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         return 0;
 114 }
 115 EXPORT_SYMBOL(inet_frags_init);
 116
 117 void inet_frags_fini(struct inet_frags *f)
 118 {
 119         /* We must wait that all inet_frag_destroy_rcu() have completed. */
 120         rcu_barrier();
 121
 122         kmem_cache_destroy(f->frags_cachep);
 123         f->frags_cachep = NULL;
 124 }
 125 EXPORT_SYMBOL(inet_frags_fini);
 126
 127 static void inet_frags_free_cb(void *ptr, void *arg)
 128 {
 129         struct inet_frag_queue *fq = ptr;
 130
 131         /* If we can not cancel the timer, it means this frag_queue
 132          * is already disappearing, we have nothing to do.
 133          * Otherwise, we own a refcount until the end of this function.
 134          */
 135         if (!del_timer(&fq->timer))
 136                 return;
 137
 138         spin_lock_bh(&fq->lock);
 139         if (!(fq->flags & INET_FRAG_COMPLETE)) {
 140                 fq->flags |= INET_FRAG_COMPLETE;
 141                 refcount_dec(&fq->refcnt);
 142         }
 143         spin_unlock_bh(&fq->lock);
 144
 145         inet_frag_put(fq);
 146 }
 147
 148 void inet_frags_exit_net(struct netns_frags *nf)
 149 {
 150         nf->high_thresh = 0; /* prevent creation of new frags */
 151
 152         rhashtable_free_and_destroy(&nf->rhashtable, inet_frags_free_cb, NULL);
 153 }
 154 EXPORT_SYMBOL(inet_frags_exit_net);
 155
 156 void inet_frag_kill(struct inet_frag_queue *fq)
 157 {
 158         if (del_timer(&fq->timer))
 159                 refcount_dec(&fq->refcnt);
 160
 161         if (!(fq->flags & INET_FRAG_COMPLETE)) {
 162                 struct netns_frags *nf = fq->net;
 163
 164                 fq->flags |= INET_FRAG_COMPLETE;
 165                 rhashtable_remove_fast(&nf->rhashtable, &fq->node, nf->f->rhash_params);
 166                 refcount_dec(&fq->refcnt);
 167         }
 168 }
 169 EXPORT_SYMBOL(inet_frag_kill);
 170
 171 static void inet_frag_destroy_rcu(struct rcu_head *head)
 172 {
 173         struct inet_frag_queue *q = container_of(head, struct inet_frag_queue,
 174                                                  rcu);
 175         struct inet_frags *f = q->net->f;
 176
 177         if (f->destructor)
 178                 f->destructor(q);
 179         kmem_cache_free(f->frags_cachep, q);
 180 }
 181
 182 unsigned int inet_frag_rbtree_purge(struct rb_root *root)
 183 {
 184         struct rb_node *p = rb_first(root);
 185         unsigned int sum = 0;
 186
 187         while (p) {
 188                 struct sk_buff *skb = rb_entry(p, struct sk_buff, rbnode);
 189
 190                 p = rb_next(p);
 191                 rb_erase(&skb->rbnode, root);
 192                 while (skb) {
 193                         struct sk_buff *next = FRAG_CB(skb)->next_frag;
 194
 195                         sum += skb->truesize;
 196                         kfree_skb(skb);
 197                         skb = next;
 198                 }
 199         }
 200         return sum;
 201 }
 202 EXPORT_SYMBOL(inet_frag_rbtree_purge);
 203
 204 void inet_frag_destroy(struct inet_frag_queue *q)
 205 {
 206         struct sk_buff *fp;
 207         struct netns_frags *nf;
 208         unsigned int sum, sum_truesize = 0;
 209         struct inet_frags *f;
 210
 211         WARN_ON(!(q->flags & INET_FRAG_COMPLETE));
 212         WARN_ON(del_timer(&q->timer) != 0);
 213
 214         /* Release all fragment data. */
 215         fp = q->fragments;
 216         nf = q->net;
 217         f = nf->f;
 218         if (fp) {
 219                 do {
 220                         struct sk_buff *xp = fp->next;
 221
 222                         sum_truesize += fp->truesize;
 223                         kfree_skb(fp);
 224                         fp = xp;
 225                 } while (fp);
 226         } else {
 227                 sum_truesize = inet_frag_rbtree_purge(&q->rb_fragments);
 228         }
 229         sum = sum_truesize + f->qsize;
 230
 231         call_rcu(&q->rcu, inet_frag_destroy_rcu);
 232
 233         sub_frag_mem_limit(nf, sum);
 234 }
 235 EXPORT_SYMBOL(inet_frag_destroy);
 236
 237 static struct inet_frag_queue *inet_frag_alloc(struct netns_frags *nf,
 238                                                struct inet_frags *f,
 239                                                void *arg)
 240 {
 241         struct inet_frag_queue *q;
 242
 243         q = kmem_cache_zalloc(f->frags_cachep, GFP_ATOMIC);
 244         if (!q)
 245                 return NULL;
 246
 247         q->net = nf;
 248         f->constructor(q, arg);
 249         add_frag_mem_limit(nf, f->qsize);
 250
 251         timer_setup(&q->timer, f->frag_expire, 0);
 252         spin_lock_init(&q->lock);
 253         refcount_set(&q->refcnt, 3);
 254
 255         return q;
 256 }
 257
 258 static struct inet_frag_queue *inet_frag_create(struct netns_frags *nf,
 259                                                 void *arg,
 260                                                 struct inet_frag_queue **prev)
 261 {
 262         struct inet_frags *f = nf->f;
 263         struct inet_frag_queue *q;
 264
 265         q = inet_frag_alloc(nf, f, arg);
 266         if (!q) {
 267                 *prev = ERR_PTR(-ENOMEM);
 268                 return NULL;
 269         }
 270         mod_timer(&q->timer, jiffies + nf->timeout);
 271
 272         *prev = rhashtable_lookup_get_insert_key(&nf->rhashtable, &q->key,
 273                                                  &q->node, f->rhash_params);
 274         if (*prev) {
 275                 q->flags |= INET_FRAG_COMPLETE;
 276                 inet_frag_kill(q);
 277                 inet_frag_destroy(q);
 278                 return NULL;
 279         }
 280         return q;
 281 }
 282
 283 /* TODO : call from rcu_read_lock() and no longer use refcount_inc_not_zero() */
 284 struct inet_frag_queue *inet_frag_find(struct netns_frags *nf, void *key)
 285 {
 286         struct inet_frag_queue *fq = NULL, *prev;
 287
 288         if (!nf->high_thresh || frag_mem_limit(nf) > nf->high_thresh)
 289                 return NULL;
 290
 291         rcu_read_lock();
 292
 293         prev = rhashtable_lookup(&nf->rhashtable, key, nf->f->rhash_params);
 294         if (!prev)
 295                 fq = inet_frag_create(nf, key, &prev);
 296         if (prev && !IS_ERR(prev)) {
 297                 fq = prev;
 298                 if (!refcount_inc_not_zero(&fq->refcnt))
 299                         fq = NULL;
 300         }
 301         rcu_read_unlock();
 302         return fq;
 303 }
 304 EXPORT_SYMBOL(inet_frag_find);
 305
 306 int inet_frag_queue_insert(struct inet_frag_queue *q, struct sk_buff *skb,
 307                            int offset, int end)
 308 {
 309         struct sk_buff *last = q->fragments_tail;
 310
 311         /* RFC5722, Section 4, amended by Errata ID : 3089
 312          *                          When reassembling an IPv6 datagram, if
 313          *   one or more its constituent fragments is determined to be an
 314          *   overlapping fragment, the entire datagram (and any constituent
 315          *   fragments) MUST be silently discarded.
 316          *
 317          * Duplicates, however, should be ignored (i.e. skb dropped, but the
 318          * queue/fragments kept for later reassembly).
 319          */
 320         if (!last)
 321                 fragrun_create(q, skb);  /* First fragment. */
 322         else if (last->ip_defrag_offset + last->len < end) {
 323                 /* This is the common case: skb goes to the end. */
 324                 /* Detect and discard overlaps. */
 325                 if (offset < last->ip_defrag_offset + last->len)
 326                         return IPFRAG_OVERLAP;
 327                 if (offset == last->ip_defrag_offset + last->len)
 328                         fragrun_append_to_last(q, skb);
 329                 else
 330                         fragrun_create(q, skb);
 331         } else {
 332                 /* Binary search. Note that skb can become the first fragment,
 333                  * but not the last (covered above).
 334                  */
 335                 struct rb_node **rbn, *parent;
 336
 337                 rbn = &q->rb_fragments.rb_node;
 338                 do {
 339                         struct sk_buff *curr;
 340                         int curr_run_end;
 341
 342                         parent = *rbn;
 343                         curr = rb_to_skb(parent);
 344                         curr_run_end = curr->ip_defrag_offset +
 345                                         FRAG_CB(curr)->frag_run_len;
 346                         if (end <= curr->ip_defrag_offset)
 347                                 rbn = &parent->rb_left;
 348                         else if (offset >= curr_run_end)
 349                                 rbn = &parent->rb_right;
 350                         else if (offset >= curr->ip_defrag_offset &&
 351                                  end <= curr_run_end)
 352                                 return IPFRAG_DUP;
 353                         else
 354                                 return IPFRAG_OVERLAP;
 355                 } while (*rbn);
 356                 /* Here we have parent properly set, and rbn pointing to
 357                  * one of its NULL left/right children. Insert skb.
 358                  */
 359                 fragcb_clear(skb);
 360                 rb_link_node(&skb->rbnode, parent, rbn);
 361                 rb_insert_color(&skb->rbnode, &q->rb_fragments);
 362         }
 363
 364         skb->ip_defrag_offset = offset;
 365
 366         return IPFRAG_OK;
 367 }
 368 EXPORT_SYMBOL(inet_frag_queue_insert);
 369
 370 void *inet_frag_reasm_prepare(struct inet_frag_queue *q, struct sk_buff *skb,
 371                               struct sk_buff *parent)
 372 {
 373         struct sk_buff *fp, *head = skb_rb_first(&q->rb_fragments);
 374         struct sk_buff **nextp;
 375         int delta;
 376
 377         if (head != skb) {
 378                 fp = skb_clone(skb, GFP_ATOMIC);
 379                 if (!fp)
 380                         return NULL;
 381                 FRAG_CB(fp)->next_frag = FRAG_CB(skb)->next_frag;
 382                 if (RB_EMPTY_NODE(&skb->rbnode))
 383                         FRAG_CB(parent)->next_frag = fp;
 384                 else
 385                         rb_replace_node(&skb->rbnode, &fp->rbnode,
 386                                         &q->rb_fragments);
 387                 if (q->fragments_tail == skb)
 388                         q->fragments_tail = fp;
 389                 skb_morph(skb, head);
 390                 FRAG_CB(skb)->next_frag = FRAG_CB(head)->next_frag;
 391                 rb_replace_node(&head->rbnode, &skb->rbnode,
 392                                 &q->rb_fragments);
 393                 consume_skb(head);
 394                 head = skb;
 395         }
 396         WARN_ON(head->ip_defrag_offset != 0);
 397
 398         delta = -head->truesize;
 399
 400         /* Head of list must not be cloned. */
 401         if (skb_unclone(head, GFP_ATOMIC))
 402                 return NULL;
 403
 404         delta += head->truesize;
 405         if (delta)
 406                 add_frag_mem_limit(q->net, delta);
 407
 408         /* If the first fragment is fragmented itself, we split
 409          * it to two chunks: the first with data and paged part
 410          * and the second, holding only fragments.
 411          */
 412         if (skb_has_frag_list(head)) {
 413                 struct sk_buff *clone;
 414                 int i, plen = 0;
 415
 416                 clone = alloc_skb(0, GFP_ATOMIC);
 417                 if (!clone)
 418                         return NULL;
 419                 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
 420                 skb_frag_list_init(head);
 421                 for (i = 0; i < skb_shinfo(head)->nr_frags; i++)
 422                         plen += skb_frag_size(&skb_shinfo(head)->frags[i]);
 423                 clone->data_len = head->data_len - plen;
 424                 clone->len = clone->data_len;
 425                 head->truesize += clone->truesize;
 426                 clone->csum = 0;
 427                 clone->ip_summed = head->ip_summed;
 428                 add_frag_mem_limit(q->net, clone->truesize);
 429                 skb_shinfo(head)->frag_list = clone;
 430                 nextp = &clone->next;
 431         } else {
 432                 nextp = &skb_shinfo(head)->frag_list;
 433         }
 434
 435         return nextp;
 436 }
 437 EXPORT_SYMBOL(inet_frag_reasm_prepare);
 438
 439 void inet_frag_reasm_finish(struct inet_frag_queue *q, struct sk_buff *head,
 440                             void *reasm_data)
 441 {
 442         struct sk_buff **nextp = (struct sk_buff **)reasm_data;
 443         struct rb_node *rbn;
 444         struct sk_buff *fp;
 445
 446         skb_push(head, head->data - skb_network_header(head));
 447
 448         /* Traverse the tree in order, to build frag_list. */
 449         fp = FRAG_CB(head)->next_frag;
 450         rbn = rb_next(&head->rbnode);
 451         rb_erase(&head->rbnode, &q->rb_fragments);
 452         while (rbn || fp) {
 453                 /* fp points to the next sk_buff in the current run;
 454                  * rbn points to the next run.
 455                  */
 456                 /* Go through the current run. */
 457                 while (fp) {
 458                         *nextp = fp;
 459                         nextp = &fp->next;
 460                         fp->prev = NULL;
 461                         memset(&fp->rbnode, 0, sizeof(fp->rbnode));
 462                         fp->sk = NULL;
 463                         head->data_len += fp->len;
 464                         head->len += fp->len;
 465                         if (head->ip_summed != fp->ip_summed)
 466                                 head->ip_summed = CHECKSUM_NONE;
 467                         else if (head->ip_summed == CHECKSUM_COMPLETE)
 468                                 head->csum = csum_add(head->csum, fp->csum);
 469                         head->truesize += fp->truesize;
 470                         fp = FRAG_CB(fp)->next_frag;
 471                 }
 472                 /* Move to the next run. */
 473                 if (rbn) {
 474                         struct rb_node *rbnext = rb_next(rbn);
 475
 476                         fp = rb_to_skb(rbn);
 477                         rb_erase(rbn, &q->rb_fragments);
 478                         rbn = rbnext;
 479                 }
 480         }
 481         sub_frag_mem_limit(q->net, head->truesize);
 482
 483         *nextp = NULL;
 484         skb_mark_not_on_list(head);
 485         head->prev = NULL;
 486         head->tstamp = q->stamp;
 487 }
 488 EXPORT_SYMBOL(inet_frag_reasm_finish);
 489
 490 struct sk_buff *inet_frag_pull_head(struct inet_frag_queue *q)
 491 {
 492         struct sk_buff *head;
 493
 494         if (q->fragments) {
 495                 head = q->fragments;
 496                 q->fragments = head->next;
 497         } else {
 498                 struct sk_buff *skb;
 499
 500                 head = skb_rb_first(&q->rb_fragments);
 501                 if (!head)
 502                         return NULL;
 503                 skb = FRAG_CB(head)->next_frag;
 504                 if (skb)
 505                         rb_replace_node(&head->rbnode, &skb->rbnode,
 506                                         &q->rb_fragments);
 507                 else
 508                         rb_erase(&head->rbnode, &q->rb_fragments);
 509                 memset(&head->rbnode, 0, sizeof(head->rbnode));
 510                 barrier();
 511         }
 512         if (head == q->fragments_tail)
 513                 q->fragments_tail = NULL;
 514
 515         sub_frag_mem_limit(q->net, head->truesize);
 516
 517         return head;
 518 }
 519 EXPORT_SYMBOL(inet_frag_pull_head);