1 // SPDX-License-Identifier: GPL-2.0
3 * INET An implementation of the TCP/IP protocol suite for the LINUX
4 * operating system. INET is implemented using the BSD Socket
5 * interface as the means of communication with the user level.
7 * The IP fragmentation functionality.
9 * Authors: Fred N. van Kempen <waltje@uWalt.NL.Mugnet.ORG>
10 * Alan Cox <alan@lxorguk.ukuu.org.uk>
13 * Alan Cox : Split from ip.c , see ip_input.c for history.
14 * David S. Miller : Begin massive cleanup...
15 * Andi Kleen : Add sysctls.
16 * xxxx : Overlapfrag bug.
17 * Ultima : ip_expire() kernel panic.
18 * Bill Hawes : Frag accounting and evictor fixes.
19 * John McDonald : 0 length frag bug.
20 * Alexey Kuznetsov: SMP races, threading, cleanup.
21 * Patrick McHardy : LRU queue of frag heads for evictor.
24 #define pr_fmt(fmt) "IPv4: " fmt
26 #include <linux/compiler.h>
27 #include <linux/module.h>
28 #include <linux/types.h>
30 #include <linux/jiffies.h>
31 #include <linux/skbuff.h>
32 #include <linux/list.h>
34 #include <linux/icmp.h>
35 #include <linux/netdevice.h>
36 #include <linux/jhash.h>
37 #include <linux/random.h>
38 #include <linux/slab.h>
39 #include <net/route.h>
44 #include <net/checksum.h>
45 #include <net/inetpeer.h>
46 #include <net/inet_frag.h>
47 #include <linux/tcp.h>
48 #include <linux/udp.h>
49 #include <linux/inet.h>
50 #include <linux/netfilter_ipv4.h>
51 #include <net/inet_ecn.h>
52 #include <net/l3mdev.h>
54 /* NOTE. Logic of IP defragmentation is parallel to corresponding IPv6
55 * code now. If you change something here, _PLEASE_ update ipv6/reassembly.c
56 * as well. Or notify me, at least. --ANK
58 static const char ip_frag_cache_name[] = "ip4-frags";
60 /* Describe an entry in the "incomplete datagrams" queue. */
62 struct inet_frag_queue q;
64 u8 ecn; /* RFC3168 support */
65 u16 max_df_size; /* largest frag with DF set seen */
68 struct inet_peer *peer;
71 static u8 ip4_frag_ecn(u8 tos)
73 return 1 << (tos & INET_ECN_MASK);
76 static struct inet_frags ip4_frags;
78 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
79 struct sk_buff *prev_tail, struct net_device *dev);
82 static void ip4_frag_init(struct inet_frag_queue *q, const void *a)
84 struct ipq *qp = container_of(q, struct ipq, q);
85 struct net *net = q->fqdir->net;
87 const struct frag_v4_compare_key *key = a;
91 qp->peer = q->fqdir->max_dist ?
92 inet_getpeer_v4(net->ipv4.peers, key->saddr, key->vif, 1) :
96 static void ip4_frag_free(struct inet_frag_queue *q)
100 qp = container_of(q, struct ipq, q);
102 inet_putpeer(qp->peer);
106 /* Destruction primitives. */
108 static void ipq_put(struct ipq *ipq)
110 inet_frag_put(&ipq->q);
113 /* Kill ipq entry. It is not destroyed immediately,
114 * because caller (and someone more) holds reference count.
116 static void ipq_kill(struct ipq *ipq)
118 inet_frag_kill(&ipq->q);
121 static bool frag_expire_skip_icmp(u32 user)
123 return user == IP_DEFRAG_AF_PACKET ||
124 ip_defrag_user_in_between(user, IP_DEFRAG_CONNTRACK_IN,
125 __IP_DEFRAG_CONNTRACK_IN_END) ||
126 ip_defrag_user_in_between(user, IP_DEFRAG_CONNTRACK_BRIDGE_IN,
127 __IP_DEFRAG_CONNTRACK_BRIDGE_IN);
131 * Oops, a fragment queue timed out. Kill it and send an ICMP reply.
133 static void ip_expire(struct timer_list *t)
135 struct inet_frag_queue *frag = from_timer(frag, t, timer);
136 const struct iphdr *iph;
137 struct sk_buff *head = NULL;
142 qp = container_of(frag, struct ipq, q);
143 net = qp->q.fqdir->net;
147 /* Paired with WRITE_ONCE() in fqdir_pre_exit(). */
148 if (READ_ONCE(qp->q.fqdir->dead))
151 spin_lock(&qp->q.lock);
153 if (qp->q.flags & INET_FRAG_COMPLETE)
157 __IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
158 __IP_INC_STATS(net, IPSTATS_MIB_REASMTIMEOUT);
160 if (!(qp->q.flags & INET_FRAG_FIRST_IN))
163 /* sk_buff::dev and sk_buff::rbnode are unionized. So we
164 * pull the head out of the tree in order to be able to
165 * deal with head->dev.
167 head = inet_frag_pull_head(&qp->q);
170 head->dev = dev_get_by_index_rcu(net, qp->iif);
175 /* skb has no dst, perform route lookup again */
177 err = ip_route_input_noref(head, iph->daddr, iph->saddr,
178 iph->tos, head->dev);
182 /* Only an end host needs to send an ICMP
183 * "Fragment Reassembly Timeout" message, per RFC792.
185 if (frag_expire_skip_icmp(qp->q.key.v4.user) &&
186 (skb_rtable(head)->rt_type != RTN_LOCAL))
189 spin_unlock(&qp->q.lock);
190 icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
194 spin_unlock(&qp->q.lock);
201 /* Find the correct entry in the "incomplete datagrams" queue for
202 * this IP datagram, and create new one, if nothing is found.
204 static struct ipq *ip_find(struct net *net, struct iphdr *iph,
207 struct frag_v4_compare_key key = {
213 .protocol = iph->protocol,
215 struct inet_frag_queue *q;
217 q = inet_frag_find(net->ipv4.fqdir, &key);
221 return container_of(q, struct ipq, q);
224 /* Is the fragment too far ahead to be part of ipq? */
225 static int ip_frag_too_far(struct ipq *qp)
227 struct inet_peer *peer = qp->peer;
228 unsigned int max = qp->q.fqdir->max_dist;
229 unsigned int start, end;
237 end = atomic_inc_return(&peer->rid);
240 rc = qp->q.fragments_tail && (end - start) > max;
243 __IP_INC_STATS(qp->q.fqdir->net, IPSTATS_MIB_REASMFAILS);
248 static int ip_frag_reinit(struct ipq *qp)
250 unsigned int sum_truesize = 0;
252 if (!mod_timer(&qp->q.timer, jiffies + qp->q.fqdir->timeout)) {
253 refcount_inc(&qp->q.refcnt);
257 sum_truesize = inet_frag_rbtree_purge(&qp->q.rb_fragments);
258 sub_frag_mem_limit(qp->q.fqdir, sum_truesize);
263 qp->q.rb_fragments = RB_ROOT;
264 qp->q.fragments_tail = NULL;
265 qp->q.last_run_head = NULL;
272 /* Add new segment to existing queue. */
273 static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
275 struct net *net = qp->q.fqdir->net;
276 int ihl, end, flags, offset;
277 struct sk_buff *prev_tail;
278 struct net_device *dev;
279 unsigned int fragsize;
283 if (qp->q.flags & INET_FRAG_COMPLETE)
286 if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) &&
287 unlikely(ip_frag_too_far(qp)) &&
288 unlikely(err = ip_frag_reinit(qp))) {
293 ecn = ip4_frag_ecn(ip_hdr(skb)->tos);
294 offset = ntohs(ip_hdr(skb)->frag_off);
295 flags = offset & ~IP_OFFSET;
297 offset <<= 3; /* offset is in 8-byte chunks */
298 ihl = ip_hdrlen(skb);
300 /* Determine the position of this fragment. */
301 end = offset + skb->len - skb_network_offset(skb) - ihl;
304 /* Is this the final fragment? */
305 if ((flags & IP_MF) == 0) {
306 /* If we already have some bits beyond end
307 * or have different end, the segment is corrupted.
309 if (end < qp->q.len ||
310 ((qp->q.flags & INET_FRAG_LAST_IN) && end != qp->q.len))
312 qp->q.flags |= INET_FRAG_LAST_IN;
317 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
318 skb->ip_summed = CHECKSUM_NONE;
320 if (end > qp->q.len) {
321 /* Some bits beyond end -> corruption. */
322 if (qp->q.flags & INET_FRAG_LAST_IN)
331 if (!pskb_pull(skb, skb_network_offset(skb) + ihl))
334 err = pskb_trim_rcsum(skb, end - offset);
338 /* Note : skb->rbnode and skb->dev share the same location. */
340 /* Makes sure compiler wont do silly aliasing games */
343 prev_tail = qp->q.fragments_tail;
344 err = inet_frag_queue_insert(&qp->q, skb, offset, end);
349 qp->iif = dev->ifindex;
351 qp->q.stamp = skb->tstamp;
352 qp->q.meat += skb->len;
354 add_frag_mem_limit(qp->q.fqdir, skb->truesize);
356 qp->q.flags |= INET_FRAG_FIRST_IN;
358 fragsize = skb->len + ihl;
360 if (fragsize > qp->q.max_size)
361 qp->q.max_size = fragsize;
363 if (ip_hdr(skb)->frag_off & htons(IP_DF) &&
364 fragsize > qp->max_df_size)
365 qp->max_df_size = fragsize;
367 if (qp->q.flags == (INET_FRAG_FIRST_IN | INET_FRAG_LAST_IN) &&
368 qp->q.meat == qp->q.len) {
369 unsigned long orefdst = skb->_skb_refdst;
371 skb->_skb_refdst = 0UL;
372 err = ip_frag_reasm(qp, skb, prev_tail, dev);
373 skb->_skb_refdst = orefdst;
375 inet_frag_kill(&qp->q);
383 if (err == IPFRAG_DUP) {
388 __IP_INC_STATS(net, IPSTATS_MIB_REASM_OVERLAPS);
390 inet_frag_kill(&qp->q);
391 __IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
397 static bool ip_frag_coalesce_ok(const struct ipq *qp)
399 return qp->q.key.v4.user == IP_DEFRAG_LOCAL_DELIVER;
402 /* Build a new IP datagram from all its fragments. */
403 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
404 struct sk_buff *prev_tail, struct net_device *dev)
406 struct net *net = qp->q.fqdir->net;
414 ecn = ip_frag_ecn_table[qp->ecn];
415 if (unlikely(ecn == 0xff)) {
420 /* Make the one we just received the head. */
421 reasm_data = inet_frag_reasm_prepare(&qp->q, skb, prev_tail);
425 len = ip_hdrlen(skb) + qp->q.len;
430 inet_frag_reasm_finish(&qp->q, skb, reasm_data,
431 ip_frag_coalesce_ok(qp));
434 IPCB(skb)->frag_max_size = max(qp->max_df_size, qp->q.max_size);
437 iph->tot_len = htons(len);
440 /* When we set IP_DF on a refragmented skb we must also force a
441 * call to ip_fragment to avoid forwarding a DF-skb of size s while
442 * original sender only sent fragments of size f (where f < s).
444 * We only set DF/IPSKB_FRAG_PMTU if such DF fragment was the largest
445 * frag seen to avoid sending tiny DF-fragments in case skb was built
446 * from one very small df-fragment and one large non-df frag.
448 if (qp->max_df_size == qp->q.max_size) {
449 IPCB(skb)->flags |= IPSKB_FRAG_PMTU;
450 iph->frag_off = htons(IP_DF);
457 __IP_INC_STATS(net, IPSTATS_MIB_REASMOKS);
458 qp->q.rb_fragments = RB_ROOT;
459 qp->q.fragments_tail = NULL;
460 qp->q.last_run_head = NULL;
464 net_dbg_ratelimited("queue_glue: no memory for gluing queue %p\n", qp);
468 net_info_ratelimited("Oversized IP packet from %pI4\n", &qp->q.key.v4.saddr);
470 __IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
474 /* Process an incoming IP datagram fragment. */
475 int ip_defrag(struct net *net, struct sk_buff *skb, u32 user)
477 struct net_device *dev = skb->dev ? : skb_dst(skb)->dev;
478 int vif = l3mdev_master_ifindex_rcu(dev);
481 __IP_INC_STATS(net, IPSTATS_MIB_REASMREQDS);
484 /* Lookup (or create) queue header */
485 qp = ip_find(net, ip_hdr(skb), user, vif);
489 spin_lock(&qp->q.lock);
491 ret = ip_frag_queue(qp, skb);
493 spin_unlock(&qp->q.lock);
498 __IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
502 EXPORT_SYMBOL(ip_defrag);
504 struct sk_buff *ip_check_defrag(struct net *net, struct sk_buff *skb, u32 user)
510 if (skb->protocol != htons(ETH_P_IP))
513 netoff = skb_network_offset(skb);
515 if (skb_copy_bits(skb, netoff, &iph, sizeof(iph)) < 0)
518 if (iph.ihl < 5 || iph.version != 4)
521 len = ntohs(iph.tot_len);
522 if (skb->len < netoff + len || len < (iph.ihl * 4))
525 if (ip_is_fragment(&iph)) {
526 skb = skb_share_check(skb, GFP_ATOMIC);
528 if (!pskb_may_pull(skb, netoff + iph.ihl * 4)) {
532 if (pskb_trim_rcsum(skb, netoff + len)) {
536 memset(IPCB(skb), 0, sizeof(struct inet_skb_parm));
537 if (ip_defrag(net, skb, user))
544 EXPORT_SYMBOL(ip_check_defrag);
549 static struct ctl_table ip4_frags_ns_ctl_table[] = {
551 .procname = "ipfrag_high_thresh",
552 .maxlen = sizeof(unsigned long),
554 .proc_handler = proc_doulongvec_minmax,
557 .procname = "ipfrag_low_thresh",
558 .maxlen = sizeof(unsigned long),
560 .proc_handler = proc_doulongvec_minmax,
563 .procname = "ipfrag_time",
564 .maxlen = sizeof(int),
566 .proc_handler = proc_dointvec_jiffies,
569 .procname = "ipfrag_max_dist",
570 .maxlen = sizeof(int),
572 .proc_handler = proc_dointvec_minmax,
578 /* secret interval has been deprecated */
579 static int ip4_frags_secret_interval_unused;
580 static struct ctl_table ip4_frags_ctl_table[] = {
582 .procname = "ipfrag_secret_interval",
583 .data = &ip4_frags_secret_interval_unused,
584 .maxlen = sizeof(int),
586 .proc_handler = proc_dointvec_jiffies,
591 static int __net_init ip4_frags_ns_ctl_register(struct net *net)
593 struct ctl_table *table;
594 struct ctl_table_header *hdr;
596 table = ip4_frags_ns_ctl_table;
597 if (!net_eq(net, &init_net)) {
598 table = kmemdup(table, sizeof(ip4_frags_ns_ctl_table), GFP_KERNEL);
603 table[0].data = &net->ipv4.fqdir->high_thresh;
604 table[0].extra1 = &net->ipv4.fqdir->low_thresh;
605 table[1].data = &net->ipv4.fqdir->low_thresh;
606 table[1].extra2 = &net->ipv4.fqdir->high_thresh;
607 table[2].data = &net->ipv4.fqdir->timeout;
608 table[3].data = &net->ipv4.fqdir->max_dist;
610 hdr = register_net_sysctl(net, "net/ipv4", table);
614 net->ipv4.frags_hdr = hdr;
618 if (!net_eq(net, &init_net))
624 static void __net_exit ip4_frags_ns_ctl_unregister(struct net *net)
626 struct ctl_table *table;
628 table = net->ipv4.frags_hdr->ctl_table_arg;
629 unregister_net_sysctl_table(net->ipv4.frags_hdr);
633 static void __init ip4_frags_ctl_register(void)
635 register_net_sysctl(&init_net, "net/ipv4", ip4_frags_ctl_table);
638 static int ip4_frags_ns_ctl_register(struct net *net)
643 static void ip4_frags_ns_ctl_unregister(struct net *net)
647 static void __init ip4_frags_ctl_register(void)
652 static int __net_init ipv4_frags_init_net(struct net *net)
656 res = fqdir_init(&net->ipv4.fqdir, &ip4_frags, net);
659 /* Fragment cache limits.
661 * The fragment memory accounting code, (tries to) account for
662 * the real memory usage, by measuring both the size of frag
663 * queue struct (inet_frag_queue (ipv4:ipq/ipv6:frag_queue))
664 * and the SKB's truesize.
666 * A 64K fragment consumes 129736 bytes (44*2944)+200
667 * (1500 truesize == 2944, sizeof(struct ipq) == 200)
669 * We will commit 4MB at one time. Should we cross that limit
670 * we will prune down to 3MB, making room for approx 8 big 64K
673 net->ipv4.fqdir->high_thresh = 4 * 1024 * 1024;
674 net->ipv4.fqdir->low_thresh = 3 * 1024 * 1024;
676 * Important NOTE! Fragment queue must be destroyed before MSL expires.
677 * RFC791 is wrong proposing to prolongate timer each fragment arrival
680 net->ipv4.fqdir->timeout = IP_FRAG_TIME;
682 net->ipv4.fqdir->max_dist = 64;
684 res = ip4_frags_ns_ctl_register(net);
686 fqdir_exit(net->ipv4.fqdir);
690 static void __net_exit ipv4_frags_pre_exit_net(struct net *net)
692 fqdir_pre_exit(net->ipv4.fqdir);
695 static void __net_exit ipv4_frags_exit_net(struct net *net)
697 ip4_frags_ns_ctl_unregister(net);
698 fqdir_exit(net->ipv4.fqdir);
701 static struct pernet_operations ip4_frags_ops = {
702 .init = ipv4_frags_init_net,
703 .pre_exit = ipv4_frags_pre_exit_net,
704 .exit = ipv4_frags_exit_net,
708 static u32 ip4_key_hashfn(const void *data, u32 len, u32 seed)
711 sizeof(struct frag_v4_compare_key) / sizeof(u32), seed);
714 static u32 ip4_obj_hashfn(const void *data, u32 len, u32 seed)
716 const struct inet_frag_queue *fq = data;
718 return jhash2((const u32 *)&fq->key.v4,
719 sizeof(struct frag_v4_compare_key) / sizeof(u32), seed);
722 static int ip4_obj_cmpfn(struct rhashtable_compare_arg *arg, const void *ptr)
724 const struct frag_v4_compare_key *key = arg->key;
725 const struct inet_frag_queue *fq = ptr;
727 return !!memcmp(&fq->key, key, sizeof(*key));
730 static const struct rhashtable_params ip4_rhash_params = {
731 .head_offset = offsetof(struct inet_frag_queue, node),
732 .key_offset = offsetof(struct inet_frag_queue, key),
733 .key_len = sizeof(struct frag_v4_compare_key),
734 .hashfn = ip4_key_hashfn,
735 .obj_hashfn = ip4_obj_hashfn,
736 .obj_cmpfn = ip4_obj_cmpfn,
737 .automatic_shrinking = true,
740 void __init ipfrag_init(void)
742 ip4_frags.constructor = ip4_frag_init;
743 ip4_frags.destructor = ip4_frag_free;
744 ip4_frags.qsize = sizeof(struct ipq);
745 ip4_frags.frag_expire = ip_expire;
746 ip4_frags.frags_cache_name = ip_frag_cache_name;
747 ip4_frags.rhash_params = ip4_rhash_params;
748 if (inet_frags_init(&ip4_frags))
749 panic("IP: failed to allocate ip4_frags cache\n");
750 ip4_frags_ctl_register();
751 register_pernet_subsys(&ip4_frags_ops);