net/ipv4/inet_fragment.c

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