net/xdp/xsk.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* XDP sockets
   3  *
   4  * AF_XDP sockets allows a channel between XDP programs and userspace
   5  * applications.
   6  * Copyright(c) 2018 Intel Corporation.
   7  *
   8  * Author(s): Björn Töpel <[email protected]>
   9  *            Magnus Karlsson <[email protected]>
  10  */
  11
  12 #define pr_fmt(fmt) "AF_XDP: %s: " fmt, __func__
  13
  14 #include <linux/if_xdp.h>
  15 #include <linux/init.h>
  16 #include <linux/sched/mm.h>
  17 #include <linux/sched/signal.h>
  18 #include <linux/sched/task.h>
  19 #include <linux/socket.h>
  20 #include <linux/file.h>
  21 #include <linux/uaccess.h>
  22 #include <linux/net.h>
  23 #include <linux/netdevice.h>
  24 #include <linux/rculist.h>
  25 #include <linux/vmalloc.h>
  26 #include <net/xdp_sock_drv.h>
  27 #include <net/busy_poll.h>
  28 #include <net/netdev_rx_queue.h>
  29 #include <net/xdp.h>
  30
  31 #include "xsk_queue.h"
  32 #include "xdp_umem.h"
  33 #include "xsk.h"
  34
  35 #define TX_BATCH_SIZE 32
  36 #define MAX_PER_SOCKET_BUDGET (TX_BATCH_SIZE)
  37
  38 void xsk_set_rx_need_wakeup(struct xsk_buff_pool *pool)
  39 {
  40         if (pool->cached_need_wakeup & XDP_WAKEUP_RX)
  41                 return;
  42
  43         pool->fq->ring->flags |= XDP_RING_NEED_WAKEUP;
  44         pool->cached_need_wakeup |= XDP_WAKEUP_RX;
  45 }
  46 EXPORT_SYMBOL(xsk_set_rx_need_wakeup);
  47
  48 void xsk_set_tx_need_wakeup(struct xsk_buff_pool *pool)
  49 {
  50         struct xdp_sock *xs;
  51
  52         if (pool->cached_need_wakeup & XDP_WAKEUP_TX)
  53                 return;
  54
  55         rcu_read_lock();
  56         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
  57                 xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
  58         }
  59         rcu_read_unlock();
  60
  61         pool->cached_need_wakeup |= XDP_WAKEUP_TX;
  62 }
  63 EXPORT_SYMBOL(xsk_set_tx_need_wakeup);
  64
  65 void xsk_clear_rx_need_wakeup(struct xsk_buff_pool *pool)
  66 {
  67         if (!(pool->cached_need_wakeup & XDP_WAKEUP_RX))
  68                 return;
  69
  70         pool->fq->ring->flags &= ~XDP_RING_NEED_WAKEUP;
  71         pool->cached_need_wakeup &= ~XDP_WAKEUP_RX;
  72 }
  73 EXPORT_SYMBOL(xsk_clear_rx_need_wakeup);
  74
  75 void xsk_clear_tx_need_wakeup(struct xsk_buff_pool *pool)
  76 {
  77         struct xdp_sock *xs;
  78
  79         if (!(pool->cached_need_wakeup & XDP_WAKEUP_TX))
  80                 return;
  81
  82         rcu_read_lock();
  83         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
  84                 xs->tx->ring->flags &= ~XDP_RING_NEED_WAKEUP;
  85         }
  86         rcu_read_unlock();
  87
  88         pool->cached_need_wakeup &= ~XDP_WAKEUP_TX;
  89 }
  90 EXPORT_SYMBOL(xsk_clear_tx_need_wakeup);
  91
  92 bool xsk_uses_need_wakeup(struct xsk_buff_pool *pool)
  93 {
  94         return pool->uses_need_wakeup;
  95 }
  96 EXPORT_SYMBOL(xsk_uses_need_wakeup);
  97
  98 struct xsk_buff_pool *xsk_get_pool_from_qid(struct net_device *dev,
  99                                             u16 queue_id)
 100 {
 101         if (queue_id < dev->real_num_rx_queues)
 102                 return dev->_rx[queue_id].pool;
 103         if (queue_id < dev->real_num_tx_queues)
 104                 return dev->_tx[queue_id].pool;
 105
 106         return NULL;
 107 }
 108 EXPORT_SYMBOL(xsk_get_pool_from_qid);
 109
 110 void xsk_clear_pool_at_qid(struct net_device *dev, u16 queue_id)
 111 {
 112         if (queue_id < dev->num_rx_queues)
 113                 dev->_rx[queue_id].pool = NULL;
 114         if (queue_id < dev->num_tx_queues)
 115                 dev->_tx[queue_id].pool = NULL;
 116 }
 117
 118 /* The buffer pool is stored both in the _rx struct and the _tx struct as we do
 119  * not know if the device has more tx queues than rx, or the opposite.
 120  * This might also change during run time.
 121  */
 122 int xsk_reg_pool_at_qid(struct net_device *dev, struct xsk_buff_pool *pool,
 123                         u16 queue_id)
 124 {
 125         if (queue_id >= max_t(unsigned int,
 126                               dev->real_num_rx_queues,
 127                               dev->real_num_tx_queues))
 128                 return -EINVAL;
 129
 130         if (queue_id < dev->real_num_rx_queues)
 131                 dev->_rx[queue_id].pool = pool;
 132         if (queue_id < dev->real_num_tx_queues)
 133                 dev->_tx[queue_id].pool = pool;
 134
 135         return 0;
 136 }
 137
 138 static int __xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff_xsk *xskb, u32 len,
 139                         u32 flags)
 140 {
 141         u64 addr;
 142         int err;
 143
 144         addr = xp_get_handle(xskb, xskb->pool);
 145         err = xskq_prod_reserve_desc(xs->rx, addr, len, flags);
 146         if (err) {
 147                 xs->rx_queue_full++;
 148                 return err;
 149         }
 150
 151         xp_release(xskb);
 152         return 0;
 153 }
 154
 155 static int xsk_rcv_zc(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
 156 {
 157         struct xdp_buff_xsk *xskb = container_of(xdp, struct xdp_buff_xsk, xdp);
 158         u32 frags = xdp_buff_has_frags(xdp);
 159         struct xdp_buff_xsk *pos, *tmp;
 160         struct list_head *xskb_list;
 161         u32 contd = 0;
 162         int err;
 163
 164         if (frags)
 165                 contd = XDP_PKT_CONTD;
 166
 167         err = __xsk_rcv_zc(xs, xskb, len, contd);
 168         if (err)
 169                 goto err;
 170         if (likely(!frags))
 171                 return 0;
 172
 173         xskb_list = &xskb->pool->xskb_list;
 174         list_for_each_entry_safe(pos, tmp, xskb_list, list_node) {
 175                 if (list_is_singular(xskb_list))
 176                         contd = 0;
 177                 len = pos->xdp.data_end - pos->xdp.data;
 178                 err = __xsk_rcv_zc(xs, pos, len, contd);
 179                 if (err)
 180                         goto err;
 181                 list_del(&pos->list_node);
 182         }
 183
 184         return 0;
 185 err:
 186         xsk_buff_free(xdp);
 187         return err;
 188 }
 189
 190 static void *xsk_copy_xdp_start(struct xdp_buff *from)
 191 {
 192         if (unlikely(xdp_data_meta_unsupported(from)))
 193                 return from->data;
 194         else
 195                 return from->data_meta;
 196 }
 197
 198 static u32 xsk_copy_xdp(void *to, void **from, u32 to_len,
 199                         u32 *from_len, skb_frag_t **frag, u32 rem)
 200 {
 201         u32 copied = 0;
 202
 203         while (1) {
 204                 u32 copy_len = min_t(u32, *from_len, to_len);
 205
 206                 memcpy(to, *from, copy_len);
 207                 copied += copy_len;
 208                 if (rem == copied)
 209                         return copied;
 210
 211                 if (*from_len == copy_len) {
 212                         *from = skb_frag_address(*frag);
 213                         *from_len = skb_frag_size((*frag)++);
 214                 } else {
 215                         *from += copy_len;
 216                         *from_len -= copy_len;
 217                 }
 218                 if (to_len == copy_len)
 219                         return copied;
 220
 221                 to_len -= copy_len;
 222                 to += copy_len;
 223         }
 224 }
 225
 226 static int __xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
 227 {
 228         u32 frame_size = xsk_pool_get_rx_frame_size(xs->pool);
 229         void *copy_from = xsk_copy_xdp_start(xdp), *copy_to;
 230         u32 from_len, meta_len, rem, num_desc;
 231         struct xdp_buff_xsk *xskb;
 232         struct xdp_buff *xsk_xdp;
 233         skb_frag_t *frag;
 234
 235         from_len = xdp->data_end - copy_from;
 236         meta_len = xdp->data - copy_from;
 237         rem = len + meta_len;
 238
 239         if (len <= frame_size && !xdp_buff_has_frags(xdp)) {
 240                 int err;
 241
 242                 xsk_xdp = xsk_buff_alloc(xs->pool);
 243                 if (!xsk_xdp) {
 244                         xs->rx_dropped++;
 245                         return -ENOMEM;
 246                 }
 247                 memcpy(xsk_xdp->data - meta_len, copy_from, rem);
 248                 xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp);
 249                 err = __xsk_rcv_zc(xs, xskb, len, 0);
 250                 if (err) {
 251                         xsk_buff_free(xsk_xdp);
 252                         return err;
 253                 }
 254
 255                 return 0;
 256         }
 257
 258         num_desc = (len - 1) / frame_size + 1;
 259
 260         if (!xsk_buff_can_alloc(xs->pool, num_desc)) {
 261                 xs->rx_dropped++;
 262                 return -ENOMEM;
 263         }
 264         if (xskq_prod_nb_free(xs->rx, num_desc) < num_desc) {
 265                 xs->rx_queue_full++;
 266                 return -ENOBUFS;
 267         }
 268
 269         if (xdp_buff_has_frags(xdp)) {
 270                 struct skb_shared_info *sinfo;
 271
 272                 sinfo = xdp_get_shared_info_from_buff(xdp);
 273                 frag =  &sinfo->frags[0];
 274         }
 275
 276         do {
 277                 u32 to_len = frame_size + meta_len;
 278                 u32 copied;
 279
 280                 xsk_xdp = xsk_buff_alloc(xs->pool);
 281                 copy_to = xsk_xdp->data - meta_len;
 282
 283                 copied = xsk_copy_xdp(copy_to, &copy_from, to_len, &from_len, &frag, rem);
 284                 rem -= copied;
 285
 286                 xskb = container_of(xsk_xdp, struct xdp_buff_xsk, xdp);
 287                 __xsk_rcv_zc(xs, xskb, copied - meta_len, rem ? XDP_PKT_CONTD : 0);
 288                 meta_len = 0;
 289         } while (rem);
 290
 291         return 0;
 292 }
 293
 294 static bool xsk_tx_writeable(struct xdp_sock *xs)
 295 {
 296         if (xskq_cons_present_entries(xs->tx) > xs->tx->nentries / 2)
 297                 return false;
 298
 299         return true;
 300 }
 301
 302 static bool xsk_is_bound(struct xdp_sock *xs)
 303 {
 304         if (READ_ONCE(xs->state) == XSK_BOUND) {
 305                 /* Matches smp_wmb() in bind(). */
 306                 smp_rmb();
 307                 return true;
 308         }
 309         return false;
 310 }
 311
 312 static int xsk_rcv_check(struct xdp_sock *xs, struct xdp_buff *xdp, u32 len)
 313 {
 314         if (!xsk_is_bound(xs))
 315                 return -ENXIO;
 316
 317         if (xs->dev != xdp->rxq->dev || xs->queue_id != xdp->rxq->queue_index)
 318                 return -EINVAL;
 319
 320         if (len > xsk_pool_get_rx_frame_size(xs->pool) && !xs->sg) {
 321                 xs->rx_dropped++;
 322                 return -ENOSPC;
 323         }
 324
 325         sk_mark_napi_id_once_xdp(&xs->sk, xdp);
 326         return 0;
 327 }
 328
 329 static void xsk_flush(struct xdp_sock *xs)
 330 {
 331         xskq_prod_submit(xs->rx);
 332         __xskq_cons_release(xs->pool->fq);
 333         sock_def_readable(&xs->sk);
 334 }
 335
 336 int xsk_generic_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 337 {
 338         u32 len = xdp_get_buff_len(xdp);
 339         int err;
 340
 341         spin_lock_bh(&xs->rx_lock);
 342         err = xsk_rcv_check(xs, xdp, len);
 343         if (!err) {
 344                 err = __xsk_rcv(xs, xdp, len);
 345                 xsk_flush(xs);
 346         }
 347         spin_unlock_bh(&xs->rx_lock);
 348         return err;
 349 }
 350
 351 static int xsk_rcv(struct xdp_sock *xs, struct xdp_buff *xdp)
 352 {
 353         u32 len = xdp_get_buff_len(xdp);
 354         int err;
 355
 356         err = xsk_rcv_check(xs, xdp, len);
 357         if (err)
 358                 return err;
 359
 360         if (xdp->rxq->mem.type == MEM_TYPE_XSK_BUFF_POOL) {
 361                 len = xdp->data_end - xdp->data;
 362                 return xsk_rcv_zc(xs, xdp, len);
 363         }
 364
 365         err = __xsk_rcv(xs, xdp, len);
 366         if (!err)
 367                 xdp_return_buff(xdp);
 368         return err;
 369 }
 370
 371 int __xsk_map_redirect(struct xdp_sock *xs, struct xdp_buff *xdp)
 372 {
 373         int err;
 374
 375         err = xsk_rcv(xs, xdp);
 376         if (err)
 377                 return err;
 378
 379         if (!xs->flush_node.prev) {
 380                 struct list_head *flush_list = bpf_net_ctx_get_xskmap_flush_list();
 381
 382                 list_add(&xs->flush_node, flush_list);
 383         }
 384
 385         return 0;
 386 }
 387
 388 void __xsk_map_flush(struct list_head *flush_list)
 389 {
 390         struct xdp_sock *xs, *tmp;
 391
 392         list_for_each_entry_safe(xs, tmp, flush_list, flush_node) {
 393                 xsk_flush(xs);
 394                 __list_del_clearprev(&xs->flush_node);
 395         }
 396 }
 397
 398 void xsk_tx_completed(struct xsk_buff_pool *pool, u32 nb_entries)
 399 {
 400         xskq_prod_submit_n(pool->cq, nb_entries);
 401 }
 402 EXPORT_SYMBOL(xsk_tx_completed);
 403
 404 void xsk_tx_release(struct xsk_buff_pool *pool)
 405 {
 406         struct xdp_sock *xs;
 407
 408         rcu_read_lock();
 409         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
 410                 __xskq_cons_release(xs->tx);
 411                 if (xsk_tx_writeable(xs))
 412                         xs->sk.sk_write_space(&xs->sk);
 413         }
 414         rcu_read_unlock();
 415 }
 416 EXPORT_SYMBOL(xsk_tx_release);
 417
 418 bool xsk_tx_peek_desc(struct xsk_buff_pool *pool, struct xdp_desc *desc)
 419 {
 420         bool budget_exhausted = false;
 421         struct xdp_sock *xs;
 422
 423         rcu_read_lock();
 424 again:
 425         list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list) {
 426                 if (xs->tx_budget_spent >= MAX_PER_SOCKET_BUDGET) {
 427                         budget_exhausted = true;
 428                         continue;
 429                 }
 430
 431                 if (!xskq_cons_peek_desc(xs->tx, desc, pool)) {
 432                         if (xskq_has_descs(xs->tx))
 433                                 xskq_cons_release(xs->tx);
 434                         continue;
 435                 }
 436
 437                 xs->tx_budget_spent++;
 438
 439                 /* This is the backpressure mechanism for the Tx path.
 440                  * Reserve space in the completion queue and only proceed
 441                  * if there is space in it. This avoids having to implement
 442                  * any buffering in the Tx path.
 443                  */
 444                 if (xskq_prod_reserve_addr(pool->cq, desc->addr))
 445                         goto out;
 446
 447                 xskq_cons_release(xs->tx);
 448                 rcu_read_unlock();
 449                 return true;
 450         }
 451
 452         if (budget_exhausted) {
 453                 list_for_each_entry_rcu(xs, &pool->xsk_tx_list, tx_list)
 454                         xs->tx_budget_spent = 0;
 455
 456                 budget_exhausted = false;
 457                 goto again;
 458         }
 459
 460 out:
 461         rcu_read_unlock();
 462         return false;
 463 }
 464 EXPORT_SYMBOL(xsk_tx_peek_desc);
 465
 466 static u32 xsk_tx_peek_release_fallback(struct xsk_buff_pool *pool, u32 max_entries)
 467 {
 468         struct xdp_desc *descs = pool->tx_descs;
 469         u32 nb_pkts = 0;
 470
 471         while (nb_pkts < max_entries && xsk_tx_peek_desc(pool, &descs[nb_pkts]))
 472                 nb_pkts++;
 473
 474         xsk_tx_release(pool);
 475         return nb_pkts;
 476 }
 477
 478 u32 xsk_tx_peek_release_desc_batch(struct xsk_buff_pool *pool, u32 nb_pkts)
 479 {
 480         struct xdp_sock *xs;
 481
 482         rcu_read_lock();
 483         if (!list_is_singular(&pool->xsk_tx_list)) {
 484                 /* Fallback to the non-batched version */
 485                 rcu_read_unlock();
 486                 return xsk_tx_peek_release_fallback(pool, nb_pkts);
 487         }
 488
 489         xs = list_first_or_null_rcu(&pool->xsk_tx_list, struct xdp_sock, tx_list);
 490         if (!xs) {
 491                 nb_pkts = 0;
 492                 goto out;
 493         }
 494
 495         nb_pkts = xskq_cons_nb_entries(xs->tx, nb_pkts);
 496
 497         /* This is the backpressure mechanism for the Tx path. Try to
 498          * reserve space in the completion queue for all packets, but
 499          * if there are fewer slots available, just process that many
 500          * packets. This avoids having to implement any buffering in
 501          * the Tx path.
 502          */
 503         nb_pkts = xskq_prod_nb_free(pool->cq, nb_pkts);
 504         if (!nb_pkts)
 505                 goto out;
 506
 507         nb_pkts = xskq_cons_read_desc_batch(xs->tx, pool, nb_pkts);
 508         if (!nb_pkts) {
 509                 xs->tx->queue_empty_descs++;
 510                 goto out;
 511         }
 512
 513         __xskq_cons_release(xs->tx);
 514         xskq_prod_write_addr_batch(pool->cq, pool->tx_descs, nb_pkts);
 515         xs->sk.sk_write_space(&xs->sk);
 516
 517 out:
 518         rcu_read_unlock();
 519         return nb_pkts;
 520 }
 521 EXPORT_SYMBOL(xsk_tx_peek_release_desc_batch);
 522
 523 static int xsk_wakeup(struct xdp_sock *xs, u8 flags)
 524 {
 525         struct net_device *dev = xs->dev;
 526
 527         return dev->netdev_ops->ndo_xsk_wakeup(dev, xs->queue_id, flags);
 528 }
 529
 530 static int xsk_cq_reserve_addr_locked(struct xsk_buff_pool *pool, u64 addr)
 531 {
 532         unsigned long flags;
 533         int ret;
 534
 535         spin_lock_irqsave(&pool->cq_lock, flags);
 536         ret = xskq_prod_reserve_addr(pool->cq, addr);
 537         spin_unlock_irqrestore(&pool->cq_lock, flags);
 538
 539         return ret;
 540 }
 541
 542 static void xsk_cq_submit_locked(struct xsk_buff_pool *pool, u32 n)
 543 {
 544         unsigned long flags;
 545
 546         spin_lock_irqsave(&pool->cq_lock, flags);
 547         xskq_prod_submit_n(pool->cq, n);
 548         spin_unlock_irqrestore(&pool->cq_lock, flags);
 549 }
 550
 551 static void xsk_cq_cancel_locked(struct xsk_buff_pool *pool, u32 n)
 552 {
 553         unsigned long flags;
 554
 555         spin_lock_irqsave(&pool->cq_lock, flags);
 556         xskq_prod_cancel_n(pool->cq, n);
 557         spin_unlock_irqrestore(&pool->cq_lock, flags);
 558 }
 559
 560 static u32 xsk_get_num_desc(struct sk_buff *skb)
 561 {
 562         return skb ? (long)skb_shinfo(skb)->destructor_arg : 0;
 563 }
 564
 565 static void xsk_destruct_skb(struct sk_buff *skb)
 566 {
 567         struct xsk_tx_metadata_compl *compl = &skb_shinfo(skb)->xsk_meta;
 568
 569         if (compl->tx_timestamp) {
 570                 /* sw completion timestamp, not a real one */
 571                 *compl->tx_timestamp = ktime_get_tai_fast_ns();
 572         }
 573
 574         xsk_cq_submit_locked(xdp_sk(skb->sk)->pool, xsk_get_num_desc(skb));
 575         sock_wfree(skb);
 576 }
 577
 578 static void xsk_set_destructor_arg(struct sk_buff *skb)
 579 {
 580         long num = xsk_get_num_desc(xdp_sk(skb->sk)->skb) + 1;
 581
 582         skb_shinfo(skb)->destructor_arg = (void *)num;
 583 }
 584
 585 static void xsk_consume_skb(struct sk_buff *skb)
 586 {
 587         struct xdp_sock *xs = xdp_sk(skb->sk);
 588
 589         skb->destructor = sock_wfree;
 590         xsk_cq_cancel_locked(xs->pool, xsk_get_num_desc(skb));
 591         /* Free skb without triggering the perf drop trace */
 592         consume_skb(skb);
 593         xs->skb = NULL;
 594 }
 595
 596 static void xsk_drop_skb(struct sk_buff *skb)
 597 {
 598         xdp_sk(skb->sk)->tx->invalid_descs += xsk_get_num_desc(skb);
 599         xsk_consume_skb(skb);
 600 }
 601
 602 static struct sk_buff *xsk_build_skb_zerocopy(struct xdp_sock *xs,
 603                                               struct xdp_desc *desc)
 604 {
 605         struct xsk_buff_pool *pool = xs->pool;
 606         u32 hr, len, ts, offset, copy, copied;
 607         struct sk_buff *skb = xs->skb;
 608         struct page *page;
 609         void *buffer;
 610         int err, i;
 611         u64 addr;
 612
 613         if (!skb) {
 614                 hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(xs->dev->needed_headroom));
 615
 616                 skb = sock_alloc_send_skb(&xs->sk, hr, 1, &err);
 617                 if (unlikely(!skb))
 618                         return ERR_PTR(err);
 619
 620                 skb_reserve(skb, hr);
 621         }
 622
 623         addr = desc->addr;
 624         len = desc->len;
 625         ts = pool->unaligned ? len : pool->chunk_size;
 626
 627         buffer = xsk_buff_raw_get_data(pool, addr);
 628         offset = offset_in_page(buffer);
 629         addr = buffer - pool->addrs;
 630
 631         for (copied = 0, i = skb_shinfo(skb)->nr_frags; copied < len; i++) {
 632                 if (unlikely(i >= MAX_SKB_FRAGS))
 633                         return ERR_PTR(-EOVERFLOW);
 634
 635                 page = pool->umem->pgs[addr >> PAGE_SHIFT];
 636                 get_page(page);
 637
 638                 copy = min_t(u32, PAGE_SIZE - offset, len - copied);
 639                 skb_fill_page_desc(skb, i, page, offset, copy);
 640
 641                 copied += copy;
 642                 addr += copy;
 643                 offset = 0;
 644         }
 645
 646         skb->len += len;
 647         skb->data_len += len;
 648         skb->truesize += ts;
 649
 650         refcount_add(ts, &xs->sk.sk_wmem_alloc);
 651
 652         return skb;
 653 }
 654
 655 static struct sk_buff *xsk_build_skb(struct xdp_sock *xs,
 656                                      struct xdp_desc *desc)
 657 {
 658         struct xsk_tx_metadata *meta = NULL;
 659         struct net_device *dev = xs->dev;
 660         struct sk_buff *skb = xs->skb;
 661         bool first_frag = false;
 662         int err;
 663
 664         if (dev->priv_flags & IFF_TX_SKB_NO_LINEAR) {
 665                 skb = xsk_build_skb_zerocopy(xs, desc);
 666                 if (IS_ERR(skb)) {
 667                         err = PTR_ERR(skb);
 668                         goto free_err;
 669                 }
 670         } else {
 671                 u32 hr, tr, len;
 672                 void *buffer;
 673
 674                 buffer = xsk_buff_raw_get_data(xs->pool, desc->addr);
 675                 len = desc->len;
 676
 677                 if (!skb) {
 678                         first_frag = true;
 679
 680                         hr = max(NET_SKB_PAD, L1_CACHE_ALIGN(dev->needed_headroom));
 681                         tr = dev->needed_tailroom;
 682                         skb = sock_alloc_send_skb(&xs->sk, hr + len + tr, 1, &err);
 683                         if (unlikely(!skb))
 684                                 goto free_err;
 685
 686                         skb_reserve(skb, hr);
 687                         skb_put(skb, len);
 688
 689                         err = skb_store_bits(skb, 0, buffer, len);
 690                         if (unlikely(err))
 691                                 goto free_err;
 692                 } else {
 693                         int nr_frags = skb_shinfo(skb)->nr_frags;
 694                         struct page *page;
 695                         u8 *vaddr;
 696
 697                         if (unlikely(nr_frags == (MAX_SKB_FRAGS - 1) && xp_mb_desc(desc))) {
 698                                 err = -EOVERFLOW;
 699                                 goto free_err;
 700                         }
 701
 702                         page = alloc_page(xs->sk.sk_allocation);
 703                         if (unlikely(!page)) {
 704                                 err = -EAGAIN;
 705                                 goto free_err;
 706                         }
 707
 708                         vaddr = kmap_local_page(page);
 709                         memcpy(vaddr, buffer, len);
 710                         kunmap_local(vaddr);
 711
 712                         skb_add_rx_frag(skb, nr_frags, page, 0, len, PAGE_SIZE);
 713                         refcount_add(PAGE_SIZE, &xs->sk.sk_wmem_alloc);
 714                 }
 715
 716                 if (first_frag && desc->options & XDP_TX_METADATA) {
 717                         if (unlikely(xs->pool->tx_metadata_len == 0)) {
 718                                 err = -EINVAL;
 719                                 goto free_err;
 720                         }
 721
 722                         meta = buffer - xs->pool->tx_metadata_len;
 723                         if (unlikely(!xsk_buff_valid_tx_metadata(meta))) {
 724                                 err = -EINVAL;
 725                                 goto free_err;
 726                         }
 727
 728                         if (meta->flags & XDP_TXMD_FLAGS_CHECKSUM) {
 729                                 if (unlikely(meta->request.csum_start +
 730                                              meta->request.csum_offset +
 731                                              sizeof(__sum16) > len)) {
 732                                         err = -EINVAL;
 733                                         goto free_err;
 734                                 }
 735
 736                                 skb->csum_start = hr + meta->request.csum_start;
 737                                 skb->csum_offset = meta->request.csum_offset;
 738                                 skb->ip_summed = CHECKSUM_PARTIAL;
 739
 740                                 if (unlikely(xs->pool->tx_sw_csum)) {
 741                                         err = skb_checksum_help(skb);
 742                                         if (err)
 743                                                 goto free_err;
 744                                 }
 745                         }
 746                 }
 747         }
 748
 749         skb->dev = dev;
 750         skb->priority = READ_ONCE(xs->sk.sk_priority);
 751         skb->mark = READ_ONCE(xs->sk.sk_mark);
 752         skb->destructor = xsk_destruct_skb;
 753         xsk_tx_metadata_to_compl(meta, &skb_shinfo(skb)->xsk_meta);
 754         xsk_set_destructor_arg(skb);
 755
 756         return skb;
 757
 758 free_err:
 759         if (first_frag && skb)
 760                 kfree_skb(skb);
 761
 762         if (err == -EOVERFLOW) {
 763                 /* Drop the packet */
 764                 xsk_set_destructor_arg(xs->skb);
 765                 xsk_drop_skb(xs->skb);
 766                 xskq_cons_release(xs->tx);
 767         } else {
 768                 /* Let application retry */
 769                 xsk_cq_cancel_locked(xs->pool, 1);
 770         }
 771
 772         return ERR_PTR(err);
 773 }
 774
 775 static int __xsk_generic_xmit(struct sock *sk)
 776 {
 777         struct xdp_sock *xs = xdp_sk(sk);
 778         u32 max_batch = TX_BATCH_SIZE;
 779         bool sent_frame = false;
 780         struct xdp_desc desc;
 781         struct sk_buff *skb;
 782         int err = 0;
 783
 784         mutex_lock(&xs->mutex);
 785
 786         /* Since we dropped the RCU read lock, the socket state might have changed. */
 787         if (unlikely(!xsk_is_bound(xs))) {
 788                 err = -ENXIO;
 789                 goto out;
 790         }
 791
 792         if (xs->queue_id >= xs->dev->real_num_tx_queues)
 793                 goto out;
 794
 795         while (xskq_cons_peek_desc(xs->tx, &desc, xs->pool)) {
 796                 if (max_batch-- == 0) {
 797                         err = -EAGAIN;
 798                         goto out;
 799                 }
 800
 801                 /* This is the backpressure mechanism for the Tx path.
 802                  * Reserve space in the completion queue and only proceed
 803                  * if there is space in it. This avoids having to implement
 804                  * any buffering in the Tx path.
 805                  */
 806                 if (xsk_cq_reserve_addr_locked(xs->pool, desc.addr))
 807                         goto out;
 808
 809                 skb = xsk_build_skb(xs, &desc);
 810                 if (IS_ERR(skb)) {
 811                         err = PTR_ERR(skb);
 812                         if (err != -EOVERFLOW)
 813                                 goto out;
 814                         err = 0;
 815                         continue;
 816                 }
 817
 818                 xskq_cons_release(xs->tx);
 819
 820                 if (xp_mb_desc(&desc)) {
 821                         xs->skb = skb;
 822                         continue;
 823                 }
 824
 825                 err = __dev_direct_xmit(skb, xs->queue_id);
 826                 if  (err == NETDEV_TX_BUSY) {
 827                         /* Tell user-space to retry the send */
 828                         xskq_cons_cancel_n(xs->tx, xsk_get_num_desc(skb));
 829                         xsk_consume_skb(skb);
 830                         err = -EAGAIN;
 831                         goto out;
 832                 }
 833
 834                 /* Ignore NET_XMIT_CN as packet might have been sent */
 835                 if (err == NET_XMIT_DROP) {
 836                         /* SKB completed but not sent */
 837                         err = -EBUSY;
 838                         xs->skb = NULL;
 839                         goto out;
 840                 }
 841
 842                 sent_frame = true;
 843                 xs->skb = NULL;
 844         }
 845
 846         if (xskq_has_descs(xs->tx)) {
 847                 if (xs->skb)
 848                         xsk_drop_skb(xs->skb);
 849                 xskq_cons_release(xs->tx);
 850         }
 851
 852 out:
 853         if (sent_frame)
 854                 if (xsk_tx_writeable(xs))
 855                         sk->sk_write_space(sk);
 856
 857         mutex_unlock(&xs->mutex);
 858         return err;
 859 }
 860
 861 static int xsk_generic_xmit(struct sock *sk)
 862 {
 863         int ret;
 864
 865         /* Drop the RCU lock since the SKB path might sleep. */
 866         rcu_read_unlock();
 867         ret = __xsk_generic_xmit(sk);
 868         /* Reaquire RCU lock before going into common code. */
 869         rcu_read_lock();
 870
 871         return ret;
 872 }
 873
 874 static bool xsk_no_wakeup(struct sock *sk)
 875 {
 876 #ifdef CONFIG_NET_RX_BUSY_POLL
 877         /* Prefer busy-polling, skip the wakeup. */
 878         return READ_ONCE(sk->sk_prefer_busy_poll) && READ_ONCE(sk->sk_ll_usec) &&
 879                 READ_ONCE(sk->sk_napi_id) >= MIN_NAPI_ID;
 880 #else
 881         return false;
 882 #endif
 883 }
 884
 885 static int xsk_check_common(struct xdp_sock *xs)
 886 {
 887         if (unlikely(!xsk_is_bound(xs)))
 888                 return -ENXIO;
 889         if (unlikely(!(xs->dev->flags & IFF_UP)))
 890                 return -ENETDOWN;
 891
 892         return 0;
 893 }
 894
 895 static int __xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 896 {
 897         bool need_wait = !(m->msg_flags & MSG_DONTWAIT);
 898         struct sock *sk = sock->sk;
 899         struct xdp_sock *xs = xdp_sk(sk);
 900         struct xsk_buff_pool *pool;
 901         int err;
 902
 903         err = xsk_check_common(xs);
 904         if (err)
 905                 return err;
 906         if (unlikely(need_wait))
 907                 return -EOPNOTSUPP;
 908         if (unlikely(!xs->tx))
 909                 return -ENOBUFS;
 910
 911         if (sk_can_busy_loop(sk)) {
 912                 if (xs->zc)
 913                         __sk_mark_napi_id_once(sk, xsk_pool_get_napi_id(xs->pool));
 914                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 915         }
 916
 917         if (xs->zc && xsk_no_wakeup(sk))
 918                 return 0;
 919
 920         pool = xs->pool;
 921         if (pool->cached_need_wakeup & XDP_WAKEUP_TX) {
 922                 if (xs->zc)
 923                         return xsk_wakeup(xs, XDP_WAKEUP_TX);
 924                 return xsk_generic_xmit(sk);
 925         }
 926         return 0;
 927 }
 928
 929 static int xsk_sendmsg(struct socket *sock, struct msghdr *m, size_t total_len)
 930 {
 931         int ret;
 932
 933         rcu_read_lock();
 934         ret = __xsk_sendmsg(sock, m, total_len);
 935         rcu_read_unlock();
 936
 937         return ret;
 938 }
 939
 940 static int __xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 941 {
 942         bool need_wait = !(flags & MSG_DONTWAIT);
 943         struct sock *sk = sock->sk;
 944         struct xdp_sock *xs = xdp_sk(sk);
 945         int err;
 946
 947         err = xsk_check_common(xs);
 948         if (err)
 949                 return err;
 950         if (unlikely(!xs->rx))
 951                 return -ENOBUFS;
 952         if (unlikely(need_wait))
 953                 return -EOPNOTSUPP;
 954
 955         if (sk_can_busy_loop(sk))
 956                 sk_busy_loop(sk, 1); /* only support non-blocking sockets */
 957
 958         if (xsk_no_wakeup(sk))
 959                 return 0;
 960
 961         if (xs->pool->cached_need_wakeup & XDP_WAKEUP_RX && xs->zc)
 962                 return xsk_wakeup(xs, XDP_WAKEUP_RX);
 963         return 0;
 964 }
 965
 966 static int xsk_recvmsg(struct socket *sock, struct msghdr *m, size_t len, int flags)
 967 {
 968         int ret;
 969
 970         rcu_read_lock();
 971         ret = __xsk_recvmsg(sock, m, len, flags);
 972         rcu_read_unlock();
 973
 974         return ret;
 975 }
 976
 977 static __poll_t xsk_poll(struct file *file, struct socket *sock,
 978                              struct poll_table_struct *wait)
 979 {
 980         __poll_t mask = 0;
 981         struct sock *sk = sock->sk;
 982         struct xdp_sock *xs = xdp_sk(sk);
 983         struct xsk_buff_pool *pool;
 984
 985         sock_poll_wait(file, sock, wait);
 986
 987         rcu_read_lock();
 988         if (xsk_check_common(xs))
 989                 goto out;
 990
 991         pool = xs->pool;
 992
 993         if (pool->cached_need_wakeup) {
 994                 if (xs->zc)
 995                         xsk_wakeup(xs, pool->cached_need_wakeup);
 996                 else if (xs->tx)
 997                         /* Poll needs to drive Tx also in copy mode */
 998                         xsk_generic_xmit(sk);
 999         }
1000
1001         if (xs->rx && !xskq_prod_is_empty(xs->rx))
1002                 mask |= EPOLLIN | EPOLLRDNORM;
1003         if (xs->tx && xsk_tx_writeable(xs))
1004                 mask |= EPOLLOUT | EPOLLWRNORM;
1005 out:
1006         rcu_read_unlock();
1007         return mask;
1008 }
1009
1010 static int xsk_init_queue(u32 entries, struct xsk_queue **queue,
1011                           bool umem_queue)
1012 {
1013         struct xsk_queue *q;
1014
1015         if (entries == 0 || *queue || !is_power_of_2(entries))
1016                 return -EINVAL;
1017
1018         q = xskq_create(entries, umem_queue);
1019         if (!q)
1020                 return -ENOMEM;
1021
1022         /* Make sure queue is ready before it can be seen by others */
1023         smp_wmb();
1024         WRITE_ONCE(*queue, q);
1025         return 0;
1026 }
1027
1028 static void xsk_unbind_dev(struct xdp_sock *xs)
1029 {
1030         struct net_device *dev = xs->dev;
1031
1032         if (xs->state != XSK_BOUND)
1033                 return;
1034         WRITE_ONCE(xs->state, XSK_UNBOUND);
1035
1036         /* Wait for driver to stop using the xdp socket. */
1037         xp_del_xsk(xs->pool, xs);
1038         synchronize_net();
1039         dev_put(dev);
1040 }
1041
1042 static struct xsk_map *xsk_get_map_list_entry(struct xdp_sock *xs,
1043                                               struct xdp_sock __rcu ***map_entry)
1044 {
1045         struct xsk_map *map = NULL;
1046         struct xsk_map_node *node;
1047
1048         *map_entry = NULL;
1049
1050         spin_lock_bh(&xs->map_list_lock);
1051         node = list_first_entry_or_null(&xs->map_list, struct xsk_map_node,
1052                                         node);
1053         if (node) {
1054                 bpf_map_inc(&node->map->map);
1055                 map = node->map;
1056                 *map_entry = node->map_entry;
1057         }
1058         spin_unlock_bh(&xs->map_list_lock);
1059         return map;
1060 }
1061
1062 static void xsk_delete_from_maps(struct xdp_sock *xs)
1063 {
1064         /* This function removes the current XDP socket from all the
1065          * maps it resides in. We need to take extra care here, due to
1066          * the two locks involved. Each map has a lock synchronizing
1067          * updates to the entries, and each socket has a lock that
1068          * synchronizes access to the list of maps (map_list). For
1069          * deadlock avoidance the locks need to be taken in the order
1070          * "map lock"->"socket map list lock". We start off by
1071          * accessing the socket map list, and take a reference to the
1072          * map to guarantee existence between the
1073          * xsk_get_map_list_entry() and xsk_map_try_sock_delete()
1074          * calls. Then we ask the map to remove the socket, which
1075          * tries to remove the socket from the map. Note that there
1076          * might be updates to the map between
1077          * xsk_get_map_list_entry() and xsk_map_try_sock_delete().
1078          */
1079         struct xdp_sock __rcu **map_entry = NULL;
1080         struct xsk_map *map;
1081
1082         while ((map = xsk_get_map_list_entry(xs, &map_entry))) {
1083                 xsk_map_try_sock_delete(map, xs, map_entry);
1084                 bpf_map_put(&map->map);
1085         }
1086 }
1087
1088 static int xsk_release(struct socket *sock)
1089 {
1090         struct sock *sk = sock->sk;
1091         struct xdp_sock *xs = xdp_sk(sk);
1092         struct net *net;
1093
1094         if (!sk)
1095                 return 0;
1096
1097         net = sock_net(sk);
1098
1099         if (xs->skb)
1100                 xsk_drop_skb(xs->skb);
1101
1102         mutex_lock(&net->xdp.lock);
1103         sk_del_node_init_rcu(sk);
1104         mutex_unlock(&net->xdp.lock);
1105
1106         sock_prot_inuse_add(net, sk->sk_prot, -1);
1107
1108         xsk_delete_from_maps(xs);
1109         mutex_lock(&xs->mutex);
1110         xsk_unbind_dev(xs);
1111         mutex_unlock(&xs->mutex);
1112
1113         xskq_destroy(xs->rx);
1114         xskq_destroy(xs->tx);
1115         xskq_destroy(xs->fq_tmp);
1116         xskq_destroy(xs->cq_tmp);
1117
1118         sock_orphan(sk);
1119         sock->sk = NULL;
1120
1121         sock_put(sk);
1122
1123         return 0;
1124 }
1125
1126 static struct socket *xsk_lookup_xsk_from_fd(int fd)
1127 {
1128         struct socket *sock;
1129         int err;
1130
1131         sock = sockfd_lookup(fd, &err);
1132         if (!sock)
1133                 return ERR_PTR(-ENOTSOCK);
1134
1135         if (sock->sk->sk_family != PF_XDP) {
1136                 sockfd_put(sock);
1137                 return ERR_PTR(-ENOPROTOOPT);
1138         }
1139
1140         return sock;
1141 }
1142
1143 static bool xsk_validate_queues(struct xdp_sock *xs)
1144 {
1145         return xs->fq_tmp && xs->cq_tmp;
1146 }
1147
1148 static int xsk_bind(struct socket *sock, struct sockaddr *addr, int addr_len)
1149 {
1150         struct sockaddr_xdp *sxdp = (struct sockaddr_xdp *)addr;
1151         struct sock *sk = sock->sk;
1152         struct xdp_sock *xs = xdp_sk(sk);
1153         struct net_device *dev;
1154         int bound_dev_if;
1155         u32 flags, qid;
1156         int err = 0;
1157
1158         if (addr_len < sizeof(struct sockaddr_xdp))
1159                 return -EINVAL;
1160         if (sxdp->sxdp_family != AF_XDP)
1161                 return -EINVAL;
1162
1163         flags = sxdp->sxdp_flags;
1164         if (flags & ~(XDP_SHARED_UMEM | XDP_COPY | XDP_ZEROCOPY |
1165                       XDP_USE_NEED_WAKEUP | XDP_USE_SG))
1166                 return -EINVAL;
1167
1168         bound_dev_if = READ_ONCE(sk->sk_bound_dev_if);
1169         if (bound_dev_if && bound_dev_if != sxdp->sxdp_ifindex)
1170                 return -EINVAL;
1171
1172         rtnl_lock();
1173         mutex_lock(&xs->mutex);
1174         if (xs->state != XSK_READY) {
1175                 err = -EBUSY;
1176                 goto out_release;
1177         }
1178
1179         dev = dev_get_by_index(sock_net(sk), sxdp->sxdp_ifindex);
1180         if (!dev) {
1181                 err = -ENODEV;
1182                 goto out_release;
1183         }
1184
1185         if (!xs->rx && !xs->tx) {
1186                 err = -EINVAL;
1187                 goto out_unlock;
1188         }
1189
1190         qid = sxdp->sxdp_queue_id;
1191
1192         if (flags & XDP_SHARED_UMEM) {
1193                 struct xdp_sock *umem_xs;
1194                 struct socket *sock;
1195
1196                 if ((flags & XDP_COPY) || (flags & XDP_ZEROCOPY) ||
1197                     (flags & XDP_USE_NEED_WAKEUP) || (flags & XDP_USE_SG)) {
1198                         /* Cannot specify flags for shared sockets. */
1199                         err = -EINVAL;
1200                         goto out_unlock;
1201                 }
1202
1203                 if (xs->umem) {
1204                         /* We have already our own. */
1205                         err = -EINVAL;
1206                         goto out_unlock;
1207                 }
1208
1209                 sock = xsk_lookup_xsk_from_fd(sxdp->sxdp_shared_umem_fd);
1210                 if (IS_ERR(sock)) {
1211                         err = PTR_ERR(sock);
1212                         goto out_unlock;
1213                 }
1214
1215                 umem_xs = xdp_sk(sock->sk);
1216                 if (!xsk_is_bound(umem_xs)) {
1217                         err = -EBADF;
1218                         sockfd_put(sock);
1219                         goto out_unlock;
1220                 }
1221
1222                 if (umem_xs->queue_id != qid || umem_xs->dev != dev) {
1223                         /* Share the umem with another socket on another qid
1224                          * and/or device.
1225                          */
1226                         xs->pool = xp_create_and_assign_umem(xs,
1227                                                              umem_xs->umem);
1228                         if (!xs->pool) {
1229                                 err = -ENOMEM;
1230                                 sockfd_put(sock);
1231                                 goto out_unlock;
1232                         }
1233
1234                         err = xp_assign_dev_shared(xs->pool, umem_xs, dev,
1235                                                    qid);
1236                         if (err) {
1237                                 xp_destroy(xs->pool);
1238                                 xs->pool = NULL;
1239                                 sockfd_put(sock);
1240                                 goto out_unlock;
1241                         }
1242                 } else {
1243                         /* Share the buffer pool with the other socket. */
1244                         if (xs->fq_tmp || xs->cq_tmp) {
1245                                 /* Do not allow setting your own fq or cq. */
1246                                 err = -EINVAL;
1247                                 sockfd_put(sock);
1248                                 goto out_unlock;
1249                         }
1250
1251                         xp_get_pool(umem_xs->pool);
1252                         xs->pool = umem_xs->pool;
1253
1254                         /* If underlying shared umem was created without Tx
1255                          * ring, allocate Tx descs array that Tx batching API
1256                          * utilizes
1257                          */
1258                         if (xs->tx && !xs->pool->tx_descs) {
1259                                 err = xp_alloc_tx_descs(xs->pool, xs);
1260                                 if (err) {
1261                                         xp_put_pool(xs->pool);
1262                                         xs->pool = NULL;
1263                                         sockfd_put(sock);
1264                                         goto out_unlock;
1265                                 }
1266                         }
1267                 }
1268
1269                 xdp_get_umem(umem_xs->umem);
1270                 WRITE_ONCE(xs->umem, umem_xs->umem);
1271                 sockfd_put(sock);
1272         } else if (!xs->umem || !xsk_validate_queues(xs)) {
1273                 err = -EINVAL;
1274                 goto out_unlock;
1275         } else {
1276                 /* This xsk has its own umem. */
1277                 xs->pool = xp_create_and_assign_umem(xs, xs->umem);
1278                 if (!xs->pool) {
1279                         err = -ENOMEM;
1280                         goto out_unlock;
1281                 }
1282
1283                 err = xp_assign_dev(xs->pool, dev, qid, flags);
1284                 if (err) {
1285                         xp_destroy(xs->pool);
1286                         xs->pool = NULL;
1287                         goto out_unlock;
1288                 }
1289         }
1290
1291         /* FQ and CQ are now owned by the buffer pool and cleaned up with it. */
1292         xs->fq_tmp = NULL;
1293         xs->cq_tmp = NULL;
1294
1295         xs->dev = dev;
1296         xs->zc = xs->umem->zc;
1297         xs->sg = !!(xs->umem->flags & XDP_UMEM_SG_FLAG);
1298         xs->queue_id = qid;
1299         xp_add_xsk(xs->pool, xs);
1300
1301 out_unlock:
1302         if (err) {
1303                 dev_put(dev);
1304         } else {
1305                 /* Matches smp_rmb() in bind() for shared umem
1306                  * sockets, and xsk_is_bound().
1307                  */
1308                 smp_wmb();
1309                 WRITE_ONCE(xs->state, XSK_BOUND);
1310         }
1311 out_release:
1312         mutex_unlock(&xs->mutex);
1313         rtnl_unlock();
1314         return err;
1315 }
1316
1317 struct xdp_umem_reg_v1 {
1318         __u64 addr; /* Start of packet data area */
1319         __u64 len; /* Length of packet data area */
1320         __u32 chunk_size;
1321         __u32 headroom;
1322 };
1323
1324 static int xsk_setsockopt(struct socket *sock, int level, int optname,
1325                           sockptr_t optval, unsigned int optlen)
1326 {
1327         struct sock *sk = sock->sk;
1328         struct xdp_sock *xs = xdp_sk(sk);
1329         int err;
1330
1331         if (level != SOL_XDP)
1332                 return -ENOPROTOOPT;
1333
1334         switch (optname) {
1335         case XDP_RX_RING:
1336         case XDP_TX_RING:
1337         {
1338                 struct xsk_queue **q;
1339                 int entries;
1340
1341                 if (optlen < sizeof(entries))
1342                         return -EINVAL;
1343                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1344                         return -EFAULT;
1345
1346                 mutex_lock(&xs->mutex);
1347                 if (xs->state != XSK_READY) {
1348                         mutex_unlock(&xs->mutex);
1349                         return -EBUSY;
1350                 }
1351                 q = (optname == XDP_TX_RING) ? &xs->tx : &xs->rx;
1352                 err = xsk_init_queue(entries, q, false);
1353                 if (!err && optname == XDP_TX_RING)
1354                         /* Tx needs to be explicitly woken up the first time */
1355                         xs->tx->ring->flags |= XDP_RING_NEED_WAKEUP;
1356                 mutex_unlock(&xs->mutex);
1357                 return err;
1358         }
1359         case XDP_UMEM_REG:
1360         {
1361                 size_t mr_size = sizeof(struct xdp_umem_reg);
1362                 struct xdp_umem_reg mr = {};
1363                 struct xdp_umem *umem;
1364
1365                 if (optlen < sizeof(struct xdp_umem_reg_v1))
1366                         return -EINVAL;
1367                 else if (optlen < sizeof(mr))
1368                         mr_size = sizeof(struct xdp_umem_reg_v1);
1369
1370                 BUILD_BUG_ON(sizeof(struct xdp_umem_reg_v1) >= sizeof(struct xdp_umem_reg));
1371
1372                 /* Make sure the last field of the struct doesn't have
1373                  * uninitialized padding. All padding has to be explicit
1374                  * and has to be set to zero by the userspace to make
1375                  * struct xdp_umem_reg extensible in the future.
1376                  */
1377                 BUILD_BUG_ON(offsetof(struct xdp_umem_reg, tx_metadata_len) +
1378                              sizeof_field(struct xdp_umem_reg, tx_metadata_len) !=
1379                              sizeof(struct xdp_umem_reg));
1380
1381                 if (copy_from_sockptr(&mr, optval, mr_size))
1382                         return -EFAULT;
1383
1384                 mutex_lock(&xs->mutex);
1385                 if (xs->state != XSK_READY || xs->umem) {
1386                         mutex_unlock(&xs->mutex);
1387                         return -EBUSY;
1388                 }
1389
1390                 umem = xdp_umem_create(&mr);
1391                 if (IS_ERR(umem)) {
1392                         mutex_unlock(&xs->mutex);
1393                         return PTR_ERR(umem);
1394                 }
1395
1396                 /* Make sure umem is ready before it can be seen by others */
1397                 smp_wmb();
1398                 WRITE_ONCE(xs->umem, umem);
1399                 mutex_unlock(&xs->mutex);
1400                 return 0;
1401         }
1402         case XDP_UMEM_FILL_RING:
1403         case XDP_UMEM_COMPLETION_RING:
1404         {
1405                 struct xsk_queue **q;
1406                 int entries;
1407
1408                 if (optlen < sizeof(entries))
1409                         return -EINVAL;
1410                 if (copy_from_sockptr(&entries, optval, sizeof(entries)))
1411                         return -EFAULT;
1412
1413                 mutex_lock(&xs->mutex);
1414                 if (xs->state != XSK_READY) {
1415                         mutex_unlock(&xs->mutex);
1416                         return -EBUSY;
1417                 }
1418
1419                 q = (optname == XDP_UMEM_FILL_RING) ? &xs->fq_tmp :
1420                         &xs->cq_tmp;
1421                 err = xsk_init_queue(entries, q, true);
1422                 mutex_unlock(&xs->mutex);
1423                 return err;
1424         }
1425         default:
1426                 break;
1427         }
1428
1429         return -ENOPROTOOPT;
1430 }
1431
1432 static void xsk_enter_rxtx_offsets(struct xdp_ring_offset_v1 *ring)
1433 {
1434         ring->producer = offsetof(struct xdp_rxtx_ring, ptrs.producer);
1435         ring->consumer = offsetof(struct xdp_rxtx_ring, ptrs.consumer);
1436         ring->desc = offsetof(struct xdp_rxtx_ring, desc);
1437 }
1438
1439 static void xsk_enter_umem_offsets(struct xdp_ring_offset_v1 *ring)
1440 {
1441         ring->producer = offsetof(struct xdp_umem_ring, ptrs.producer);
1442         ring->consumer = offsetof(struct xdp_umem_ring, ptrs.consumer);
1443         ring->desc = offsetof(struct xdp_umem_ring, desc);
1444 }
1445
1446 struct xdp_statistics_v1 {
1447         __u64 rx_dropped;
1448         __u64 rx_invalid_descs;
1449         __u64 tx_invalid_descs;
1450 };
1451
1452 static int xsk_getsockopt(struct socket *sock, int level, int optname,
1453                           char __user *optval, int __user *optlen)
1454 {
1455         struct sock *sk = sock->sk;
1456         struct xdp_sock *xs = xdp_sk(sk);
1457         int len;
1458
1459         if (level != SOL_XDP)
1460                 return -ENOPROTOOPT;
1461
1462         if (get_user(len, optlen))
1463                 return -EFAULT;
1464         if (len < 0)
1465                 return -EINVAL;
1466
1467         switch (optname) {
1468         case XDP_STATISTICS:
1469         {
1470                 struct xdp_statistics stats = {};
1471                 bool extra_stats = true;
1472                 size_t stats_size;
1473
1474                 if (len < sizeof(struct xdp_statistics_v1)) {
1475                         return -EINVAL;
1476                 } else if (len < sizeof(stats)) {
1477                         extra_stats = false;
1478                         stats_size = sizeof(struct xdp_statistics_v1);
1479                 } else {
1480                         stats_size = sizeof(stats);
1481                 }
1482
1483                 mutex_lock(&xs->mutex);
1484                 stats.rx_dropped = xs->rx_dropped;
1485                 if (extra_stats) {
1486                         stats.rx_ring_full = xs->rx_queue_full;
1487                         stats.rx_fill_ring_empty_descs =
1488                                 xs->pool ? xskq_nb_queue_empty_descs(xs->pool->fq) : 0;
1489                         stats.tx_ring_empty_descs = xskq_nb_queue_empty_descs(xs->tx);
1490                 } else {
1491                         stats.rx_dropped += xs->rx_queue_full;
1492                 }
1493                 stats.rx_invalid_descs = xskq_nb_invalid_descs(xs->rx);
1494                 stats.tx_invalid_descs = xskq_nb_invalid_descs(xs->tx);
1495                 mutex_unlock(&xs->mutex);
1496
1497                 if (copy_to_user(optval, &stats, stats_size))
1498                         return -EFAULT;
1499                 if (put_user(stats_size, optlen))
1500                         return -EFAULT;
1501
1502                 return 0;
1503         }
1504         case XDP_MMAP_OFFSETS:
1505         {
1506                 struct xdp_mmap_offsets off;
1507                 struct xdp_mmap_offsets_v1 off_v1;
1508                 bool flags_supported = true;
1509                 void *to_copy;
1510
1511                 if (len < sizeof(off_v1))
1512                         return -EINVAL;
1513                 else if (len < sizeof(off))
1514                         flags_supported = false;
1515
1516                 if (flags_supported) {
1517                         /* xdp_ring_offset is identical to xdp_ring_offset_v1
1518                          * except for the flags field added to the end.
1519                          */
1520                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1521                                                &off.rx);
1522                         xsk_enter_rxtx_offsets((struct xdp_ring_offset_v1 *)
1523                                                &off.tx);
1524                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1525                                                &off.fr);
1526                         xsk_enter_umem_offsets((struct xdp_ring_offset_v1 *)
1527                                                &off.cr);
1528                         off.rx.flags = offsetof(struct xdp_rxtx_ring,
1529                                                 ptrs.flags);
1530                         off.tx.flags = offsetof(struct xdp_rxtx_ring,
1531                                                 ptrs.flags);
1532                         off.fr.flags = offsetof(struct xdp_umem_ring,
1533                                                 ptrs.flags);
1534                         off.cr.flags = offsetof(struct xdp_umem_ring,
1535                                                 ptrs.flags);
1536
1537                         len = sizeof(off);
1538                         to_copy = &off;
1539                 } else {
1540                         xsk_enter_rxtx_offsets(&off_v1.rx);
1541                         xsk_enter_rxtx_offsets(&off_v1.tx);
1542                         xsk_enter_umem_offsets(&off_v1.fr);
1543                         xsk_enter_umem_offsets(&off_v1.cr);
1544
1545                         len = sizeof(off_v1);
1546                         to_copy = &off_v1;
1547                 }
1548
1549                 if (copy_to_user(optval, to_copy, len))
1550                         return -EFAULT;
1551                 if (put_user(len, optlen))
1552                         return -EFAULT;
1553
1554                 return 0;
1555         }
1556         case XDP_OPTIONS:
1557         {
1558                 struct xdp_options opts = {};
1559
1560                 if (len < sizeof(opts))
1561                         return -EINVAL;
1562
1563                 mutex_lock(&xs->mutex);
1564                 if (xs->zc)
1565                         opts.flags |= XDP_OPTIONS_ZEROCOPY;
1566                 mutex_unlock(&xs->mutex);
1567
1568                 len = sizeof(opts);
1569                 if (copy_to_user(optval, &opts, len))
1570                         return -EFAULT;
1571                 if (put_user(len, optlen))
1572                         return -EFAULT;
1573
1574                 return 0;
1575         }
1576         default:
1577                 break;
1578         }
1579
1580         return -EOPNOTSUPP;
1581 }
1582
1583 static int xsk_mmap(struct file *file, struct socket *sock,
1584                     struct vm_area_struct *vma)
1585 {
1586         loff_t offset = (loff_t)vma->vm_pgoff << PAGE_SHIFT;
1587         unsigned long size = vma->vm_end - vma->vm_start;
1588         struct xdp_sock *xs = xdp_sk(sock->sk);
1589         int state = READ_ONCE(xs->state);
1590         struct xsk_queue *q = NULL;
1591
1592         if (state != XSK_READY && state != XSK_BOUND)
1593                 return -EBUSY;
1594
1595         if (offset == XDP_PGOFF_RX_RING) {
1596                 q = READ_ONCE(xs->rx);
1597         } else if (offset == XDP_PGOFF_TX_RING) {
1598                 q = READ_ONCE(xs->tx);
1599         } else {
1600                 /* Matches the smp_wmb() in XDP_UMEM_REG */
1601                 smp_rmb();
1602                 if (offset == XDP_UMEM_PGOFF_FILL_RING)
1603                         q = state == XSK_READY ? READ_ONCE(xs->fq_tmp) :
1604                                                  READ_ONCE(xs->pool->fq);
1605                 else if (offset == XDP_UMEM_PGOFF_COMPLETION_RING)
1606                         q = state == XSK_READY ? READ_ONCE(xs->cq_tmp) :
1607                                                  READ_ONCE(xs->pool->cq);
1608         }
1609
1610         if (!q)
1611                 return -EINVAL;
1612
1613         /* Matches the smp_wmb() in xsk_init_queue */
1614         smp_rmb();
1615         if (size > q->ring_vmalloc_size)
1616                 return -EINVAL;
1617
1618         return remap_vmalloc_range(vma, q->ring, 0);
1619 }
1620
1621 static int xsk_notifier(struct notifier_block *this,
1622                         unsigned long msg, void *ptr)
1623 {
1624         struct net_device *dev = netdev_notifier_info_to_dev(ptr);
1625         struct net *net = dev_net(dev);
1626         struct sock *sk;
1627
1628         switch (msg) {
1629         case NETDEV_UNREGISTER:
1630                 mutex_lock(&net->xdp.lock);
1631                 sk_for_each(sk, &net->xdp.list) {
1632                         struct xdp_sock *xs = xdp_sk(sk);
1633
1634                         mutex_lock(&xs->mutex);
1635                         if (xs->dev == dev) {
1636                                 sk->sk_err = ENETDOWN;
1637                                 if (!sock_flag(sk, SOCK_DEAD))
1638                                         sk_error_report(sk);
1639
1640                                 xsk_unbind_dev(xs);
1641
1642                                 /* Clear device references. */
1643                                 xp_clear_dev(xs->pool);
1644                         }
1645                         mutex_unlock(&xs->mutex);
1646                 }
1647                 mutex_unlock(&net->xdp.lock);
1648                 break;
1649         }
1650         return NOTIFY_DONE;
1651 }
1652
1653 static struct proto xsk_proto = {
1654         .name =         "XDP",
1655         .owner =        THIS_MODULE,
1656         .obj_size =     sizeof(struct xdp_sock),
1657 };
1658
1659 static const struct proto_ops xsk_proto_ops = {
1660         .family         = PF_XDP,
1661         .owner          = THIS_MODULE,
1662         .release        = xsk_release,
1663         .bind           = xsk_bind,
1664         .connect        = sock_no_connect,
1665         .socketpair     = sock_no_socketpair,
1666         .accept         = sock_no_accept,
1667         .getname        = sock_no_getname,
1668         .poll           = xsk_poll,
1669         .ioctl          = sock_no_ioctl,
1670         .listen         = sock_no_listen,
1671         .shutdown       = sock_no_shutdown,
1672         .setsockopt     = xsk_setsockopt,
1673         .getsockopt     = xsk_getsockopt,
1674         .sendmsg        = xsk_sendmsg,
1675         .recvmsg        = xsk_recvmsg,
1676         .mmap           = xsk_mmap,
1677 };
1678
1679 static void xsk_destruct(struct sock *sk)
1680 {
1681         struct xdp_sock *xs = xdp_sk(sk);
1682
1683         if (!sock_flag(sk, SOCK_DEAD))
1684                 return;
1685
1686         if (!xp_put_pool(xs->pool))
1687                 xdp_put_umem(xs->umem, !xs->pool);
1688 }
1689
1690 static int xsk_create(struct net *net, struct socket *sock, int protocol,
1691                       int kern)
1692 {
1693         struct xdp_sock *xs;
1694         struct sock *sk;
1695
1696         if (!ns_capable(net->user_ns, CAP_NET_RAW))
1697                 return -EPERM;
1698         if (sock->type != SOCK_RAW)
1699                 return -ESOCKTNOSUPPORT;
1700
1701         if (protocol)
1702                 return -EPROTONOSUPPORT;
1703
1704         sock->state = SS_UNCONNECTED;
1705
1706         sk = sk_alloc(net, PF_XDP, GFP_KERNEL, &xsk_proto, kern);
1707         if (!sk)
1708                 return -ENOBUFS;
1709
1710         sock->ops = &xsk_proto_ops;
1711
1712         sock_init_data(sock, sk);
1713
1714         sk->sk_family = PF_XDP;
1715
1716         sk->sk_destruct = xsk_destruct;
1717
1718         sock_set_flag(sk, SOCK_RCU_FREE);
1719
1720         xs = xdp_sk(sk);
1721         xs->state = XSK_READY;
1722         mutex_init(&xs->mutex);
1723         spin_lock_init(&xs->rx_lock);
1724
1725         INIT_LIST_HEAD(&xs->map_list);
1726         spin_lock_init(&xs->map_list_lock);
1727
1728         mutex_lock(&net->xdp.lock);
1729         sk_add_node_rcu(sk, &net->xdp.list);
1730         mutex_unlock(&net->xdp.lock);
1731
1732         sock_prot_inuse_add(net, &xsk_proto, 1);
1733
1734         return 0;
1735 }
1736
1737 static const struct net_proto_family xsk_family_ops = {
1738         .family = PF_XDP,
1739         .create = xsk_create,
1740         .owner  = THIS_MODULE,
1741 };
1742
1743 static struct notifier_block xsk_netdev_notifier = {
1744         .notifier_call  = xsk_notifier,
1745 };
1746
1747 static int __net_init xsk_net_init(struct net *net)
1748 {
1749         mutex_init(&net->xdp.lock);
1750         INIT_HLIST_HEAD(&net->xdp.list);
1751         return 0;
1752 }
1753
1754 static void __net_exit xsk_net_exit(struct net *net)
1755 {
1756         WARN_ON_ONCE(!hlist_empty(&net->xdp.list));
1757 }
1758
1759 static struct pernet_operations xsk_net_ops = {
1760         .init = xsk_net_init,
1761         .exit = xsk_net_exit,
1762 };
1763
1764 static int __init xsk_init(void)
1765 {
1766         int err;
1767
1768         err = proto_register(&xsk_proto, 0 /* no slab */);
1769         if (err)
1770                 goto out;
1771
1772         err = sock_register(&xsk_family_ops);
1773         if (err)
1774                 goto out_proto;
1775
1776         err = register_pernet_subsys(&xsk_net_ops);
1777         if (err)
1778                 goto out_sk;
1779
1780         err = register_netdevice_notifier(&xsk_netdev_notifier);
1781         if (err)
1782                 goto out_pernet;
1783
1784         return 0;
1785
1786 out_pernet:
1787         unregister_pernet_subsys(&xsk_net_ops);
1788 out_sk:
1789         sock_unregister(PF_XDP);
1790 out_proto:
1791         proto_unregister(&xsk_proto);
1792 out:
1793         return err;
1794 }
1795
1796 fs_initcall(xsk_init);