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 1.53 2000/12/08 17:15:53 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  */
  23
  24 #include <linux/config.h>
  25 #include <linux/types.h>
  26 #include <linux/mm.h>
  27 #include <linux/sched.h>
  28 #include <linux/skbuff.h>
  29 #include <linux/ip.h>
  30 #include <linux/icmp.h>
  31 #include <linux/netdevice.h>
  32 #include <net/sock.h>
  33 #include <net/ip.h>
  34 #include <net/icmp.h>
  35 #include <net/checksum.h>
  36 #include <linux/tcp.h>
  37 #include <linux/udp.h>
  38 #include <linux/inet.h>
  39 #include <linux/netfilter_ipv4.h>
  40
  41 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
  42  * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
  43  * as well. Or notify me, at least. --ANK
  44  */
  45
  46 /* Fragment cache limits. We will commit 256K at one time. Should we
  47  * cross that limit we will prune down to 192K. This should cope with
  48  * even the most extreme cases without allowing an attacker to measurably
  49  * harm machine performance.
  50  */
  51 int sysctl_ipfrag_high_thresh = 256*1024;
  52 int sysctl_ipfrag_low_thresh = 192*1024;
  53
  54 /* Important NOTE! Fragment queue must be destroyed before MSL expires.
  55  * RFC791 is wrong proposing to prolongate timer each fragment arrival by TTL.
  56  */
  57 int sysctl_ipfrag_time = IP_FRAG_TIME;
  58
  59 struct ipfrag_skb_cb
  60 {
  61         struct inet_skb_parm    h;
  62         int                     offset;
  63 };
  64
  65 #define FRAG_CB(skb)    ((struct ipfrag_skb_cb*)((skb)->cb))
  66
  67 /* Describe an entry in the "incomplete datagrams" queue. */
  68 struct ipq {
  69         struct ipq      *next;          /* linked list pointers                 */
  70         u32             saddr;
  71         u32             daddr;
  72         u16             id;
  73         u8              protocol;
  74         u8              last_in;
  75 #define COMPLETE                4
  76 #define FIRST_IN                2
  77 #define LAST_IN                 1
  78
  79         struct sk_buff  *fragments;     /* linked list of received fragments    */
  80         int             len;            /* total length of original datagram    */
  81         int             meat;
  82         spinlock_t      lock;
  83         atomic_t        refcnt;
  84         struct timer_list timer;        /* when will this queue expire?         */
  85         struct ipq      **pprev;
  86         int             iif;            /* Device index - for icmp replies      */
  87 };
  88
  89 /* Hash table. */
  90
  91 #define IPQ_HASHSZ      64
  92
  93 /* Per-bucket lock is easy to add now. */
  94 static struct ipq *ipq_hash[IPQ_HASHSZ];
  95 static rwlock_t ipfrag_lock = RW_LOCK_UNLOCKED;
  96 int ip_frag_nqueues = 0;
  97
  98 static __inline__ void __ipq_unlink(struct ipq *qp)
  99 {
 100         if(qp->next)
 101                 qp->next->pprev = qp->pprev;
 102         *qp->pprev = qp->next;
 103         ip_frag_nqueues--;
 104 }
 105
 106 static __inline__ void ipq_unlink(struct ipq *ipq)
 107 {
 108         write_lock(&ipfrag_lock);
 109         __ipq_unlink(ipq);
 110         write_unlock(&ipfrag_lock);
 111 }
 112
 113 /*
 114  * Was: ((((id) >> 1) ^ (saddr) ^ (daddr) ^ (prot)) & (IPQ_HASHSZ - 1))
 115  *
 116  * I see, I see evil hand of bigendian mafia. On Intel all the packets hit
 117  * one hash bucket with this hash function. 8)
 118  */
 119 static __inline__ unsigned int ipqhashfn(u16 id, u32 saddr, u32 daddr, u8 prot)
 120 {
 121         unsigned int h = saddr ^ daddr;
 122
 123         h ^= (h>>16)^id;
 124         h ^= (h>>8)^prot;
 125         return h & (IPQ_HASHSZ - 1);
 126 }
 127
 128
 129 atomic_t ip_frag_mem = ATOMIC_INIT(0);  /* Memory used for fragments */
 130
 131 /* Memory Tracking Functions. */
 132 extern __inline__ void frag_kfree_skb(struct sk_buff *skb)
 133 {
 134         atomic_sub(skb->truesize, &ip_frag_mem);
 135         kfree_skb(skb);
 136 }
 137
 138 extern __inline__ void frag_free_queue(struct ipq *qp)
 139 {
 140         atomic_sub(sizeof(struct ipq), &ip_frag_mem);
 141         kfree(qp);
 142 }
 143
 144 extern __inline__ struct ipq *frag_alloc_queue(void)
 145 {
 146         struct ipq *qp = kmalloc(sizeof(struct ipq), GFP_ATOMIC);
 147
 148         if(!qp)
 149                 return NULL;
 150         atomic_add(sizeof(struct ipq), &ip_frag_mem);
 151         return qp;
 152 }
 153
 154
 155 /* Destruction primitives. */
 156
 157 /* Complete destruction of ipq. */
 158 static void ip_frag_destroy(struct ipq *qp)
 159 {
 160         struct sk_buff *fp;
 161
 162         BUG_TRAP(qp->last_in&COMPLETE);
 163         BUG_TRAP(del_timer(&qp->timer) == 0);
 164
 165         /* Release all fragment data. */
 166         fp = qp->fragments;
 167         while (fp) {
 168                 struct sk_buff *xp = fp->next;
 169
 170                 frag_kfree_skb(fp);
 171                 fp = xp;
 172         }
 173
 174         /* Finally, release the queue descriptor itself. */
 175         frag_free_queue(qp);
 176 }
 177
 178 static __inline__ void ipq_put(struct ipq *ipq)
 179 {
 180         if (atomic_dec_and_test(&ipq->refcnt))
 181                 ip_frag_destroy(ipq);
 182 }
 183
 184 /* Kill ipq entry. It is not destroyed immediately,
 185  * because caller (and someone more) holds reference count.
 186  */
 187 static __inline__ void ipq_kill(struct ipq *ipq)
 188 {
 189         if (del_timer(&ipq->timer))
 190                 atomic_dec(&ipq->refcnt);
 191
 192         if (!(ipq->last_in & COMPLETE)) {
 193                 ipq_unlink(ipq);
 194                 atomic_dec(&ipq->refcnt);
 195                 ipq->last_in |= COMPLETE;
 196         }
 197 }
 198
 199 /* Memory limiting on fragments.  Evictor trashes the oldest
 200  * fragment queue until we are back under the low threshold.
 201  */
 202 static void ip_evictor(void)
 203 {
 204         int i, progress;
 205
 206         do {
 207                 if (atomic_read(&ip_frag_mem) <= sysctl_ipfrag_low_thresh)
 208                         return;
 209                 progress = 0;
 210                 /* FIXME: Make LRU queue of frag heads. -DaveM */
 211                 for (i = 0; i < IPQ_HASHSZ; i++) {
 212                         struct ipq *qp;
 213                         if (ipq_hash[i] == NULL)
 214                                 continue;
 215
 216                         write_lock(&ipfrag_lock);
 217                         if ((qp = ipq_hash[i]) != NULL) {
 218                                 /* find the oldest queue for this hash bucket */
 219                                 while (qp->next)
 220                                         qp = qp->next;
 221                                 __ipq_unlink(qp);
 222                                 write_unlock(&ipfrag_lock);
 223
 224                                 spin_lock(&qp->lock);
 225                                 if (del_timer(&qp->timer))
 226                                         atomic_dec(&qp->refcnt);
 227                                 qp->last_in |= COMPLETE;
 228                                 spin_unlock(&qp->lock);
 229
 230                                 ipq_put(qp);
 231                                 IP_INC_STATS_BH(IpReasmFails);
 232                                 progress = 1;
 233                                 continue;
 234                         }
 235                         write_unlock(&ipfrag_lock);
 236                 }
 237         } while (progress);
 238 }
 239
 240 /*
 241  * Oops, a fragment queue timed out.  Kill it and send an ICMP reply.
 242  */
 243 static void ip_expire(unsigned long arg)
 244 {
 245         struct ipq *qp = (struct ipq *) arg;
 246
 247         spin_lock(&qp->lock);
 248
 249         if (qp->last_in & COMPLETE)
 250                 goto out;
 251
 252         ipq_kill(qp);
 253
 254         IP_INC_STATS_BH(IpReasmTimeout);
 255         IP_INC_STATS_BH(IpReasmFails);
 256
 257         if ((qp->last_in&FIRST_IN) && qp->fragments != NULL) {
 258                 struct sk_buff *head = qp->fragments;
 259
 260                 /* Send an ICMP "Fragment Reassembly Timeout" message. */
 261                 if ((head->dev = dev_get_by_index(qp->iif)) != NULL) {
 262                         icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
 263                         dev_put(head->dev);
 264                 }
 265         }
 266 out:
 267         spin_unlock(&qp->lock);
 268         ipq_put(qp);
 269 }
 270
 271 /* Creation primitives. */
 272
 273 static struct ipq *ip_frag_intern(unsigned int hash, struct ipq *qp_in)
 274 {
 275         struct ipq *qp;
 276
 277         write_lock(&ipfrag_lock);
 278 #ifdef CONFIG_SMP
 279         /* With SMP race we have to recheck hash table, because
 280          * such entry could be created on other cpu, while we
 281          * promoted read lock to write lock.
 282          */
 283         for(qp = ipq_hash[hash]; qp; qp = qp->next) {
 284                 if(qp->id == qp_in->id          &&
 285                    qp->saddr == qp_in->saddr    &&
 286                    qp->daddr == qp_in->daddr    &&
 287                    qp->protocol == qp_in->protocol) {
 288                         atomic_inc(&qp->refcnt);
 289                         write_unlock(&ipfrag_lock);
 290                         qp_in->last_in |= COMPLETE;
 291                         ipq_put(qp_in);
 292                         return qp;
 293                 }
 294         }
 295 #endif
 296         qp = qp_in;
 297
 298         if (!mod_timer(&qp->timer, jiffies + sysctl_ipfrag_time))
 299                 atomic_inc(&qp->refcnt);
 300
 301         atomic_inc(&qp->refcnt);
 302         if((qp->next = ipq_hash[hash]) != NULL)
 303                 qp->next->pprev = &qp->next;
 304         ipq_hash[hash] = qp;
 305         qp->pprev = &ipq_hash[hash];
 306         ip_frag_nqueues++;
 307         write_unlock(&ipfrag_lock);
 308         return qp;
 309 }
 310
 311 /* Add an entry to the 'ipq' queue for a newly received IP datagram. */
 312 static struct ipq *ip_frag_create(unsigned hash, struct iphdr *iph)
 313 {
 314         struct ipq *qp;
 315
 316         if ((qp = frag_alloc_queue()) == NULL)
 317                 goto out_nomem;
 318
 319         qp->protocol = iph->protocol;
 320         qp->last_in = 0;
 321         qp->id = iph->id;
 322         qp->saddr = iph->saddr;
 323         qp->daddr = iph->daddr;
 324         qp->len = 0;
 325         qp->meat = 0;
 326         qp->fragments = NULL;
 327         qp->iif = 0;
 328
 329         /* Initialize a timer for this entry. */
 330         init_timer(&qp->timer);
 331         qp->timer.data = (unsigned long) qp;    /* pointer to queue     */
 332         qp->timer.function = ip_expire;         /* expire function      */
 333         qp->lock = SPIN_LOCK_UNLOCKED;
 334         atomic_set(&qp->refcnt, 1);
 335
 336         return ip_frag_intern(hash, qp);
 337
 338 out_nomem:
 339         NETDEBUG(printk(KERN_ERR "ip_frag_create: no memory left !\n"));
 340         return NULL;
 341 }
 342
 343 /* Find the correct entry in the "incomplete datagrams" queue for
 344  * this IP datagram, and create new one, if nothing is found.
 345  */
 346 static inline struct ipq *ip_find(struct iphdr *iph)
 347 {
 348         __u16 id = iph->id;
 349         __u32 saddr = iph->saddr;
 350         __u32 daddr = iph->daddr;
 351         __u8 protocol = iph->protocol;
 352         unsigned int hash = ipqhashfn(id, saddr, daddr, protocol);
 353         struct ipq *qp;
 354
 355         read_lock(&ipfrag_lock);
 356         for(qp = ipq_hash[hash]; qp; qp = qp->next) {
 357                 if(qp->id == id         &&
 358                    qp->saddr == saddr   &&
 359                    qp->daddr == daddr   &&
 360                    qp->protocol == protocol) {
 361                         atomic_inc(&qp->refcnt);
 362                         read_unlock(&ipfrag_lock);
 363                         return qp;
 364                 }
 365         }
 366         read_unlock(&ipfrag_lock);
 367
 368         return ip_frag_create(hash, iph);
 369 }
 370
 371 /* Add new segment to existing queue. */
 372 static void ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
 373 {
 374         struct iphdr *iph = skb->nh.iph;
 375         struct sk_buff *prev, *next;
 376         int flags, offset;
 377         int ihl, end;
 378
 379         if (qp->last_in & COMPLETE)
 380                 goto err;
 381
 382         offset = ntohs(iph->frag_off);
 383         flags = offset & ~IP_OFFSET;
 384         offset &= IP_OFFSET;
 385         offset <<= 3;           /* offset is in 8-byte chunks */
 386         ihl = iph->ihl * 4;
 387
 388         /* Determine the position of this fragment. */
 389         end = offset + (ntohs(iph->tot_len) - ihl);
 390
 391         /* Is this the final fragment? */
 392         if ((flags & IP_MF) == 0) {
 393                 /* If we already have some bits beyond end
 394                  * or have different end, the segment is corrrupted.
 395                  */
 396                 if (end < qp->len ||
 397                     ((qp->last_in & LAST_IN) && end != qp->len))
 398                         goto err;
 399                 qp->last_in |= LAST_IN;
 400                 qp->len = end;
 401         } else {
 402                 if (end&7) {
 403                         end &= ~7;
 404                         if (skb->ip_summed != CHECKSUM_UNNECESSARY)
 405                                 skb->ip_summed = CHECKSUM_NONE;
 406                 }
 407                 if (end > qp->len) {
 408                         /* Some bits beyond end -> corruption. */
 409                         if (qp->last_in & LAST_IN)
 410                                 goto err;
 411                         qp->len = end;
 412                 }
 413         }
 414         if (end == offset)
 415                 goto err;
 416
 417         /* Point into the IP datagram 'data' part. */
 418         skb_pull(skb, (skb->nh.raw+ihl) - skb->data);
 419         skb_trim(skb, end - offset);
 420
 421         /* Find out which fragments are in front and at the back of us
 422          * in the chain of fragments so far.  We must know where to put
 423          * this fragment, right?
 424          */
 425         prev = NULL;
 426         for(next = qp->fragments; next != NULL; next = next->next) {
 427                 if (FRAG_CB(next)->offset >= offset)
 428                         break;  /* bingo! */
 429                 prev = next;
 430         }
 431
 432         /* We found where to put this one.  Check for overlap with
 433          * preceding fragment, and, if needed, align things so that
 434          * any overlaps are eliminated.
 435          */
 436         if (prev) {
 437                 int i = (FRAG_CB(prev)->offset + prev->len) - offset;
 438
 439                 if (i > 0) {
 440                         offset += i;
 441                         if (end <= offset)
 442                                 goto err;
 443                         skb_pull(skb, i);
 444                         if (skb->ip_summed != CHECKSUM_UNNECESSARY)
 445                                 skb->ip_summed = CHECKSUM_NONE;
 446                 }
 447         }
 448
 449         while (next && FRAG_CB(next)->offset < end) {
 450                 int i = end - FRAG_CB(next)->offset; /* overlap is 'i' bytes */
 451
 452                 if (i < next->len) {
 453                         /* Eat head of the next overlapped fragment
 454                          * and leave the loop. The next ones cannot overlap.
 455                          */
 456                         FRAG_CB(next)->offset += i;
 457                         skb_pull(next, i);
 458                         qp->meat -= i;
 459                         if (next->ip_summed != CHECKSUM_UNNECESSARY)
 460                                 next->ip_summed = CHECKSUM_NONE;
 461                         break;
 462                 } else {
 463                         struct sk_buff *free_it = next;
 464
 465                         /* Old fragmnet is completely overridden with
 466                          * new one drop it.
 467                          */
 468                         next = next->next;
 469
 470                         if (prev)
 471                                 prev->next = next;
 472                         else
 473                                 qp->fragments = next;
 474
 475                         qp->meat -= free_it->len;
 476                         frag_kfree_skb(free_it);
 477                 }
 478         }
 479
 480         FRAG_CB(skb)->offset = offset;
 481
 482         /* Insert this fragment in the chain of fragments. */
 483         skb->next = next;
 484         if (prev)
 485                 prev->next = skb;
 486         else
 487                 qp->fragments = skb;
 488
 489         if (skb->dev)
 490                 qp->iif = skb->dev->ifindex;
 491         skb->dev = NULL;
 492         qp->meat += skb->len;
 493         atomic_add(skb->truesize, &ip_frag_mem);
 494         if (offset == 0)
 495                 qp->last_in |= FIRST_IN;
 496
 497         return;
 498
 499 err:
 500         kfree_skb(skb);
 501 }
 502
 503
 504 /* Build a new IP datagram from all its fragments.
 505  *
 506  * FIXME: We copy here because we lack an effective way of handling lists
 507  * of bits on input. Until the new skb data handling is in I'm not going
 508  * to touch this with a bargepole.
 509  */
 510 static struct sk_buff *ip_frag_reasm(struct ipq *qp, struct net_device *dev)
 511 {
 512         struct sk_buff *skb;
 513         struct iphdr *iph;
 514         struct sk_buff *fp, *head = qp->fragments;
 515         int len;
 516         int ihlen;
 517
 518         ipq_kill(qp);
 519
 520         BUG_TRAP(head != NULL);
 521         BUG_TRAP(FRAG_CB(head)->offset == 0);
 522
 523         /* Allocate a new buffer for the datagram. */
 524         ihlen = head->nh.iph->ihl*4;
 525         len = ihlen + qp->len;
 526
 527         if(len > 65535)
 528                 goto out_oversize;
 529
 530         skb = dev_alloc_skb(len);
 531         if (!skb)
 532                 goto out_nomem;
 533
 534         /* Fill in the basic details. */
 535         skb->mac.raw = skb->data;
 536         skb->nh.raw = skb->data;
 537         FRAG_CB(skb)->h = FRAG_CB(head)->h;
 538         skb->ip_summed = head->ip_summed;
 539         skb->csum = 0;
 540
 541         /* Copy the original IP headers into the new buffer. */
 542         memcpy(skb_put(skb, ihlen), head->nh.iph, ihlen);
 543
 544         /* Copy the data portions of all fragments into the new buffer. */
 545         for (fp=head; fp; fp = fp->next) {
 546                 memcpy(skb_put(skb, fp->len), fp->data, fp->len);
 547
 548                 if (skb->ip_summed != fp->ip_summed)
 549                         skb->ip_summed = CHECKSUM_NONE;
 550                 else if (skb->ip_summed == CHECKSUM_HW)
 551                         skb->csum = csum_add(skb->csum, fp->csum);
 552         }
 553
 554         skb->dst = dst_clone(head->dst);
 555         skb->pkt_type = head->pkt_type;
 556         skb->protocol = head->protocol;
 557         skb->dev = dev;
 558
 559         /*
 560         *  Clearly bogus, because security markings of the individual
 561         *  fragments should have been checked for consistency before
 562         *  gluing, and intermediate coalescing of fragments may have
 563         *  taken place in ip_defrag() before ip_glue() ever got called.
 564         *  If we're not going to do the consistency checking, we might
 565         *  as well take the value associated with the first fragment.
 566         *       --rct
 567         */
 568         skb->security = head->security;
 569
 570 #ifdef CONFIG_NETFILTER
 571         /* Connection association is same as fragment (if any). */
 572         skb->nfct = head->nfct;
 573         nf_conntrack_get(skb->nfct);
 574 #ifdef CONFIG_NETFILTER_DEBUG
 575         skb->nf_debug = head->nf_debug;
 576 #endif
 577 #endif
 578
 579         /* Done with all fragments. Fixup the new IP header. */
 580         iph = skb->nh.iph;
 581         iph->frag_off = 0;
 582         iph->tot_len = htons(len);
 583         IP_INC_STATS_BH(IpReasmOKs);
 584         return skb;
 585
 586 out_nomem:
 587         NETDEBUG(printk(KERN_ERR
 588                         "IP: queue_glue: no memory for gluing queue %p\n",
 589                         qp));
 590         goto out_fail;
 591 out_oversize:
 592         if (net_ratelimit())
 593                 printk(KERN_INFO
 594                         "Oversized IP packet from %d.%d.%d.%d.\n",
 595                         NIPQUAD(qp->saddr));
 596 out_fail:
 597         IP_INC_STATS_BH(IpReasmFails);
 598         return NULL;
 599 }
 600
 601 /* Process an incoming IP datagram fragment. */
 602 struct sk_buff *ip_defrag(struct sk_buff *skb)
 603 {
 604         struct iphdr *iph = skb->nh.iph;
 605         struct ipq *qp;
 606         struct net_device *dev;
 607
 608         IP_INC_STATS_BH(IpReasmReqds);
 609
 610         /* Start by cleaning up the memory. */
 611         if (atomic_read(&ip_frag_mem) > sysctl_ipfrag_high_thresh)
 612                 ip_evictor();
 613
 614         dev = skb->dev;
 615
 616         /* Lookup (or create) queue header */
 617         if ((qp = ip_find(iph)) != NULL) {
 618                 struct sk_buff *ret = NULL;
 619
 620                 spin_lock(&qp->lock);
 621
 622                 ip_frag_queue(qp, skb);
 623
 624                 if (qp->last_in == (FIRST_IN|LAST_IN) &&
 625                     qp->meat == qp->len)
 626                         ret = ip_frag_reasm(qp, dev);
 627
 628                 spin_unlock(&qp->lock);
 629                 ipq_put(qp);
 630                 return ret;
 631         }
 632
 633         IP_INC_STATS_BH(IpReasmFails);
 634         kfree_skb(skb);
 635         return NULL;
 636 }