cached_need_wakeup |= XDP_WAKEUP_RX;
  44 }
  45 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
  46
  47 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool)
  48 {
  49         struct xdp_sock *xs;
  50
  51         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
  52                 return;
  53
  54         rcu_read_lock();
  55         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
  56                 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
  57         }
  58         rcu_read_unlock();
  59
  60         pool->cached_need_wakeup |= XDP_WAKEUP_TX;
  61 }
  62 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
  63
  64 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool)
  65 {
  66         if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX))
  67                 return;
  68
  69         pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
  70         pool->cached_need_wakeup &= ~XDP_WAKEUP_RX;
  71 }
  72 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
  73
  74 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool)
  75 {
  76         struct xdp_sock *xs;
  77
  78         if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX))
  79                 return;
  80
  81         rcu_read_lock();
  82         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
  83                 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
  84         }
  85         rcu_read_unlock();
  86
  87         pool->cached_need_wakeup &= ~XDP_WAKEUP_TX;
  88 }
  89 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
  90
  91 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool)
  92 {
  93         return pool->uses_need_wakeup;
  94 }
  95 EXPORT_SYMBOL(xsk_uses_need_wakeup);
  96
  97 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev,
  98                                             u16 queue_id)
  99 {
 100         if (queue_id < dev->real_num_rx_queues)
 101                 return dev->_rx[queue_id].pool;
 102         if (queue_id < dev->real_num_tx_queues)
 103                 return dev->_tx[queue_id].pool;
 104
 105         return NULL;
 106 }
 107 EXPORT_SYMBOL(xsk_get_pool_from_qid);
 108
 109 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id)
 110 {
 111         if (queue_id < dev->num_rx_queues)
 112                 dev->_rx[queue_id].pool = NULL;
 113         if (queue_id < dev->num_tx_queues)
 114                 dev->_tx[queue_id].pool = NULL;
 115 }
 116
 117 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do
 118  * not know if the device has more tx queues than rx, or the opposite.
 119  * This might also change during run time.
 120  */
 121 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool,
 122                         u16 queue_id)
 123 {
 124         if (queue_id >= max_t(unsigned int,
 125                               dev->real_num_rx_queues,
 126                               dev->real_num_tx_queues))
 127                 return -EINVAL;
 128
 129         if (queue_id < dev->real_num_rx_queues)
 130                 dev->_rx[queue_id].pool = pool;
 131         if (queue_id < dev->real_num_tx_queues)
 132                 dev->_tx[queue_id].pool = pool;
 133
 134         return 0;
 135 }
 136
 137 void xp_release(struct xdp_buff_xsk *xskb)
 138 {
 139         xskb->pool->free_heads[xskb->pool->free_heads_cnt++] = xskb;
 140 }
 141
 142 static u64 xp_get_handle(struct xdp_buff_xsk *xskb)
 143 {
 144         u64 offset = xskb->xdp.data - xskb->xdp.data_hard_start;
 145
 146         offset += xskb->pool->headroom;
 147         if (!xskb->pool->unaligned)
 148                 return xskb->orig_addr + offset;
 149         return xskb->orig_addr + (offset << XSK_UNALIGNED_BUF_OFFSET_SHIFT);
 150 }
 151
 152 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
 153 {
 154         struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
 155         u64 addr;
 156         int err;
 157
 158         addr = xp_get_handle(xskb);
 159         err = xskq_prod_reserve_desc(xs->rx, addr, len);
 160         if (err) {
 161                 xs->rx_queue_full++;
 162                 return err;
 163         }
 164
 165         xp_release(xskb);
 166         return 0;
 167 }
 168
 169 static void xsk_copy_xdp(struct xdp_buff *to, struct xdp_buff *from, u32 len)
 170 {
 171         void *from_buf, *to_buf;
 172         u32 metalen;
 173
 174         if (unlikely(xdp_data_meta_unsupported(from))) {
 175                 from_buf = from->data;
 176                 to_buf = to->data;
 177                 metalen = 0;
 178         } else {
 179                 from_buf = from->data_meta;
 180                 metalen = from->data - from->data_meta;
 181                 to_buf = to->data - metalen;
 182         }
 183
 184         memcpy(to_buf, from_buf, len + metalen);
 185 }
 186
 187 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 188 {
 189         struct xdp_buff *xsk_xdp;
 190         int err;
 191         u32 len;
 192
 193         len = xdp->data_end - xdp->data;
 194         if (len > xsk_pool_get_rx_frame_size(xs->pool)) {
 195                 xs->rx_dropped++;
 196                 return -ENOSPC;
 197         }
 198
 199         xsk_xdp = xsk_buff_alloc(xs->pool);
 200         if (!xsk_xdp) {
 201                 xs->rx_dropped++;
 202                 return -ENOSPC;
 203         }
 204
 205         xsk_copy_xdp(xsk_xdp, xdp, len);
 206         err = __xsk_rcv_zc(xs, xsk_xdp, len);
 207         if (err) {
 208                 xsk_buff_free(xsk_xdp);
 209                 return err;
 210         }
 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_check(struct xdp_sock *xs, struct xdp_buff *xdp)
 233 {
 234         if (!xsk_is_bound(xs))
 235                 return -EINVAL;
 236
 237         if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
 238                 return -EINVAL;
 239
 240         sk_mark_napi_id_once_xdp(&xs->sk, xdp);
 241         return 0;
 242 }
 243
 244 static void xsk_flush(struct xdp_sock *xs)
 245 {
 246         xskq_prod_submit(xs->rx);
 247         __xskq_cons_release(xs->pool->fq);
 248         sock_def_readable(&xs->sk);
 249 }
 250
 251 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 252 {
 253         int err;
 254
 255         spin_lock_bh(&xs->rx_lock);
 256         err = xsk_rcv_check(xs, xdp);
 257         if (!err) {
 258                 err = __xsk_rcv(xs, xdp);
 259                 xsk_flush(xs);
 260         }
 261         spin_unlock_bh(&xs->rx_lock);
 262         return err;
 263 }
 264
 265 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 266 {
 267         int err;
 268         u32 len;
 269
 270         err = xsk_rcv_check(xs, xdp);
 271         if (err)
 272                 return err;
 273
 274         if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) {
 275                 len = xdp->data_end - xdp->data;
 276                 return __xsk_rcv_zc(xs, xdp, len);
 277         }
 278
 279         err = __xsk_rcv(xs, xdp);
 280         if (!err)
 281                 xdp_return_buff(xdp);
 282         return err;
 283 }
 284
 285 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
 286 {
 287         struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
 288         int err;
 289
 290         err = xsk_rcv(xs, xdp);
 291         if (err)
 292                 return err;
 293
 294         if (!xs->flush_node.prev)
 295                 list_add(&xs->flush_node, flush_list);
 296
 297         return 0;
 298 }
 299
 300 void __xsk_map_flush(void)
 301 {
 302         struct list_head *flush_list = this_cpu_ptr(&xskmap_flush_list);
 303         struct xdp_sock *xs, *tmp;
 304
 305         list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
 306                 xsk_flush(xs);
 307                 __list_del_clearprev(&xs->flush_node);
 308         }
 309 }
 310
 311 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
 312 {
 313         xskq_prod_submit_n(pool->cq, nb_entries);
 314 }
 315 EXPORT_SYMBOL(xsk_tx_completed);
 316
 317 void xsk_tx_release(struct xsk_buff_pool *pool)
 318 {
 319         struct xdp_sock *xs;
 320
 321         rcu_read_lock();
 322         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
 323                 __xskq_cons_release(xs->tx);
 324                 if (xsk_tx_writeable(xs))
 325                         xs->sk.sk_write_space(&xs->sk);
 326         }
 327         rcu_read_unlock();
 328 }
 329 EXPORT_SYMBOL(xsk_tx_release);
 330
 331 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
 332 {
 333         struct xdp_sock *xs;
 334
 335         rcu_read_lock();
 336         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
 337                 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
 338                         xs->tx->queue_empty_descs++;
 339                         continue;
 340                 }
 341
 342                 /* This is the backpressure mechanism for the Tx path.
 343                  * Reserve space in the completion queue and only proceed
 344                  * if there is space in it. This avoids having to implement
 345                  * any buffering in the Tx path.
 346                  */
 347                 if (xskq_prod_reserve_addr(pool->cq, desc->addr))
 348                         goto out;
 349
 350                 xskq_cons_release(xs->tx);
 351                 rcu_read_unlock();
 352                 return true;
 353         }
 354
 355 out:
 356         rcu_read_unlock();
 357         return false;
 358 }
 359 EXPORT_SYMBOL(xsk_tx_peek_desc);
 360
 361 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, u32 max_entries)
 362 {
 363         struct xdp_desc *descs = pool->tx_descs;
 364         u32 nb_pkts = 0;
 365
 366         while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts]))
 367                 nb_pkts++;
 368
 369         xsk_tx_release(pool);
 370         return nb_pkts;
 371 }
 372
 373 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, u32 nb_pkts)
 374 {
 375         struct xdp_sock *xs;
 376
 377         rcu_read_lock();
 378         if (!list_is_singular(&pool->xsk_tx_list)) {
 379                 /* Fallback to the non-batched version */
 380                 rcu_read_unlock();
 381                 return xsk_tx_peek_release_fallback(pool, nb_pkts);
 382         }
 383
 384         xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
 385         if (!xs) {
 386                 nb_pkts = 0;
 387                 goto out;
 388         }
 389
 390         nb_pkts = xskq_cons_nb_entries(xs->tx, nb_pkts);
 391
 392         /* This is the backpressure mechanism for the Tx path. Try to
 393          * reserve space in the completion queue for all packets, but
 394          * if there are fewer slots available, just process that many
 395          * packets. This avoids having to implement any buffering in
 396          * the Tx path.
 397          */
 398         nb_pkts = xskq_prod_nb_free(pool->cq, nb_pkts);
 399         if (!nb_pkts)
 400                 goto out;
 401
 402         nb_pkts = xskq_cons_read_desc_batch(xs->tx, pool, nb_pkts);
 403         if (!nb_pkts) {
 404                 xs->tx->queue_empty_descs++;
 405                 goto out;
 406         }
 407
 408         __xskq_cons_release(xs->tx);
 409         xskq_prod_write_addr_batch(pool->cq, pool->tx_descs, nb_pkts);
 410         xs->sk.sk_write_space(&xs->sk);
 411
 412 out:
 413         rcu_read_unlock();
 414         return nb_pkts;
 415 }
 416 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
 417
 418 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
 419 {
 420         struct net_device *dev = xs->dev;
 421
 422         return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 423 }
 424
 425 static void xsk_destruct_skb(struct sk_buff *skb)
 426 {
 427         u64 addr = (u64)(long)skb_shinfo(skb)->destructor_arg;
 428         struct xdp_sock *xs = xdp_sk(skb->sk);
 429         unsigned long flags;
 430
 431         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 432         xskq_prod_submit_addr(xs->pool->cq, addr);
 433         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 434
 435         sock_wfree(skb);
 436 }
 437
 438 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
 439                                               struct xdp_desc *desc)
 440 {
 441         struct xsk_buff_pool *pool = xs->pool;
 442         u32 hr, len, ts, offset, copy, copied;
 443         struct sk_buff *skb;
 444         struct page *page;
 445         void *buffer;
 446         int err, i;
 447         u64 addr;
 448
 449         hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
 450
 451         skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
 452         if (unlikely(!skb))
 453                 return ERR_PTR(err);
 454
 455         skb_reserve(skb, hr);
 456
 457         addr = desc->addr;
 458         len = desc->len;
 459         ts = pool->unaligned ? len : pool->chunk_size;
 460
 461         buffer = xsk_buff_raw_get_data(pool, addr);
 462         offset = offset_in_page(buffer);
 463         addr = buffer - pool->addrs;
 464
 465         for (copied = 0, i = 0; copied < len; i++) {
 466                 page = pool->umem->pgs[addr >> PAGE_SHIFT];
 467                 get_page(page);
 468
 469                 copy = min_t(u32, PAGE_SIZE - offset, len - copied);
 470                 skb_fill_page_desc(skb, i, page, offset, copy);
 471
 472                 copied += copy;
 473                 addr += copy;
 474                 offset = 0;
 475         }
 476
 477         skb->len += len;
 478         skb->data_len += len;
 479         skb->truesize += ts;
 480
 481         refcount_add(ts, &xs->sk.sk_wmem_alloc);
 482
 483         return skb;
 484 }
 485
 486 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
 487                                      struct xdp_desc *desc)
 488 {
 489         struct net_device *dev = xs->dev;
 490         struct sk_buff *skb;
 491
 492         if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
 493                 skb = xsk_build_skb_zerocopy(xs, desc);
 494                 if (IS_ERR(skb))
 495                         return skb;
 496         } else {
 497                 u32 hr, tr, len;
 498                 void *buffer;
 499                 int err;
 500
 501                 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
 502                 tr = dev->needed_tailroom;
 503                 len = desc->len;
 504
 505                 skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
 506                 if (unlikely(!skb))
 507                         return ERR_PTR(err);
 508
 509                 skb_reserve(skb, hr);
 510                 skb_put(skb, len);
 511
 512                 buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
 513                 err = skb_store_bits(skb, 0, buffer, len);
 514                 if (unlikely(err)) {
 515                         kfree_skb(skb);
 516                         return ERR_PTR(err);
 517                 }
 518         }
 519
 520         skb->dev = dev;
 521         skb->priority = xs->sk.sk_priority;
 522         skb->mark = xs->sk.sk_mark;
 523         skb_shinfo(skb)->destructor_arg = (void *)(long)desc->addr;
 524         skb->destructor = xsk_destruct_skb;
 525
 526         return skb;
 527 }
 528
 529 static int xsk_generic_xmit(struct sock *sk)
 530 {
 531         struct xdp_sock *xs = xdp_sk(sk);
 532         u32 max_batch = TX_BATCH_SIZE;
 533         bool sent_frame = false;
 534         struct xdp_desc desc;
 535         struct sk_buff *skb;
 536         unsigned long flags;
 537         int err = 0;
 538
 539         mutex_lock(&xs->mutex);
 540
 541         /* Since we dropped the RCU read lock, the socket state might have changed. */
 542         if (unlikely(!xsk_is_bound(xs))) {
 543                 err = -ENXIO;
 544                 goto out;
 545         }
 546
 547         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 548                 goto out;
 549
 550         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 551                 if (max_batch-- == 0) {
 552                         err = -EAGAIN;
 553                         goto out;
 554                 }
 555
 556                 /* This is the backpressure mechanism for the Tx path.
 557                  * Reserve space in the completion queue and only proceed
 558                  * if there is space in it. This avoids having to implement
 559                  * any buffering in the Tx path.
 560                  */
 561                 spin_lock_irqsave(&xs->pool->cq_lock, flags);
 562                 if (xskq_prod_reserve(xs->pool->cq)) {
 563                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 564                         goto out;
 565                 }
 566                 spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 567
 568                 skb = xsk_build_skb(xs, &desc);
 569                 if (IS_ERR(skb)) {
 570                         err = PTR_ERR(skb);
 571                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 572                         xskq_prod_cancel(xs->pool->cq);
 573                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 574                         goto out;
 575                 }
 576
 577                 err = __dev_direct_xmit(skb, xs->queue_id);
 578                 if  (err == NETDEV_TX_BUSY) {
 579                         /* Tell user-space to retry the send */
 580                         skb->destructor = sock_wfree;
 581                         spin_lock_irqsave(&xs->pool->cq_lock, flags);
 582                         xskq_prod_cancel(xs->pool->cq);
 583                         spin_unlock_irqrestore(&xs->pool->cq_lock, flags);
 584                         /* Free skb without triggering the perf drop trace */
 585                         consume_skb(skb);
 586                         err = -EAGAIN;
 587                         goto out;
 588                 }
 589
 590                 xskq_cons_release(xs->tx);
 591                 /* Ignore NET_XMIT_CN as packet might have been sent */
 592                 if (err == NET_XMIT_DROP) {
 593                         /* SKB completed but not sent */
 594                         err = -EBUSY;
 595                         goto out;
 596                 }
 597
 598                 sent_frame = true;
 599         }
 600
 601         xs->tx->queue_empty_descs++;
 602
 603 out:
 604         if (sent_frame)
 605                 if (xsk_tx_writeable(xs))
 606                         sk->sk_write_space(sk);
 607
 608         mutex_unlock(&xs->mutex);
 609         return err;
 610 }
 611
 612 static int xsk_xmit(struct sock *sk)
 613 {
 614         struct xdp_sock *xs = xdp_sk(sk);
 615         int ret;
 616
 617         if (unlikely(!(xs->dev->flags & IFF_UP)))
 618                 return -ENETDOWN;
 619         if (unlikely(!xs->tx))
 620                 return -ENOBUFS;
 621
 622         if (xs->zc)
 623                 return xsk_wakeup(xs, XDP_WAKEUP_TX);
 624
 625         /* Drop the RCU lock since the SKB path might sleep. */
 626         rcu_read_unlock();
 627         ret = xsk_generic_xmit(sk);
 628         /* Reaquire RCU lock before going into common code. */
 629         rcu_read_lock();
 630
 631         return ret;
 632 }
 633
 634 static bool xsk_no_wakeup(struct sock *sk)
 635 {
 636 #ifdef CONFIG_NET_RX_BUSY_POLL
 637         /* Prefer busy-polling, skip the wakeup. */
 638         return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
 639                 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
 640 #else
 641         return false;
 642 #endif
 643 }
 644
 645 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 646 {
 647         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 648         struct sock *sk = sock->sk;
 649         struct xdp_sock *xs = xdp_sk(sk);
 650         struct xsk_buff_pool *pool;
 651
 652         if (unlikely(!xsk_is_bound(xs)))
 653                 return -ENXIO;
 654         if (unlikely(need_wait))
 655                 return -EOPNOTSUPP;
 656
 657         if (sk_can_busy_loop(sk))
 658                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 659
 660         if (xs->zc && xsk_no_wakeup(sk))
 661                 return 0;
 662
 663         pool = xs->pool;
 664         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
 665                 return xsk_xmit(sk);
 666         return 0;
 667 }
 668
 669 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 670 {
 671         int ret;
 672
 673         rcu_read_lock();
 674         ret = __xsk_sendmsg(sock, m, total_len);
 675         rcu_read_unlock();
 676
 677         return ret;
 678 }
 679
 680 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 681 {
 682         bool need_wait = !(flags & MSG_DONTWAIT);
 683         struct sock *sk = sock->sk;
 684         struct xdp_sock *xs = xdp_sk(sk);
 685
 686         if (unlikely(!xsk_is_bound(xs)))
 687                 return -ENXIO;
 688         if (unlikely(!(xs->dev->flags & IFF_UP)))
 689                 return -ENETDOWN;
 690         if (unlikely(!xs->rx))
 691                 return -ENOBUFS;
 692         if (unlikely(need_wait))
 693                 return -EOPNOTSUPP;
 694
 695         if (sk_can_busy_loop(sk))
 696                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 697
 698         if (xsk_no_wakeup(sk))
 699                 return 0;
 700
 701         if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
 702                 return xsk_wakeup(xs, XDP_WAKEUP_RX);
 703         return 0;
 704 }
 705
 706 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 707 {
 708         int ret;
 709
 710         rcu_read_lock();
 711         ret = __xsk_recvmsg(sock, m, len, flags);
 712         rcu_read_unlock();
 713
 714         return ret;
 715 }
 716
 717 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 718                              struct poll_table_struct *wait)
 719 {
 720         __poll_t mask = 0;
 721         struct sock *sk = sock->sk;
 722         struct xdp_sock *xs = xdp_sk(sk);
 723         struct xsk_buff_pool *pool;
 724
 725         sock_poll_wait(file, sock, wait);
 726
 727         rcu_read_lock();
 728         if (unlikely(!xsk_is_bound(xs))) {
 729                 rcu_read_unlock();
 730                 return mask;
 731         }
 732
 733         pool = xs->pool;
 734
 735         if (pool->cached_need_wakeup) {
 736                 if (xs->zc)
 737                         xsk_wakeup(xs, pool->cached_need_wakeup);
 738                 else
 739                         /* Poll needs to drive Tx also in copy mode */
 740                         xsk_xmit(sk);
 741         }
 742
 743         if (xs->rx && !xskq_prod_is_empty(xs->rx))
 744                 mask |= EPOLLIN | EPOLLRDNORM;
 745         if (xs->tx && xsk_tx_writeable(xs))
 746                 mask |= EPOLLOUT | EPOLLWRNORM;
 747
 748         rcu_read_unlock();
 749         return mask;
 750 }
 751
 752 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
 753                           bool umem_queue)
 754 {
 755         struct xsk_queue *q;
 756
 757         if (entries == 0 || *queue || !is_power_of_2(entries))
 758                 return -EINVAL;
 759
 760         q = xskq_create(entries, umem_queue);
 761         if (!q)
 762                 return -ENOMEM;
 763
 764         /* Make sure queue is ready before it can be seen by others */
 765         smp_wmb();
 766         WRITE_ONCE(*queue, q);
 767         return 0;
 768 }
 769
 770 static void xsk_unbind_dev(struct xdp_sock *xs)
 771 {
 772         struct net_device *dev = xs->dev;
 773
 774         if (xs->state != XSK_BOUND)
 775                 return;
 776         WRITE_ONCE(xs->state, XSK_UNBOUND);
 777
 778         /* Wait for driver to stop using the xdp socket. */
 779         xp_del_xsk(xs->pool, xs);
 780         synchronize_net();
 781         dev_put(dev);
 782 }
 783
 784 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
 785                                               struct xdp_sock __rcu ***map_entry)
 786 {
 787         struct xsk_map *map = NULL;
 788         struct xsk_map_node *node;
 789
 790         *map_entry = NULL;
 791
 792         spin_lock_bh(&xs->map_list_lock);
 793         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
 794                                         node);
 795         if (node) {
 796                 bpf_map_inc(&node->map->map);
 797                 map = node->map;
 798                 *map_entry = node->map_entry;
 799         }
 800         spin_unlock_bh(&xs->map_list_lock);
 801         return map;
 802 }
 803
 804 static void xsk_delete_from_maps(struct xdp_sock *xs)
 805 {
 806         /* This function removes the current XDP socket from all the
 807          * maps it resides in. We need to take extra care here, due to
 808          * the two locks involved. Each map has a lock synchronizing
 809          * updates to the entries, and each socket has a lock that
 810          * synchronizes access to the list of maps (map_list). For
 811          * deadlock avoidance the locks need to be taken in the order
 812          * "map lock"->"socket map list lock". We start off by
 813          * accessing the socket map list, and take a reference to the
 814          * map to guarantee existence between the
 815          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
 816          * calls. Then we ask the map to remove the socket, which
 817          * tries to remove the socket from the map. Note that there
 818          * might be updates to the map between
 819          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
 820          */
 821         struct xdp_sock __rcu **map_entry = NULL;
 822         struct xsk_map *map;
 823
 824         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
 825                 xsk_map_try_sock_delete(map, xs, map_entry);
 826                 bpf_map_put(&map->map);
 827         }
 828 }
 829
 830 static int xsk_release(struct socket *sock)
 831 {
 832         struct sock *sk = sock->sk;
 833         struct xdp_sock *xs = xdp_sk(sk);
 834         struct net *net;
 835
 836         if (!sk)
 837                 return 0;
 838
 839         net = sock_net(sk);
 840
 841         mutex_lock(&net->xdp.lock);
 842         sk_del_node_init_rcu(sk);
 843         mutex_unlock(&net->xdp.lock);
 844
 845         local_bh_disable();
 846         sock_prot_inuse_add(net, sk->sk_prot, -1);
 847         local_bh_enable();
 848
 849         xsk_delete_from_maps(xs);
 850         mutex_lock(&xs->mutex);
 851         xsk_unbind_dev(xs);
 852         mutex_unlock(&xs->mutex);
 853
 854         xskq_destroy(xs->rx);
 855         xskq_destroy(xs->tx);
 856         xskq_destroy(xs->fq_tmp);
 857         xskq_destroy(xs->cq_tmp);
 858
 859         sock_orphan(sk);
 860         sock->sk = NULL;
 861
 862         sk_refcnt_debug_release(sk);
 863         sock_put(sk);
 864
 865         return 0;
 866 }
 867
 868 static struct socket *xsk_lookup_xsk_from_fd(int fd)
 869 {
 870         struct socket *sock;
 871         int err;
 872
 873         sock = sockfd_lookup(fd, &err);
 874         if (!sock)
 875                 return ERR_PTR(-ENOTSOCK);
 876
 877         if (sock->sk->sk_family != PF_XDP) {
 878                 sockfd_put(sock);
 879                 return ERR_PTR(-ENOPROTOOPT);
 880         }
 881
 882         return sock;
 883 }
 884
 885 static bool xsk_validate_queues(struct xdp_sock *xs)
 886 {
 887         return xs->fq_tmp && xs->cq_tmp;
 888 }
 889
 890 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
 891 {
 892         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
 893         struct sock *sk = sock->sk;
 894         struct xdp_sock *xs = xdp_sk(sk);
 895         struct net_device *dev;
 896         int bound_dev_if;
 897         u32 flags, qid;
 898         int err = 0;
 899
 900         if (addr_len < sizeof(struct sockaddr_xdp))
 901                 return -EINVAL;
 902         if (sxdp->sxdp_family != AF_XDP)
 903                 return -EINVAL;
 904
 905         flags = sxdp->sxdp_flags;
 906         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
 907                       XDP_USE_NEED_WAKEUP))
 908                 return -EINVAL;
 909
 910         bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
 911         if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex)
 912                 return -EINVAL;
 913
 914         rtnl_lock();
 915         mutex_lock(&xs->mutex);
 916         if (xs->state != XSK_READY) {
 917                 err = -EBUSY;
 918                 goto out_release;
 919         }
 920
 921         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
 922         if (!dev) {
 923                 err = -ENODEV;
 924                 goto out_release;
 925         }
 926
 927         if (!xs->rx && !xs->tx) {
 928                 err = -EINVAL;
 929                 goto out_unlock;
 930         }
 931
 932         qid = sxdp->sxdp_queue_id;
 933
 934         if (flags & XDP_SHARED_UMEM) {
 935                 struct xdp_sock *umem_xs;
 936                 struct socket *sock;
 937
 938                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
 939                     (flags & XDP_USE_NEED_WAKEUP)) {
 940                         /* Cannot specify flags for shared sockets. */
 941                         err = -EINVAL;
 942                         goto out_unlock;
 943                 }
 944
 945                 if (xs->umem) {
 946                         /* We have already our own. */
 947                         err = -EINVAL;
 948                         goto out_unlock;
 949                 }
 950
 951                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
 952                 if (IS_ERR(sock)) {
 953                         err = PTR_ERR(sock);
 954                         goto out_unlock;
 955                 }
 956
 957                 umem_xs = xdp_sk(sock->sk);
 958                 if (!xsk_is_bound(umem_xs)) {
 959                         err = -EBADF;
 960                         sockfd_put(sock);
 961                         goto out_unlock;
 962                 }
 963
 964                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
 965                         /* Share the umem with another socket on another qid
 966                          * and/or device.
 967                          */
 968                         xs->pool = xp_create_and_assign_umem(xs,
 969                                                              umem_xs->umem);
 970                         if (!xs->pool) {
 971                                 err = -ENOMEM;
 972                                 sockfd_put(sock);
 973                                 goto out_unlock;
 974                         }
 975
 976                         err = xp_assign_dev_shared(xs->pool, umem_xs, dev,
 977                                                    qid);
 978                         if (err) {
 979                                 xp_destroy(xs->pool);
 980                                 xs->pool = NULL;
 981                                 sockfd_put(sock);
 982                                 goto out_unlock;
 983                         }
 984                 } else {
 985                         /* Share the buffer pool with the other socket. */
 986                         if (xs->fq_tmp || xs->cq_tmp) {
 987                                 /* Do not allow setting your own fq or cq. */
 988                                 err = -EINVAL;
 989                                 sockfd_put(sock);
 990                                 goto out_unlock;
 991                         }
 992
 993                         xp_get_pool(umem_xs->pool);
 994                         xs->pool = umem_xs->pool;
 995
 996                         /* If underlying shared umem was created without Tx
 997                          * ring, allocate Tx descs array that Tx batching API
 998                          * utilizes
 999                          */
1000                         if (xs->tx && !xs->pool->tx_descs) {
1001                                 err = xp_alloc_tx_descs(xs->pool, xs);
1002                                 if (err) {
1003                                         xp_put_pool(xs->pool);
1004                                         xs->pool = NULL;
1005                                         sockfd_put(sock);
1006                                         goto out_unlock;
1007                                 }
1008                         }
1009                 }
1010
1011                 xdp_get_umem(umem_xs->umem);
1012                 WRITE_ONCE(xs->umem, umem_xs->umem);
1013                 sockfd_put(sock);
1014         } else if (!xs->umem || !xsk_validate_queues(xs)) {
1015                 err = -EINVAL;
1016                 goto out_unlock;
1017         } else {
1018                 /* This xsk has its own umem. */
1019                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
1020                 if (!xs->pool) {
1021                         err = -ENOMEM;
1022                         goto out_unlock;
1023                 }
1024
1025                 err = xp_assign_dev(xs->pool, dev, qid, flags);
1026                 if (err) {
1027                         xp_destroy(xs->pool);
1028                         xs->pool = NULL;
1029                         goto out_unlock;
1030                 }
1031         }
1032
1033         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1034         xs->fq_tmp = NULL;
1035         xs->cq_tmp = NULL;
1036
1037         xs->dev = dev;
1038         xs->zc = xs->umem->zc;
1039         xs->queue_id = qid;
1040         xp_add_xsk(xs->pool, xs);
1041
1042 out_unlock:
1043         if (err) {
1044                 dev_put(dev);
1045         } else {
1046                 /* Matches smp_rmb() in bind() for shared umem
1047                  * sockets, and xsk_is_bound().
1048                  */
1049                 smp_wmb();
1050                 WRITE_ONCE(xs->state, XSK_BOUND);
1051         }
1052 out_release:
1053         mutex_unlock(&xs->mutex);
1054         rtnl_unlock();
1055         return err;
1056 }
1057
1058 struct xdp_umem_reg_v1 {
1059         __u64 addr; /* Start of packet data area */
1060         __u64 len; /* Length of packet data area */
1061         __u32 chunk_size;
1062         __u32 headroom;
1063 };
1064
1065 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1066                           sockptr_t optval, unsigned int optlen)
1067 {
1068         struct sock *sk = sock->sk;
1069         struct xdp_sock *xs = xdp_sk(sk);
1070         int err;
1071
1072         if (level != SOL_XDP)
1073                 return -ENOPROTOOPT;
1074
1075         switch (optname) {
1076         case XDP_RX_RING:
1077         case XDP_TX_RING:
1078         {
1079                 struct xsk_queue **q;
1080                 int entries;
1081
1082                 if (optlen < sizeof(entries))
1083                         return -EINVAL;
1084                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1085                         return -EFAULT;
1086
1087                 mutex_lock(&xs->mutex);
1088                 if (xs->state != XSK_READY) {
1089                         mutex_unlock(&xs->mutex);
1090                         return -EBUSY;
1091                 }
1092                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1093                 err = xsk_init_queue(entries, q, false);
1094                 if (!err && optname == XDP_TX_RING)
1095                         /* Tx needs to be explicitly woken up the first time */
1096                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1097                 mutex_unlock(&xs->mutex);
1098                 return err;
1099         }
1100         case XDP_UMEM_REG:
1101         {
1102                 size_t mr_size = sizeof(struct xdp_umem_reg);
1103                 struct xdp_umem_reg mr = {};
1104                 struct xdp_umem *umem;
1105
1106                 if (optlen < sizeof(struct xdp_umem_reg_v1))
1107                         return -EINVAL;
1108                 else if (optlen < sizeof(mr))
1109                         mr_size = sizeof(struct xdp_umem_reg_v1);
1110
1111                 if (copy_from_sockptr(&mr, optval, mr_size))
1112                         return -EFAULT;
1113
1114                 mutex_lock(&xs->mutex);
1115                 if (xs->state != XSK_READY || xs->umem) {
1116                         mutex_unlock(&xs->mutex);
1117                         return -EBUSY;
1118                 }
1119
1120                 umem = xdp_umem_create(&mr);
1121                 if (IS_ERR(umem)) {
1122                         mutex_unlock(&xs->mutex);
1123                         return PTR_ERR(umem);
1124                 }
1125
1126                 /* Make sure umem is ready before it can be seen by others */
1127                 smp_wmb();
1128                 WRITE_ONCE(xs->umem, umem);
1129                 mutex_unlock(&xs->mutex);
1130                 return 0;
1131         }
1132         case XDP_UMEM_FILL_RING:
1133         case XDP_UMEM_COMPLETION_RING:
1134         {
1135                 struct xsk_queue **q;
1136                 int entries;
1137
1138                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1139                         return -EFAULT;
1140
1141                 mutex_lock(&xs->mutex);
1142                 if (xs->state != XSK_READY) {
1143                         mutex_unlock(&xs->mutex);
1144                         return -EBUSY;
1145                 }
1146
1147                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1148                         &xs->cq_tmp;
1149                 err = xsk_init_queue(entries, q, true);
1150                 mutex_unlock(&xs->mutex);
1151                 return err;
1152         }
1153         default:
1154                 break;
1155         }
1156
1157         return -ENOPROTOOPT;
1158 }
1159
1160 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1161 {
1162         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1163         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1164         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1165 }
1166
1167 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1168 {
1169         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1170         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1171         ring->desc = offsetof(struct xdp_umem_ring, desc);
1172 }
1173
1174 struct xdp_statistics_v1 {
1175         __u64 rx_dropped;
1176         __u64 rx_invalid_descs;
1177         __u64 tx_invalid_descs;
1178 };
1179
1180 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1181                           char __user *optval, int __user *optlen)
1182 {
1183         struct sock *sk = sock->sk;
1184         struct xdp_sock *xs = xdp_sk(sk);
1185         int len;
1186
1187         if (level != SOL_XDP)
1188                 return -ENOPROTOOPT;
1189
1190         if (get_user(len, optlen))
1191                 return -EFAULT;
1192         if (len < 0)
1193                 return -EINVAL;
1194
1195         switch (optname) {
1196         case XDP_STATISTICS:
1197         {
1198                 struct xdp_statistics stats = {};
1199                 bool extra_stats = true;
1200                 size_t stats_size;
1201
1202                 if (len < sizeof(struct xdp_statistics_v1)) {
1203                         return -EINVAL;
1204                 } else if (len < sizeof(stats)) {
1205                         extra_stats = false;
1206                         stats_size = sizeof(struct xdp_statistics_v1);
1207                 } else {
1208                         stats_size = sizeof(stats);
1209                 }
1210
1211                 mutex_lock(&xs->mutex);
1212                 stats.rx_dropped = xs->rx_dropped;
1213                 if (extra_stats) {
1214                         stats.rx_ring_full = xs->rx_queue_full;
1215                         stats.rx_fill_ring_empty_descs =
1216                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1217                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1218                 } else {
1219                         stats.rx_dropped += xs->rx_queue_full;
1220                 }
1221                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1222                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1223                 mutex_unlock(&xs->mutex);
1224
1225                 if (copy_to_user(optval, &stats, stats_size))
1226                         return -EFAULT;
1227                 if (put_user(stats_size, optlen))
1228                         return -EFAULT;
1229
1230                 return 0;
1231         }
1232         case XDP_MMAP_OFFSETS:
1233         {
1234                 struct xdp_mmap_offsets off;
1235                 struct xdp_mmap_offsets_v1 off_v1;
1236                 bool flags_supported = true;
1237                 void *to_copy;
1238
1239                 if (len < sizeof(off_v1))
1240                         return -EINVAL;
1241                 else if (len < sizeof(off))
1242                         flags_supported = false;
1243
1244                 if (flags_supported) {
1245                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1246                          * except for the flags field added to the end.
1247                          */
1248                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1249                                                &off.rx);
1250                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1251                                                &off.tx);
1252                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1253                                                &off.fr);
1254                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1255                                                &off.cr);
1256                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1257                                                 ptrs.flags);
1258                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1259                                                 ptrs.flags);
1260                         off.fr.flags = offsetof(struct xdp_umem_ring,
1261                                                 ptrs.flags);
1262                         off.cr.flags = offsetof(struct xdp_umem_ring,
1263                                                 ptrs.flags);
1264
1265                         len = sizeof(off);
1266                         to_copy = &off;
1267                 } else {
1268                         xsk_enter_rxtx_offsets(&off_v1.rx);
1269                         xsk_enter_rxtx_offsets(&off_v1.tx);
1270                         xsk_enter_umem_offsets(&off_v1.fr);
1271                         xsk_enter_umem_offsets(&off_v1.cr);
1272
1273                         len = sizeof(off_v1);
1274                         to_copy = &off_v1;
1275                 }
1276
1277                 if (copy_to_user(optval, to_copy, len))
1278                         return -EFAULT;
1279                 if (put_user(len, optlen))
1280                         return -EFAULT;
1281
1282                 return 0;
1283         }
1284         case XDP_OPTIONS:
1285         {
1286                 struct xdp_options opts = {};
1287
1288                 if (len < sizeof(opts))
1289                         return -EINVAL;
1290
1291                 mutex_lock(&xs->mutex);
1292                 if (xs->zc)
1293                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1294                 mutex_unlock(&xs->mutex);
1295
1296                 len = sizeof(opts);
1297                 if (copy_to_user(optval, &opts, len))
1298                         return -EFAULT;
1299                 if (put_user(len, optlen))
1300                         return -EFAULT;
1301
1302                 return 0;
1303         }
1304         default:
1305                 break;
1306         }
1307
1308         return -EOPNOTSUPP;
1309 }
1310
1311 static int xsk_mmap(struct file *file, struct socket *sock,
1312                     struct vm_area_struct *vma)
1313 {
1314         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1315         unsigned long size = vma->vm_end - vma->vm_start;
1316         struct xdp_sock *xs = xdp_sk(sock->sk);
1317         struct xsk_queue *q = NULL;
1318         unsigned long pfn;
1319         struct page *qpg;
1320
1321         if (READ_ONCE(xs->state) != XSK_READY)
1322                 return -EBUSY;
1323
1324         if (offset == XDP_PGOFF_RX_RING) {
1325                 q = READ_ONCE(xs->rx);
1326         } else if (offset == XDP_PGOFF_TX_RING) {
1327                 q = READ_ONCE(xs->tx);
1328         } else {
1329                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1330                 smp_rmb();
1331                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1332                         q = READ_ONCE(xs->fq_tmp);
1333                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1334                         q = READ_ONCE(xs->cq_tmp);
1335         }
1336
1337         if (!q)
1338                 return -EINVAL;
1339
1340         /* Matches the smp_wmb() in xsk_init_queue */
1341         smp_rmb();
1342         qpg = virt_to_head_page(q->ring);
1343         if (size > page_size(qpg))
1344                 return -EINVAL;
1345
1346         pfn = virt_to_phys(q->ring) >> PAGE_SHIFT;
1347         return remap_pfn_range(vma, vma->vm_start, pfn,
1348                                size, vma->vm_page_prot);
1349 }
1350
1351 static int xsk_notifier(struct notifier_block *this,
1352                         unsigned long msg, void *ptr)
1353 {
1354         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1355         struct net *net = dev_net(dev);
1356         struct sock *sk;
1357
1358         switch (msg) {
1359         case NETDEV_UNREGISTER:
1360                 mutex_lock(&net->xdp.lock);
1361                 sk_for_each(sk, &net->xdp.list) {
1362                         struct xdp_sock *xs = xdp_sk(sk);
1363
1364                         mutex_lock(&xs->mutex);
1365                         if (xs->dev == dev) {
1366                                 sk->sk_err = ENETDOWN;
1367                                 if (!sock_flag(sk, SOCK_DEAD))
1368                                         sk_error_report(sk);
1369
1370                                 xsk_unbind_dev(xs);
1371
1372                                 /* Clear device references. */
1373                                 xp_clear_dev(xs->pool);
1374                         }
1375                         mutex_unlock(&xs->mutex);
1376                 }
1377                 mutex_unlock(&net->xdp.lock);
1378                 break;
1379         }
1380         return NOTIFY_DONE;
1381 }
1382
1383 static struct proto xsk_proto = {
1384         .name =         "XDP",
1385         .owner =        THIS_MODULE,
1386         .obj_size =     sizeof(struct xdp_sock),
1387 };
1388
1389 static const struct proto_ops xsk_proto_ops = {
1390         .family         = PF_XDP,
1391         .owner          = THIS_MODULE,
1392         .release        = xsk_release,
1393         .bind           = xsk_bind,
1394         .connect        = sock_no_connect,
1395         .socketpair     = sock_no_socketpair,
1396         .accept         = sock_no_accept,
1397         .getname        = sock_no_getname,
1398         .poll           = xsk_poll,
1399         .ioctl          = sock_no_ioctl,
1400         .listen         = sock_no_listen,
1401         .shutdown       = sock_no_shutdown,
1402         .setsockopt     = xsk_setsockopt,
1403         .getsockopt     = xsk_getsockopt,
1404         .sendmsg        = xsk_sendmsg,
1405         .recvmsg        = xsk_recvmsg,
1406         .mmap           = xsk_mmap,
1407         .sendpage       = sock_no_sendpage,
1408 };
1409
1410 static void xsk_destruct(struct sock *sk)
1411 {
1412         struct xdp_sock *xs = xdp_sk(sk);
1413
1414         if (!sock_flag(sk, SOCK_DEAD))
1415                 return;
1416
1417         if (!xp_put_pool(xs->pool))
1418                 xdp_put_umem(xs->umem, !xs->pool);
1419
1420         sk_refcnt_debug_dec(sk);
1421 }
1422
1423 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1424                       int kern)
1425 {
1426         struct xdp_sock *xs;
1427         struct sock *sk;
1428
1429         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1430                 return -EPERM;
1431         if (sock->type != SOCK_RAW)
1432                 return -ESOCKTNOSUPPORT;
1433
1434         if (protocol)
1435                 return -EPROTONOSUPPORT;
1436
1437         sock->state = SS_UNCONNECTED;
1438
1439         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1440         if (!sk)
1441                 return -ENOBUFS;
1442
1443         sock->ops = &xsk_proto_ops;
1444
1445         sock_init_data(sock, sk);
1446
1447         sk->sk_family = PF_XDP;
1448
1449         sk->sk_destruct = xsk_destruct;
1450         sk_refcnt_debug_inc(sk);
1451
1452         sock_set_flag(sk, SOCK_RCU_FREE);
1453
1454         xs = xdp_sk(sk);
1455         xs->state = XSK_READY;
1456         mutex_init(&xs->mutex);
1457         spin_lock_init(&xs->rx_lock);
1458
1459         INIT_LIST_HEAD(&xs->map_list);
1460         spin_lock_init(&xs->map_list_lock);
1461
1462         mutex_lock(&net->xdp.lock);
1463         sk_add_node_rcu(sk, &net->xdp.list);
1464         mutex_unlock(&net->xdp.lock);
1465
1466         local_bh_disable();
1467         sock_prot_inuse_add(net, &xsk_proto, 1);
1468         local_bh_enable();
1469
1470         return 0;
1471 }
1472
1473 static const struct net_proto_family xsk_family_ops = {
1474         .family = PF_XDP,
1475         .create = xsk_create,
1476         .owner  = THIS_MODULE,
1477 };
1478
1479 static struct notifier_block xsk_netdev_notifier = {
1480         .notifier_call  = xsk_notifier,
1481 };
1482
1483 static int __net_init xsk_net_init(struct net *net)
1484 {
1485         mutex_init(&net->xdp.lock);
1486         INIT_HLIST_HEAD(&net->xdp.list);
1487         return 0;
1488 }
1489
1490 static void __net_exit xsk_net_exit(struct net *net)
1491 {
1492         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1493 }
1494
1495 static struct pernet_operations xsk_net_ops = {
1496         .init = xsk_net_init,
1497         .exit = xsk_net_exit,
1498 };
1499
1500 static int __init xsk_init(void)
1501 {
1502         int err, cpu;
1503
1504         err = proto_register(&xsk_proto, 0 /* no slab */);
1505         if (err)
1506                 goto out;
1507
1508         err = sock_register(&xsk_family_ops);
1509         if (err)
1510                 goto out_proto;
1511
1512         err = register_pernet_subsys(&xsk_net_ops);
1513         if (err)
1514                 goto out_sk;
1515
1516         err = register_netdevice_notifier(&xsk_netdev_notifier);
1517         if (err)
1518                 goto out_pernet;
1519
1520         for_each_possible_cpu(cpu)
1521                 INIT_LIST_HEAD(&per_cpu(xskmap_flush_list, cpu));
1522         return 0;
1523
1524 out_pernet:
1525         unregister_pernet_subsys(&xsk_net_ops);
1526 out_sk:
1527         sock_unregister(PF_XDP);
1528 out_proto:
1529         proto_unregister(&xsk_proto);
1530 out:
1531         return err;
1532 }
1533
1534 fs_initcall(xsk_init);