net/ipv4/inet_fragment.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * inet fragments management
   4  *
   5  *              Authors:        Pavel Emelyanov <xemul@openvz.org>
   6  *                              Started as consolidation of ipv4/ip_fragment.c,
   7  *                              ipv6/reassembly. and ipv6 nf conntrack reassembly
   8  */
   9
  10 #include <linux/list.h>
  11 #include <linux/spinlock.h>
  12 #include <linux/module.h>
  13 #include <linux/timer.h>
  14 #include <linux/mm.h>
  15 #include <linux/random.h>
  16 #include <linux/skbuff.h>
  17 #include <linux/rtnetlink.h>
  18 #include <linux/slab.h>
  19 #include <linux/rhashtable.h>
  20
  21 #include <net/sock.h>
  22 #include <net/inet_frag.h>
  23 #include <net/inet_ecn.h>
  24 #include <net/ip.h>
  25 #include <net/ipv6.h>
  26
  27 #include "../core/sock_destructor.h"
  28
  29 /* Use skb->cb to track consecutive/adjacent fragments coming at
  30  * the end of the queue. Nodes in the rb-tree queue will
  31  * contain "runs" of one or more adjacent fragments.
  32  *
  33  * Invariants:
  34  * - next_frag is NULL at the tail of a "run";
  35  * - the head of a "run" has the sum of all fragment lengths in frag_run_len.
  36  */
  37 struct ipfrag_skb_cb {
  38         union {
  39                 struct inet_skb_parm    h4;
  40                 struct inet6_skb_parm   h6;
  41         };
  42         struct sk_buff          *next_frag;
  43         int                     frag_run_len;
  44         int                     ip_defrag_offset;
  45 };
  46
  47 #define FRAG_CB(skb)            ((struct ipfrag_skb_cb *)((skb)->cb))
  48
  49 static void fragcb_clear(struct sk_buff *skb)
  50 {
  51         RB_CLEAR_NODE(&skb->rbnode);
  52         FRAG_CB(skb)->next_frag = NULL;
  53         FRAG_CB(skb)->frag_run_len = skb->len;
  54 }
  55
  56 /* Append skb to the last "run". */
  57 static void fragrun_append_to_last(struct inet_frag_queue *q,
  58                                    struct sk_buff *skb)
  59 {
  60         fragcb_clear(skb);
  61
  62         FRAG_CB(q->last_run_head)->frag_run_len += skb->len;
  63         FRAG_CB(q->fragments_tail)->next_frag = skb;
  64         q->fragments_tail = skb;
  65 }
  66
  67 /* Create a new "run" with the skb. */
  68 static void fragrun_create(struct inet_frag_queue *q, struct sk_buff *skb)
  69 {
  70         BUILD_BUG_ON(sizeof(struct ipfrag_skb_cb) > sizeof(skb->cb));
  71         fragcb_clear(skb);
  72
  73         if (q->last_run_head)
  74                 rb_link_node(&skb->rbnode, &q->last_run_head->rbnode,
  75                              &q->last_run_head->rbnode.rb_right);
  76         else
  77                 rb_link_node(&skb->rbnode, NULL, &q->rb_fragments.rb_node);
  78         rb_insert_color(&skb->rbnode, &q->rb_fragments);
  79
  80         q->fragments_tail = skb;
  81         q->last_run_head = skb;
  82 }
  83
  84 /* Given the OR values of all fragments, apply RFC 3168 5.3 requirements
  85  * Value : 0xff if frame should be dropped.
  86  *         0 or INET_ECN_CE value, to be ORed in to final iph->tos field
  87  */
  88 const u8 ip_frag_ecn_table[16] = {
  89         /* at least one fragment had CE, and others ECT_0 or ECT_1 */
  90         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0]                      = INET_ECN_CE,
  91         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1]                      = INET_ECN_CE,
  92         [IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1]   = INET_ECN_CE,
  93
  94         /* invalid combinations : drop frame */
  95         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE] = 0xff,
  96         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0] = 0xff,
  97         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_1] = 0xff,
  98         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
  99         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0] = 0xff,
 100         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_1] = 0xff,
 101         [IPFRAG_ECN_NOT_ECT | IPFRAG_ECN_CE | IPFRAG_ECN_ECT_0 | IPFRAG_ECN_ECT_1] = 0xff,
 102 };
 103 EXPORT_SYMBOL(ip_frag_ecn_table);
 104
 105 int inet_frags_init(struct inet_frags *f)
 106 {
 107         f->frags_cachep = kmem_cache_create(f->frags_cache_name, f->qsize, 0, 0,
 108                                             NULL);
 109         if (!f->frags_cachep)
 110                 return -ENOMEM;
 111
 112         refcount_set(&f->refcnt, 1);
 113         init_completion(&f->completion);
 114         return 0;
 115 }
 116 EXPORT_SYMBOL(inet_frags_init);
 117
 118 void inet_frags_fini(struct inet_frags *f)
 119 {
 120         if (refcount_dec_and_test(&f->refcnt))
 121                 complete(&f->completion);
 122
 123         wait_for_completion(&f->completion);
 124
 125         kmem_cache_destroy(f->frags_cachep);
 126         f->frags_cachep = NULL;
 127 }
 128 EXPORT_SYMBOL(inet_frags_fini);
 129
 130 /* called from rhashtable_free_and_destroy() at netns_frags dismantle */
 131 static void inet_frags_free_cb(void *ptr, void *arg)
 132 {
 133         struct inet_frag_queue *fq = ptr;
 134         int count;
 135
 136         count = del_timer_sync(&fq->timer) ? 1 : 0;
 137
 138         spin_lock_bh(&fq->lock);
 139         if (!(fq->flags & INET_FRAG_COMPLETE)) {
 140                 fq->flags |= INET_FRAG_COMPLETE;
 141                 count++;
 142         } else if (fq->flags & INET_FRAG_HASH_DEAD) {
 143                 count++;
 144         }
 145         spin_unlock_bh(&fq->lock);
 146
 147         if (refcount_sub_and_test(count, &fq->refcnt))
 148                 inet_frag_destroy(fq);
 149 }
 150
 151 static LLIST_HEAD(fqdir_free_list);
 152
 153 static void fqdir_free_fn(struct work_struct *work)
 154 {
 155         struct llist_node *kill_list;
 156         struct fqdir *fqdir, *tmp;
 157         struct inet_frags *f;
 158
 159         /* Atomically snapshot the list of fqdirs to free */
 160         kill_list = llist_del_all(&fqdir_free_list);
 161
 162         /* We need to make sure all ongoing call_rcu(..., inet_frag_destroy_rcu)
 163          * have completed, since they need to dereference fqdir.
 164          * Would it not be nice to have kfree_rcu_barrier() ? :)
 165          */
 166         rcu_barrier();
 167
 168         llist_for_each_entry_safe(fqdir, tmp, kill_list, free_list) {
 169                 f = fqdir->f;
 170                 if (refcount_dec_and_test(&f->refcnt))
 171                         complete(&f->completion);
 172
 173                 kfree(fqdir);
 174         }
 175 }
 176
 177 static DECLARE_WORK(fqdir_free_work, fqdir_free_fn);
 178
 179 static void fqdir_work_fn(struct work_struct *work)
 180 {
 181         struct fqdir *fqdir = container_of(work, struct fqdir, destroy_work);
 182
 183         rhashtable_free_and_destroy(&fqdir->rhashtable, inet_frags_free_cb, NULL);
 184
 185         if (llist_add(&fqdir->free_list, &fqdir_free_list))
 186                 queue_work(system_wq, &fqdir_free_work);
 187 }
 188
 189 int fqdir_init(struct fqdir **fqdirp, struct inet_frags *f, struct net *net)
 190 {
 191         struct fqdir *fqdir = kzalloc(sizeof(*fqdir), GFP_KERNEL);
 192         int res;
 193
 194         if (!fqdir)
 195                 return -ENOMEM;
 196         fqdir->f = f;
 197         fqdir->net = net;
 198         res = rhashtable_init(&fqdir->rhashtable, &fqdir->f->rhash_params);
 199         if (res < 0) {
 200                 kfree(fqdir);
 201                 return res;
 202         }
 203         refcount_inc(&f->refcnt);
 204         *fqdirp = fqdir;
 205         return 0;
 206 }
 207 EXPORT_SYMBOL(fqdir_init);
 208
 209 static struct workqueue_struct *inet_frag_wq;
 210
 211 static int __init inet_frag_wq_init(void)
 212 {
 213         inet_frag_wq = create_workqueue("inet_frag_wq");
 214         if (!inet_frag_wq)
 215                 panic("Could not create inet frag workq");
 216         return 0;
 217 }
 218
 219 pure_initcall(inet_frag_wq_init);
 220
 221 void fqdir_exit(struct fqdir *fqdir)
 222 {
 223         INIT_WORK(&fqdir->destroy_work, fqdir_work_fn);
 224         queue_work(inet_frag_wq, &fqdir->destroy_work);
 225 }
 226 EXPORT_SYMBOL(fqdir_exit);
 227
 228 void inet_frag_kill(struct inet_frag_queue *fq)
 229 {
 230         if (del_timer(&fq->timer))
 231                 refcount_dec(&fq->refcnt);
 232
 233         if (!(fq->flags & INET_FRAG_COMPLETE)) {
 234                 struct fqdir *fqdir = fq->fqdir;
 235
 236                 fq->flags |= INET_FRAG_COMPLETE;
 237                 rcu_read_lock();
 238                 /* The RCU read lock provides a memory barrier
 239                  * guaranteeing that if fqdir->dead is false then
 240                  * the hash table destruction will not start until
 241                  * after we unlock.  Paired with fqdir_pre_exit().
 242                  */
 243                 if (!READ_ONCE(fqdir->dead)) {
 244                         rhashtable_remove_fast(&fqdir->rhashtable, &fq->node,
 245                                                fqdir->f->rhash_params);
 246                         refcount_dec(&fq->refcnt);
 247                 } else {
 248                         fq->flags |= INET_FRAG_HASH_DEAD;
 249                 }
 250                 rcu_read_unlock();
 251         }
 252 }
 253 EXPORT_SYMBOL(inet_frag_kill);
 254
 255 static void inet_frag_destroy_rcu(struct rcu_head *head)
 256 {
 257         struct inet_frag_queue *q = container_of(head, struct inet_frag_queue,
 258                                                  rcu);
 259         struct inet_frags *f = q->fqdir->f;
 260
 261         if (f->destructor)
 262                 f->destructor(q);
 263         kmem_cache_free(f->frags_cachep, q);
 264 }
 265
 266 unsigned int inet_frag_rbtree_purge(struct rb_root *root)
 267 {
 268         struct rb_node *p = rb_first(root);
 269         unsigned int sum = 0;
 270
 271         while (p) {
 272                 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
 273
 274                 p = rb_next(p);
 275                 rb_erase(&skb->rbnode, root);
 276                 while (skb) {
 277                         struct sk_buff *next = FRAG_CB(skb)->next_frag;
 278
 279                         sum += skb->truesize;
 280                         kfree_skb(skb);
 281                         skb = next;
 282                 }
 283         }
 284         return sum;
 285 }
 286 EXPORT_SYMBOL(inet_frag_rbtree_purge);
 287
 288 void inet_frag_destroy(struct inet_frag_queue *q)
 289 {
 290         struct fqdir *fqdir;
 291         unsigned int sum, sum_truesize = 0;
 292         struct inet_frags *f;
 293
 294         WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
 295         WARN_ON(del_timer(&q->timer) != 0);
 296
 297         /* Release all fragment data. */
 298         fqdir = q->fqdir;
 299         f = fqdir->f;
 300         sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments);
 301         sum = sum_truesize + f->qsize;
 302
 303         call_rcu(&q->rcu, inet_frag_destroy_rcu);
 304
 305         sub_frag_mem_limit(fqdir, sum);
 306 }
 307 EXPORT_SYMBOL(inet_frag_destroy);
 308
 309 static struct inet_frag_queue *inet_frag_alloc(struct fqdir *fqdir,
 310                                                struct inet_frags *f,
 311                                                void *arg)
 312 {
 313         struct inet_frag_queue *q;
 314
 315         q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
 316         if (!q)
 317                 return NULL;
 318
 319         q->fqdir = fqdir;
 320         f->constructor(q, arg);
 321         add_frag_mem_limit(fqdir, f->qsize);
 322
 323         timer_setup(&q->timer, f->frag_expire, 0);
 324         spin_lock_init(&q->lock);
 325         refcount_set(&q->refcnt, 3);
 326
 327         return q;
 328 }
 329
 330 static struct inet_frag_queue *inet_frag_create(struct fqdir *fqdir,
 331                                                 void *arg,
 332                                                 struct inet_frag_queue **prev)
 333 {
 334         struct inet_frags *f = fqdir->f;
 335         struct inet_frag_queue *q;
 336
 337         q = inet_frag_alloc(fqdir, f, arg);
 338         if (!q) {
 339                 *prev = ERR_PTR(-ENOMEM);
 340                 return NULL;
 341         }
 342         mod_timer(&q->timer, jiffies + fqdir->timeout);
 343
 344         *prev = rhashtable_lookup_get_insert_key(&fqdir->rhashtable, &q->key,
 345                                                  &q->node, f->rhash_params);
 346         if (*prev) {
 347                 q->flags |= INET_FRAG_COMPLETE;
 348                 inet_frag_kill(q);
 349                 inet_frag_destroy(q);
 350                 return NULL;
 351         }
 352         return q;
 353 }
 354
 355 /* TODO : call from rcu_read_lock() and no longer use refcount_inc_not_zero() */
 356 struct inet_frag_queue *inet_frag_find(struct fqdir *fqdir, void *key)
 357 {
 358         /* This pairs with WRITE_ONCE() in fqdir_pre_exit(). */
 359         long high_thresh = READ_ONCE(fqdir->high_thresh);
 360         struct inet_frag_queue *fq = NULL, *prev;
 361
 362         if (!high_thresh || frag_mem_limit(fqdir) > high_thresh)
 363                 return NULL;
 364
 365         rcu_read_lock();
 366
 367         prev = rhashtable_lookup(&fqdir->rhashtable, key, fqdir->f->rhash_params);
 368         if (!prev)
 369                 fq = inet_frag_create(fqdir, key, &prev);
 370         if (!IS_ERR_OR_NULL(prev)) {
 371                 fq = prev;
 372                 if (!refcount_inc_not_zero(&fq->refcnt))
 373                         fq = NULL;
 374         }
 375         rcu_read_unlock();
 376         return fq;
 377 }
 378 EXPORT_SYMBOL(inet_frag_find);
 379
 380 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb,
 381                            int offset, int end)
 382 {
 383         struct sk_buff *last = q->fragments_tail;
 384
 385         /* RFC5722, Section 4, amended by Errata ID : 3089
 386          *                          When reassembling an IPv6 datagram, if
 387          *   one or more its constituent fragments is determined to be an
 388          *   overlapping fragment, the entire datagram (and any constituent
 389          *   fragments) MUST be silently discarded.
 390          *
 391          * Duplicates, however, should be ignored (i.e. skb dropped, but the
 392          * queue/fragments kept for later reassembly).
 393          */
 394         if (!last)
 395                 fragrun_create(q, skb);  /* First fragment. */
 396         else if (FRAG_CB(last)->ip_defrag_offset + last->len < end) {
 397                 /* This is the common case: skb goes to the end. */
 398                 /* Detect and discard overlaps. */
 399                 if (offset < FRAG_CB(last)->ip_defrag_offset + last->len)
 400                         return IPFRAG_OVERLAP;
 401                 if (offset == FRAG_CB(last)->ip_defrag_offset + last->len)
 402                         fragrun_append_to_last(q, skb);
 403                 else
 404                         fragrun_create(q, skb);
 405         } else {
 406                 /* Binary search. Note that skb can become the first fragment,
 407                  * but not the last (covered above).
 408                  */
 409                 struct rb_node **rbn, *parent;
 410
 411                 rbn = &q->rb_fragments.rb_node;
 412                 do {
 413                         struct sk_buff *curr;
 414                         int curr_run_end;
 415
 416                         parent = *rbn;
 417                         curr = rb_to_skb(parent);
 418                         curr_run_end = FRAG_CB(curr)->ip_defrag_offset +
 419                                         FRAG_CB(curr)->frag_run_len;
 420                         if (end <= FRAG_CB(curr)->ip_defrag_offset)
 421                                 rbn = &parent->rb_left;
 422                         else if (offset >= curr_run_end)
 423                                 rbn = &parent->rb_right;
 424                         else if (offset >= FRAG_CB(curr)->ip_defrag_offset &&
 425                                  end <= curr_run_end)
 426                                 return IPFRAG_DUP;
 427                         else
 428                                 return IPFRAG_OVERLAP;
 429                 } while (*rbn);
 430                 /* Here we have parent properly set, and rbn pointing to
 431                  * one of its NULL left/right children. Insert skb.
 432                  */
 433                 fragcb_clear(skb);
 434                 rb_link_node(&skb->rbnode, parent, rbn);
 435                 rb_insert_color(&skb->rbnode, &q->rb_fragments);
 436         }
 437
 438         FRAG_CB(skb)->ip_defrag_offset = offset;
 439
 440         return IPFRAG_OK;
 441 }
 442 EXPORT_SYMBOL(inet_frag_queue_insert);
 443
 444 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb,
 445                               struct sk_buff *parent)
 446 {
 447         struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments);
 448         void (*destructor)(struct sk_buff *);
 449         unsigned int orig_truesize = 0;
 450         struct sk_buff **nextp = NULL;
 451         struct sock *sk = skb->sk;
 452         int delta;
 453
 454         if (sk && is_skb_wmem(skb)) {
 455                 /* TX: skb->sk might have been passed as argument to
 456                  * dst->output and must remain valid until tx completes.
 457                  *
 458                  * Move sk to reassembled skb and fix up wmem accounting.
 459                  */
 460                 orig_truesize = skb->truesize;
 461                 destructor = skb->destructor;
 462         }
 463
 464         if (head != skb) {
 465                 fp = skb_clone(skb, GFP_ATOMIC);
 466                 if (!fp) {
 467                         head = skb;
 468                         goto out_restore_sk;
 469                 }
 470                 FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag;
 471                 if (RB_EMPTY_NODE(&skb->rbnode))
 472                         FRAG_CB(parent)->next_frag = fp;
 473                 else
 474                         rb_replace_node(&skb->rbnode, &fp->rbnode,
 475                                         &q->rb_fragments);
 476                 if (q->fragments_tail == skb)
 477                         q->fragments_tail = fp;
 478
 479                 if (orig_truesize) {
 480                         /* prevent skb_morph from releasing sk */
 481                         skb->sk = NULL;
 482                         skb->destructor = NULL;
 483                 }
 484                 skb_morph(skb, head);
 485                 FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag;
 486                 rb_replace_node(&head->rbnode, &skb->rbnode,
 487                                 &q->rb_fragments);
 488                 consume_skb(head);
 489                 head = skb;
 490         }
 491         WARN_ON(FRAG_CB(head)->ip_defrag_offset != 0);
 492
 493         delta = -head->truesize;
 494
 495         /* Head of list must not be cloned. */
 496         if (skb_unclone(head, GFP_ATOMIC))
 497                 goto out_restore_sk;
 498
 499         delta += head->truesize;
 500         if (delta)
 501                 add_frag_mem_limit(q->fqdir, delta);
 502
 503         /* If the first fragment is fragmented itself, we split
 504          * it to two chunks: the first with data and paged part
 505          * and the second, holding only fragments.
 506          */
 507         if (skb_has_frag_list(head)) {
 508                 struct sk_buff *clone;
 509                 int i, plen = 0;
 510
 511                 clone = alloc_skb(0, GFP_ATOMIC);
 512                 if (!clone)
 513                         goto out_restore_sk;
 514                 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
 515                 skb_frag_list_init(head);
 516                 for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
 517                         plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
 518                 clone->data_len = head->data_len - plen;
 519                 clone->len = clone->data_len;
 520                 head->truesize += clone->truesize;
 521                 clone->csum = 0;
 522                 clone->ip_summed = head->ip_summed;
 523                 add_frag_mem_limit(q->fqdir, clone->truesize);
 524                 skb_shinfo(head)->frag_list = clone;
 525                 nextp = &clone->next;
 526         } else {
 527                 nextp = &skb_shinfo(head)->frag_list;
 528         }
 529
 530 out_restore_sk:
 531         if (orig_truesize) {
 532                 int ts_delta = head->truesize - orig_truesize;
 533
 534                 /* if this reassembled skb is fragmented later,
 535                  * fraglist skbs will get skb->sk assigned from head->sk,
 536                  * and each frag skb will be released via sock_wfree.
 537                  *
 538                  * Update sk_wmem_alloc.
 539                  */
 540                 head->sk = sk;
 541                 head->destructor = destructor;
 542                 refcount_add(ts_delta, &sk->sk_wmem_alloc);
 543         }
 544
 545         return nextp;
 546 }
 547 EXPORT_SYMBOL(inet_frag_reasm_prepare);
 548
 549 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head,
 550                             void *reasm_data, bool try_coalesce)
 551 {
 552         struct sock *sk = is_skb_wmem(head) ? head->sk : NULL;
 553         const unsigned int head_truesize = head->truesize;
 554         struct sk_buff **nextp = reasm_data;
 555         struct rb_node *rbn;
 556         struct sk_buff *fp;
 557         int sum_truesize;
 558
 559         skb_push(head, head->data - skb_network_header(head));
 560
 561         /* Traverse the tree in order, to build frag_list. */
 562         fp = FRAG_CB(head)->next_frag;
 563         rbn = rb_next(&head->rbnode);
 564         rb_erase(&head->rbnode, &q->rb_fragments);
 565
 566         sum_truesize = head->truesize;
 567         while (rbn || fp) {
 568                 /* fp points to the next sk_buff in the current run;
 569                  * rbn points to the next run.
 570                  */
 571                 /* Go through the current run. */
 572                 while (fp) {
 573                         struct sk_buff *next_frag = FRAG_CB(fp)->next_frag;
 574                         bool stolen;
 575                         int delta;
 576
 577                         sum_truesize += fp->truesize;
 578                         if (head->ip_summed != fp->ip_summed)
 579                                 head->ip_summed = CHECKSUM_NONE;
 580                         else if (head->ip_summed == CHECKSUM_COMPLETE)
 581                                 head->csum = csum_add(head->csum, fp->csum);
 582
 583                         if (try_coalesce && skb_try_coalesce(head, fp, &stolen,
 584                                                              &delta)) {
 585                                 kfree_skb_partial(fp, stolen);
 586                         } else {
 587                                 fp->prev = NULL;
 588                                 memset(&fp->rbnode, 0, sizeof(fp->rbnode));
 589                                 fp->sk = NULL;
 590
 591                                 head->data_len += fp->len;
 592                                 head->len += fp->len;
 593                                 head->truesize += fp->truesize;
 594
 595                                 *nextp = fp;
 596                                 nextp = &fp->next;
 597                         }
 598
 599                         fp = next_frag;
 600                 }
 601                 /* Move to the next run. */
 602                 if (rbn) {
 603                         struct rb_node *rbnext = rb_next(rbn);
 604
 605                         fp = rb_to_skb(rbn);
 606                         rb_erase(rbn, &q->rb_fragments);
 607                         rbn = rbnext;
 608                 }
 609         }
 610         sub_frag_mem_limit(q->fqdir, sum_truesize);
 611
 612         *nextp = NULL;
 613         skb_mark_not_on_list(head);
 614         head->prev = NULL;
 615         head->tstamp = q->stamp;
 616         head->mono_delivery_time = q->mono_delivery_time;
 617
 618         if (sk)
 619                 refcount_add(sum_truesize - head_truesize, &sk->sk_wmem_alloc);
 620 }
 621 EXPORT_SYMBOL(inet_frag_reasm_finish);
 622
 623 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q)
 624 {
 625         struct sk_buff *head, *skb;
 626
 627         head = skb_rb_first(&q->rb_fragments);
 628         if (!head)
 629                 return NULL;
 630         skb = FRAG_CB(head)->next_frag;
 631         if (skb)
 632                 rb_replace_node(&head->rbnode, &skb->rbnode,
 633                                 &q->rb_fragments);
 634         else
 635                 rb_erase(&head->rbnode, &q->rb_fragments);
 636         memset(&head->rbnode, 0, sizeof(head->rbnode));
 637         barrier();
 638
 639         if (head == q->fragments_tail)
 640                 q->fragments_tail = NULL;
 641
 642         sub_frag_mem_limit(q->fqdir, head->truesize);
 643
 644         return head;
 645 }
 646 EXPORT_SYMBOL(inet_frag_pull_head);