net/ipv4/ip_fragment.c

   1 /*
   2  * INET         An implementation of the TCP/IP protocol suite for the LINUX
   3  *              operating system.  INET is implemented using the  BSD Socket
   4  *              interface as the means of communication with the user level.
   5  *
   6  *              The IP fragmentation functionality.
   7  *
   8  * Version:     $Id: ip_fragment.c,v 1.59 2002/01/12 07:54:56 davem Exp $
   9  *
  10  * Authors:     Fred N. van Kempen <waltje@uWalt.NL.Mugnet.ORG>
  11  *              Alan Cox <Alan.Cox@linux.org>
  12  *
  13  * Fixes:
  14  *              Alan Cox        :       Split from ip.c , see ip_input.c for history.
  15  *              David S. Miller :       Begin massive cleanup...
  16  *              Andi Kleen      :       Add sysctls.
  17  *              xxxx            :       Overlapfrag bug.
  18  *              Ultima          :       ip_expire() kernel panic.
  19  *              Bill Hawes      :       Frag accounting and evictor fixes.
  20  *              John McDonald   :       0 length frag bug.
  21  *              Alexey Kuznetsov:       SMP races, threading, cleanup.
  22  *              Patrick McHardy :       LRU queue of frag heads for evictor.
  23  */
  24
  25 #include <linux/config.h>
  26 #include <linux/module.h>
  27 #include <linux/types.h>
  28 #include <linux/mm.h>
  29 #include <linux/jiffies.h>
  30 #include <linux/skbuff.h>
  31 #include <linux/list.h>
  32 #include <linux/ip.h>
  33 #include <linux/icmp.h>
  34 #include <linux/netdevice.h>
  35 #include <linux/jhash.h>
  36 #include <linux/random.h>
  37 #include <net/sock.h>
  38 #include <net/ip.h>
  39 #include <net/icmp.h>
  40 #include <net/checksum.h>
  41 #include <linux/tcp.h>
  42 #include <linux/udp.h>
  43 #include <linux/inet.h>
  44 #include <linux/netfilter_ipv4.h>
  45
  46 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
  47  * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
  48  * as well. Or notify me, at least. --ANK
  49  */
  50
  51 /* Fragment cache limits. We will commit 256K at one time. Should we
  52  * cross that limit we will prune down to 192K. This should cope with
  53  * even the most extreme cases without allowing an attacker to measurably
  54  * harm machine performance.
  55  */
  56 int sysctl_ipfrag_high_thresh = 256*1024;
  57 int sysctl_ipfrag_low_thresh = 192*1024;
  58
  59 /* Important NOTE! Fragment queue must be destroyed before MSL expires.
  60  * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
  61  */
  62 int sysctl_ipfrag_time = IP_FRAG_TIME;
  63
  64 struct ipfrag_skb_cb
  65 {
  66         struct inet_skb_parm    h;
  67         int                     offset;
  68 };
  69
  70 #define FRAG_CB(skb)    ((struct ipfrag_skb_cb*)((skb)->cb))
  71
  72 /* Describe an entry in the "incomplete datagrams" queue. */
  73 struct ipq {
  74         struct ipq      *next;          /* linked list pointers                 */
  75         struct list_head lru_list;      /* lru list member                      */
  76         u32             saddr;
  77         u32             daddr;
  78         u16             id;
  79         u8              protocol;
  80         u8              last_in;
  81 #define COMPLETE                4
  82 #define FIRST_IN                2
  83 #define LAST_IN                 1
  84
  85         struct sk_buff  *fragments;     /* linked list of received fragments    */
  86         int             len;            /* total length of original datagram    */
  87         int             meat;
  88         spinlock_t      lock;
  89         atomic_t        refcnt;
  90         struct timer_list timer;        /* when will this queue expire?         */
  91         struct ipq      **pprev;
  92         int             iif;
  93         struct timeval  stamp;
  94 };
  95
  96 /* Hash table. */
  97
  98 #define IPQ_HASHSZ      64
  99
 100 /* Per-bucket lock is easy to add now. */
 101 static struct ipq *ipq_hash[IPQ_HASHSZ];
 102 static rwlock_t ipfrag_lock = RW_LOCK_UNLOCKED;
 103 static u32 ipfrag_hash_rnd;
 104 static LIST_HEAD(ipq_lru_list);
 105 int ip_frag_nqueues = 0;
 106
 107 static __inline__ void __ipq_unlink(struct ipq *qp)
 108 {
 109         if(qp->next)
 110                 qp->next->pprev = qp->pprev;
 111         *qp->pprev = qp->next;
 112         list_del(&qp->lru_list);
 113         ip_frag_nqueues--;
 114 }
 115
 116 static __inline__ void ipq_unlink(struct ipq *ipq)
 117 {
 118         write_lock(&ipfrag_lock);
 119         __ipq_unlink(ipq);
 120         write_unlock(&ipfrag_lock);
 121 }
 122
 123 static unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
 124 {
 125         return jhash_3words((u32)id << 16 | prot, saddr, daddr,
 126                             ipfrag_hash_rnd) & (IPQ_HASHSZ - 1);
 127 }
 128
 129 static struct timer_list ipfrag_secret_timer;
 130 int sysctl_ipfrag_secret_interval = 10 * 60 * HZ;
 131
 132 static void ipfrag_secret_rebuild(unsigned long dummy)
 133 {
 134         unsigned long now = jiffies;
 135         int i;
 136
 137         write_lock(&ipfrag_lock);
 138         get_random_bytes(&ipfrag_hash_rnd, sizeof(u32));
 139         for (i = 0; i < IPQ_HASHSZ; i++) {
 140                 struct ipq *q;
 141
 142                 q = ipq_hash[i];
 143                 while (q) {
 144                         struct ipq *next = q->next;
 145                         unsigned int hval = ipqhashfn(q->id, q->saddr,
 146                                                       q->daddr, q->protocol);
 147
 148                         if (hval != i) {
 149                                 /* Unlink. */
 150                                 if (q->next)
 151                                         q->next->pprev = q->pprev;
 152                                 *q->pprev = q->next;
 153
 154                                 /* Relink to new hash chain. */
 155                                 if ((q->next = ipq_hash[hval]) != NULL)
 156                                         q->next->pprev = &q->next;
 157                                 ipq_hash[hval] = q;
 158                                 q->pprev = &ipq_hash[hval];
 159                         }
 160
 161                         q = next;
 162                 }
 163         }
 164         write_unlock(&ipfrag_lock);
 165
 166         mod_timer(&ipfrag_secret_timer, now + sysctl_ipfrag_secret_interval);
 167 }
 168
 169 atomic_t ip_frag_mem = ATOMIC_INIT(0);  /* Memory used for fragments */
 170
 171 /* Memory Tracking Functions. */
 172 static __inline__ void frag_kfree_skb(struct sk_buff *skb, int *work)
 173 {
 174         if (work)
 175                 *work -= skb->truesize;
 176         atomic_sub(skb->truesize, &ip_frag_mem);
 177         kfree_skb(skb);
 178 }
 179
 180 static __inline__ void frag_free_queue(struct ipq *qp, int *work)
 181 {
 182         if (work)
 183                 *work -= sizeof(struct ipq);
 184         atomic_sub(sizeof(struct ipq), &ip_frag_mem);
 185         kfree(qp);
 186 }
 187
 188 static __inline__ struct ipq *frag_alloc_queue(void)
 189 {
 190         struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
 191
 192         if(!qp)
 193                 return NULL;
 194         atomic_add(sizeof(struct ipq), &ip_frag_mem);
 195         return qp;
 196 }
 197
 198
 199 /* Destruction primitives. */
 200
 201 /* Complete destruction of ipq. */
 202 static void ip_frag_destroy(struct ipq *qp, int *work)
 203 {
 204         struct sk_buff *fp;
 205
 206         BUG_TRAP(qp->last_in&COMPLETE);
 207         BUG_TRAP(del_timer(&qp->timer) == 0);
 208
 209         /* Release all fragment data. */
 210         fp = qp->fragments;
 211         while (fp) {
 212                 struct sk_buff *xp = fp->next;
 213
 214                 frag_kfree_skb(fp, work);
 215                 fp = xp;
 216         }
 217
 218         /* Finally, release the queue descriptor itself. */
 219         frag_free_queue(qp, work);
 220 }
 221
 222 static __inline__ void ipq_put(struct ipq *ipq, int *work)
 223 {
 224         if (atomic_dec_and_test(&ipq->refcnt))
 225                 ip_frag_destroy(ipq, work);
 226 }
 227
 228 /* Kill ipq entry. It is not destroyed immediately,
 229  * because caller (and someone more) holds reference count.
 230  */
 231 static void ipq_kill(struct ipq *ipq)
 232 {
 233         if (del_timer(&ipq->timer))
 234                 atomic_dec(&ipq->refcnt);
 235
 236         if (!(ipq->last_in & COMPLETE)) {
 237                 ipq_unlink(ipq);
 238                 atomic_dec(&ipq->refcnt);
 239                 ipq->last_in |= COMPLETE;
 240         }
 241 }
 242
 243 /* Memory limiting on fragments.  Evictor trashes the oldest
 244  * fragment queue until we are back under the threshold.
 245  */
 246 static void __ip_evictor(int threshold)
 247 {
 248         struct ipq *qp;
 249         struct list_head *tmp;
 250         int work;
 251
 252         work = atomic_read(&ip_frag_mem) - threshold;
 253         if (work <= 0)
 254                 return;
 255
 256         while (work > 0) {
 257                 read_lock(&ipfrag_lock);
 258                 if (list_empty(&ipq_lru_list)) {
 259                         read_unlock(&ipfrag_lock);
 260                         return;
 261                 }
 262                 tmp = ipq_lru_list.next;
 263                 qp = list_entry(tmp, struct ipq, lru_list);
 264                 atomic_inc(&qp->refcnt);
 265                 read_unlock(&ipfrag_lock);
 266
 267                 spin_lock(&qp->lock);
 268                 if (!(qp->last_in&COMPLETE))
 269                         ipq_kill(qp);
 270                 spin_unlock(&qp->lock);
 271
 272                 ipq_put(qp, &work);
 273                 IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
 274         }
 275 }
 276
 277 static inline void ip_evictor(void)
 278 {
 279         __ip_evictor(sysctl_ipfrag_low_thresh);
 280 }
 281
 282 /*
 283  * Oops, a fragment queue timed out.  Kill it and send an ICMP reply.
 284  */
 285 static void ip_expire(unsigned long arg)
 286 {
 287         struct ipq *qp = (struct ipq *) arg;
 288
 289         spin_lock(&qp->lock);
 290
 291         if (qp->last_in & COMPLETE)
 292                 goto out;
 293
 294         ipq_kill(qp);
 295
 296         IP_INC_STATS_BH(IPSTATS_MIB_REASMTIMEOUT);
 297         IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
 298
 299         if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
 300                 struct sk_buff *head = qp->fragments;
 301                 /* Send an ICMP "Fragment Reassembly Timeout" message. */
 302                 if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
 303                         icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
 304                         dev_put(head->dev);
 305                 }
 306         }
 307 out:
 308         spin_unlock(&qp->lock);
 309         ipq_put(qp, NULL);
 310 }
 311
 312 /* Creation primitives. */
 313
 314 static struct ipq *ip_frag_intern(unsigned int hash, struct ipq *qp_in)
 315 {
 316         struct ipq *qp;
 317
 318         write_lock(&ipfrag_lock);
 319 #ifdef CONFIG_SMP
 320         /* With SMP race we have to recheck hash table, because
 321          * such entry could be created on other cpu, while we
 322          * promoted read lock to write lock.
 323          */
 324         for(qp = ipq_hash[hash]; qp; qp = qp->next) {
 325                 if(qp->id == qp_in->id          &&
 326                    qp->saddr == qp_in->saddr    &&
 327                    qp->daddr == qp_in->daddr    &&
 328                    qp->protocol == qp_in->protocol) {
 329                         atomic_inc(&qp->refcnt);
 330                         write_unlock(&ipfrag_lock);
 331                         qp_in->last_in |= COMPLETE;
 332                         ipq_put(qp_in, NULL);
 333                         return qp;
 334                 }
 335         }
 336 #endif
 337         qp = qp_in;
 338
 339         if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
 340                 atomic_inc(&qp->refcnt);
 341
 342         atomic_inc(&qp->refcnt);
 343         if((qp->next = ipq_hash[hash]) != NULL)
 344                 qp->next->pprev = &qp->next;
 345         ipq_hash[hash] = qp;
 346         qp->pprev = &ipq_hash[hash];
 347         INIT_LIST_HEAD(&qp->lru_list);
 348         list_add_tail(&qp->lru_list, &ipq_lru_list);
 349         ip_frag_nqueues++;
 350         write_unlock(&ipfrag_lock);
 351         return qp;
 352 }
 353
 354 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
 355 static struct ipq *ip_frag_create(unsigned hash, struct iphdr *iph)
 356 {
 357         struct ipq *qp;
 358
 359         if ((qp = frag_alloc_queue()) == NULL)
 360                 goto out_nomem;
 361
 362         qp->protocol = iph->protocol;
 363         qp->last_in = 0;
 364         qp->id = iph->id;
 365         qp->saddr = iph->saddr;
 366         qp->daddr = iph->daddr;
 367         qp->len = 0;
 368         qp->meat = 0;
 369         qp->fragments = NULL;
 370         qp->iif = 0;
 371
 372         /* Initialize a timer for this entry. */
 373         init_timer(&qp->timer);
 374         qp->timer.data = (unsigned long) qp;    /* pointer to queue     */
 375         qp->timer.function = ip_expire;         /* expire function      */
 376         qp->lock = SPIN_LOCK_UNLOCKED;
 377         atomic_set(&qp->refcnt, 1);
 378
 379         return ip_frag_intern(hash, qp);
 380
 381 out_nomem:
 382         NETDEBUG(if (net_ratelimit()) printk(KERN_ERR "ip_frag_create: no memory left !\n"));
 383         return NULL;
 384 }
 385
 386 /* Find the correct entry in the "incomplete datagrams" queue for
 387  * this IP datagram, and create new one, if nothing is found.
 388  */
 389 static inline struct ipq *ip_find(struct iphdr *iph)
 390 {
 391         __u16 id = iph->id;
 392         __u32 saddr = iph->saddr;
 393         __u32 daddr = iph->daddr;
 394         __u8 protocol = iph->protocol;
 395         unsigned int hash = ipqhashfn(id, saddr, daddr, protocol);
 396         struct ipq *qp;
 397
 398         read_lock(&ipfrag_lock);
 399         for(qp = ipq_hash[hash]; qp; qp = qp->next) {
 400                 if(qp->id == id         &&
 401                    qp->saddr == saddr   &&
 402                    qp->daddr == daddr   &&
 403                    qp->protocol == protocol) {
 404                         atomic_inc(&qp->refcnt);
 405                         read_unlock(&ipfrag_lock);
 406                         return qp;
 407                 }
 408         }
 409         read_unlock(&ipfrag_lock);
 410
 411         return ip_frag_create(hash, iph);
 412 }
 413
 414 /* Add new segment to existing queue. */
 415 static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 416 {
 417         struct sk_buff *prev, *next;
 418         int flags, offset;
 419         int ihl, end;
 420
 421         if (qp->last_in & COMPLETE)
 422                 goto err;
 423
 424         offset = ntohs(skb->nh.iph->frag_off);
 425         flags = offset & ~IP_OFFSET;
 426         offset &= IP_OFFSET;
 427         offset <<= 3;           /* offset is in 8-byte chunks */
 428         ihl = skb->nh.iph->ihl * 4;
 429
 430         /* Determine the position of this fragment. */
 431         end = offset + skb->len - ihl;
 432
 433         /* Is this the final fragment? */
 434         if ((flags & IP_MF) == 0) {
 435                 /* If we already have some bits beyond end
 436                  * or have different end, the segment is corrrupted.
 437                  */
 438                 if (end < qp->len ||
 439                     ((qp->last_in & LAST_IN) && end != qp->len))
 440                         goto err;
 441                 qp->last_in |= LAST_IN;
 442                 qp->len = end;
 443         } else {
 444                 if (end&7) {
 445                         end &= ~7;
 446                         if (skb->ip_summed != CHECKSUM_UNNECESSARY)
 447                                 skb->ip_summed = CHECKSUM_NONE;
 448                 }
 449                 if (end > qp->len) {
 450                         /* Some bits beyond end -> corruption. */
 451                         if (qp->last_in & LAST_IN)
 452                                 goto err;
 453                         qp->len = end;
 454                 }
 455         }
 456         if (end == offset)
 457                 goto err;
 458
 459         if (pskb_pull(skb, ihl) == NULL)
 460                 goto err;
 461         if (pskb_trim(skb, end-offset))
 462                 goto err;
 463
 464         /* Find out which fragments are in front and at the back of us
 465          * in the chain of fragments so far.  We must know where to put
 466          * this fragment, right?
 467          */
 468         prev = NULL;
 469         for(next = qp->fragments; next != NULL; next = next->next) {
 470                 if (FRAG_CB(next)->offset >= offset)
 471                         break;  /* bingo! */
 472                 prev = next;
 473         }
 474
 475         /* We found where to put this one.  Check for overlap with
 476          * preceding fragment, and, if needed, align things so that
 477          * any overlaps are eliminated.
 478          */
 479         if (prev) {
 480                 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
 481
 482                 if (i > 0) {
 483                         offset += i;
 484                         if (end <= offset)
 485                                 goto err;
 486                         if (!pskb_pull(skb, i))
 487                                 goto err;
 488                         if (skb->ip_summed != CHECKSUM_UNNECESSARY)
 489                                 skb->ip_summed = CHECKSUM_NONE;
 490                 }
 491         }
 492
 493         while (next && FRAG_CB(next)->offset < end) {
 494                 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
 495
 496                 if (i < next->len) {
 497                         /* Eat head of the next overlapped fragment
 498                          * and leave the loop. The next ones cannot overlap.
 499                          */
 500                         if (!pskb_pull(next, i))
 501                                 goto err;
 502                         FRAG_CB(next)->offset += i;
 503                         qp->meat -= i;
 504                         if (next->ip_summed != CHECKSUM_UNNECESSARY)
 505                                 next->ip_summed = CHECKSUM_NONE;
 506                         break;
 507                 } else {
 508                         struct sk_buff *free_it = next;
 509
 510                         /* Old fragmnet is completely overridden with
 511                          * new one drop it.
 512                          */
 513                         next = next->next;
 514
 515                         if (prev)
 516                                 prev->next = next;
 517                         else
 518                                 qp->fragments = next;
 519
 520                         qp->meat -= free_it->len;
 521                         frag_kfree_skb(free_it, NULL);
 522                 }
 523         }
 524
 525         FRAG_CB(skb)->offset = offset;
 526
 527         /* Insert this fragment in the chain of fragments. */
 528         skb->next = next;
 529         if (prev)
 530                 prev->next = skb;
 531         else
 532                 qp->fragments = skb;
 533
 534         if (skb->dev)
 535                 qp->iif = skb->dev->ifindex;
 536         skb->dev = NULL;
 537         qp->stamp = skb->stamp;
 538         qp->meat += skb->len;
 539         atomic_add(skb->truesize, &ip_frag_mem);
 540         if (offset == 0)
 541                 qp->last_in |= FIRST_IN;
 542
 543         write_lock(&ipfrag_lock);
 544         list_move_tail(&qp->lru_list, &ipq_lru_list);
 545         write_unlock(&ipfrag_lock);
 546
 547         return;
 548
 549 err:
 550         kfree_skb(skb);
 551 }
 552
 553
 554 /* Build a new IP datagram from all its fragments. */
 555
 556 static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
 557 {
 558         struct iphdr *iph;
 559         struct sk_buff *fp, *head = qp->fragments;
 560         int len;
 561         int ihlen;
 562
 563         ipq_kill(qp);
 564
 565         BUG_TRAP(head != NULL);
 566         BUG_TRAP(FRAG_CB(head)->offset == 0);
 567
 568         /* Allocate a new buffer for the datagram. */
 569         ihlen = head->nh.iph->ihl*4;
 570         len = ihlen + qp->len;
 571
 572         if(len > 65535)
 573                 goto out_oversize;
 574
 575         /* Head of list must not be cloned. */
 576         if (skb_cloned(head) && pskb_expand_head(head, 0, 0, GFP_ATOMIC))
 577                 goto out_nomem;
 578
 579         /* If the first fragment is fragmented itself, we split
 580          * it to two chunks: the first with data and paged part
 581          * and the second, holding only fragments. */
 582         if (skb_shinfo(head)->frag_list) {
 583                 struct sk_buff *clone;
 584                 int i, plen = 0;
 585
 586                 if ((clone = alloc_skb(0, GFP_ATOMIC)) == NULL)
 587                         goto out_nomem;
 588                 clone->next = head->next;
 589                 head->next = clone;
 590                 skb_shinfo(clone)->frag_list = skb_shinfo(head)->frag_list;
 591                 skb_shinfo(head)->frag_list = NULL;
 592                 for (i=0; i<skb_shinfo(head)->nr_frags; i++)
 593                         plen += skb_shinfo(head)->frags[i].size;
 594                 clone->len = clone->data_len = head->data_len - plen;
 595                 head->data_len -= clone->len;
 596                 head->len -= clone->len;
 597                 clone->csum = 0;
 598                 clone->ip_summed = head->ip_summed;
 599                 atomic_add(clone->truesize, &ip_frag_mem);
 600         }
 601
 602         skb_shinfo(head)->frag_list = head->next;
 603         skb_push(head, head->data - head->nh.raw);
 604         atomic_sub(head->truesize, &ip_frag_mem);
 605
 606         for (fp=head->next; fp; fp = fp->next) {
 607                 head->data_len += fp->len;
 608                 head->len += fp->len;
 609                 if (head->ip_summed != fp->ip_summed)
 610                         head->ip_summed = CHECKSUM_NONE;
 611                 else if (head->ip_summed == CHECKSUM_HW)
 612                         head->csum = csum_add(head->csum, fp->csum);
 613                 head->truesize += fp->truesize;
 614                 atomic_sub(fp->truesize, &ip_frag_mem);
 615         }
 616
 617         head->next = NULL;
 618         head->dev = dev;
 619         head->stamp = qp->stamp;
 620
 621         iph = head->nh.iph;
 622         iph->frag_off = 0;
 623         iph->tot_len = htons(len);
 624         IP_INC_STATS_BH(IPSTATS_MIB_REASMOKS);
 625         qp->fragments = NULL;
 626         return head;
 627
 628 out_nomem:
 629         NETDEBUG(if (net_ratelimit())
 630                  printk(KERN_ERR
 631                         "IP: queue_glue: no memory for gluing queue %p\n",
 632                         qp));
 633         goto out_fail;
 634 out_oversize:
 635         if (net_ratelimit())
 636                 printk(KERN_INFO
 637                         "Oversized IP packet from %d.%d.%d.%d.\n",
 638                         NIPQUAD(qp->saddr));
 639 out_fail:
 640         IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
 641         return NULL;
 642 }
 643
 644 /* Process an incoming IP datagram fragment. */
 645 struct sk_buff *ip_defrag(struct sk_buff *skb)
 646 {
 647         struct iphdr *iph = skb->nh.iph;
 648         struct ipq *qp;
 649         struct net_device *dev;
 650
 651         IP_INC_STATS_BH(IPSTATS_MIB_REASMREQDS);
 652
 653         /* Start by cleaning up the memory. */
 654         if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
 655                 ip_evictor();
 656
 657         dev = skb->dev;
 658
 659         /* Lookup (or create) queue header */
 660         if ((qp = ip_find(iph)) != NULL) {
 661                 struct sk_buff *ret = NULL;
 662
 663                 spin_lock(&qp->lock);
 664
 665                 ip_frag_queue(qp, skb);
 666
 667                 if (qp->last_in == (FIRST_IN|LAST_IN) &&
 668                     qp->meat == qp->len)
 669                         ret = ip_frag_reasm(qp, dev);
 670
 671                 spin_unlock(&qp->lock);
 672                 ipq_put(qp, NULL);
 673                 return ret;
 674         }
 675
 676         IP_INC_STATS_BH(IPSTATS_MIB_REASMFAILS);
 677         kfree_skb(skb);
 678         return NULL;
 679 }
 680
 681 void ipfrag_init(void)
 682 {
 683         ipfrag_hash_rnd = (u32) ((num_physpages ^ (num_physpages>>7)) ^
 684                                  (jiffies ^ (jiffies >> 6)));
 685
 686         init_timer(&ipfrag_secret_timer);
 687         ipfrag_secret_timer.function = ipfrag_secret_rebuild;
 688         ipfrag_secret_timer.expires = jiffies + sysctl_ipfrag_secret_interval;
 689         add_timer(&ipfrag_secret_timer);
 690 }
 691
 692 void ipfrag_flush(void)
 693 {
 694         __ip_evictor(0);
 695 }
 696
 697 EXPORT_SYMBOL(ip_defrag);
 698 EXPORT_SYMBOL(ipfrag_flush);