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)
156 qp->q.flags |= INET_FRAG_DROP;
158 __IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
159 __IP_INC_STATS(net, IPSTATS_MIB_REASMTIMEOUT);
161 if (!(qp->q.flags & INET_FRAG_FIRST_IN))
164 /* sk_buff::dev and sk_buff::rbnode are unionized. So we
165 * pull the head out of the tree in order to be able to
166 * deal with head->dev.
168 head = inet_frag_pull_head(&qp->q);
171 head->dev = dev_get_by_index_rcu(net, qp->iif);
176 /* skb has no dst, perform route lookup again */
178 err = ip_route_input_noref(head, iph->daddr, iph->saddr,
179 iph->tos, head->dev);
183 /* Only an end host needs to send an ICMP
184 * "Fragment Reassembly Timeout" message, per RFC792.
186 if (frag_expire_skip_icmp(qp->q.key.v4.user) &&
187 (skb_rtable(head)->rt_type != RTN_LOCAL))
190 spin_unlock(&qp->q.lock);
191 icmp_send(head, ICMP_TIME_EXCEEDED, ICMP_EXC_FRAGTIME, 0);
195 spin_unlock(&qp->q.lock);
198 kfree_skb_reason(head, SKB_DROP_REASON_FRAG_REASM_TIMEOUT);
202 /* Find the correct entry in the "incomplete datagrams" queue for
203 * this IP datagram, and create new one, if nothing is found.
205 static struct ipq *ip_find(struct net *net, struct iphdr *iph,
208 struct frag_v4_compare_key key = {
214 .protocol = iph->protocol,
216 struct inet_frag_queue *q;
218 q = inet_frag_find(net->ipv4.fqdir, &key);
222 return container_of(q, struct ipq, q);
225 /* Is the fragment too far ahead to be part of ipq? */
226 static int ip_frag_too_far(struct ipq *qp)
228 struct inet_peer *peer = qp->peer;
229 unsigned int max = qp->q.fqdir->max_dist;
230 unsigned int start, end;
238 end = atomic_inc_return(&peer->rid);
241 rc = qp->q.fragments_tail && (end - start) > max;
244 __IP_INC_STATS(qp->q.fqdir->net, IPSTATS_MIB_REASMFAILS);
249 static int ip_frag_reinit(struct ipq *qp)
251 unsigned int sum_truesize = 0;
253 if (!mod_timer(&qp->q.timer, jiffies + qp->q.fqdir->timeout)) {
254 refcount_inc(&qp->q.refcnt);
258 sum_truesize = inet_frag_rbtree_purge(&qp->q.rb_fragments,
259 SKB_DROP_REASON_FRAG_TOO_FAR);
260 sub_frag_mem_limit(qp->q.fqdir, sum_truesize);
265 qp->q.rb_fragments = RB_ROOT;
266 qp->q.fragments_tail = NULL;
267 qp->q.last_run_head = NULL;
274 /* Add new segment to existing queue. */
275 static int ip_frag_queue(struct ipq *qp, struct sk_buff *skb)
277 struct net *net = qp->q.fqdir->net;
278 int ihl, end, flags, offset;
279 struct sk_buff *prev_tail;
280 struct net_device *dev;
281 unsigned int fragsize;
286 /* If reassembly is already done, @skb must be a duplicate frag. */
287 if (qp->q.flags & INET_FRAG_COMPLETE) {
288 SKB_DR_SET(reason, DUP_FRAG);
292 if (!(IPCB(skb)->flags & IPSKB_FRAG_COMPLETE) &&
293 unlikely(ip_frag_too_far(qp)) &&
294 unlikely(err = ip_frag_reinit(qp))) {
299 ecn = ip4_frag_ecn(ip_hdr(skb)->tos);
300 offset = ntohs(ip_hdr(skb)->frag_off);
301 flags = offset & ~IP_OFFSET;
303 offset <<= 3; /* offset is in 8-byte chunks */
304 ihl = ip_hdrlen(skb);
306 /* Determine the position of this fragment. */
307 end = offset + skb->len - skb_network_offset(skb) - ihl;
310 /* Is this the final fragment? */
311 if ((flags & IP_MF) == 0) {
312 /* If we already have some bits beyond end
313 * or have different end, the segment is corrupted.
315 if (end < qp->q.len ||
316 ((qp->q.flags & INET_FRAG_LAST_IN) && end != qp->q.len))
318 qp->q.flags |= INET_FRAG_LAST_IN;
323 if (skb->ip_summed != CHECKSUM_UNNECESSARY)
324 skb->ip_summed = CHECKSUM_NONE;
326 if (end > qp->q.len) {
327 /* Some bits beyond end -> corruption. */
328 if (qp->q.flags & INET_FRAG_LAST_IN)
337 if (!pskb_pull(skb, skb_network_offset(skb) + ihl))
340 err = pskb_trim_rcsum(skb, end - offset);
344 /* Note : skb->rbnode and skb->dev share the same location. */
346 /* Makes sure compiler wont do silly aliasing games */
349 prev_tail = qp->q.fragments_tail;
350 err = inet_frag_queue_insert(&qp->q, skb, offset, end);
355 qp->iif = dev->ifindex;
357 qp->q.stamp = skb->tstamp;
358 qp->q.mono_delivery_time = skb->mono_delivery_time;
359 qp->q.meat += skb->len;
361 add_frag_mem_limit(qp->q.fqdir, skb->truesize);
363 qp->q.flags |= INET_FRAG_FIRST_IN;
365 fragsize = skb->len + ihl;
367 if (fragsize > qp->q.max_size)
368 qp->q.max_size = fragsize;
370 if (ip_hdr(skb)->frag_off & htons(IP_DF) &&
371 fragsize > qp->max_df_size)
372 qp->max_df_size = fragsize;
374 if (qp->q.flags == (INET_FRAG_FIRST_IN | INET_FRAG_LAST_IN) &&
375 qp->q.meat == qp->q.len) {
376 unsigned long orefdst = skb->_skb_refdst;
378 skb->_skb_refdst = 0UL;
379 err = ip_frag_reasm(qp, skb, prev_tail, dev);
380 skb->_skb_refdst = orefdst;
382 inet_frag_kill(&qp->q);
390 if (err == IPFRAG_DUP) {
391 SKB_DR_SET(reason, DUP_FRAG);
396 __IP_INC_STATS(net, IPSTATS_MIB_REASM_OVERLAPS);
398 inet_frag_kill(&qp->q);
399 __IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
401 kfree_skb_reason(skb, reason);
405 static bool ip_frag_coalesce_ok(const struct ipq *qp)
407 return qp->q.key.v4.user == IP_DEFRAG_LOCAL_DELIVER;
410 /* Build a new IP datagram from all its fragments. */
411 static int ip_frag_reasm(struct ipq *qp, struct sk_buff *skb,
412 struct sk_buff *prev_tail, struct net_device *dev)
414 struct net *net = qp->q.fqdir->net;
422 ecn = ip_frag_ecn_table[qp->ecn];
423 if (unlikely(ecn == 0xff)) {
428 /* Make the one we just received the head. */
429 reasm_data = inet_frag_reasm_prepare(&qp->q, skb, prev_tail);
433 len = ip_hdrlen(skb) + qp->q.len;
438 inet_frag_reasm_finish(&qp->q, skb, reasm_data,
439 ip_frag_coalesce_ok(qp));
442 IPCB(skb)->frag_max_size = max(qp->max_df_size, qp->q.max_size);
445 iph->tot_len = htons(len);
448 /* When we set IP_DF on a refragmented skb we must also force a
449 * call to ip_fragment to avoid forwarding a DF-skb of size s while
450 * original sender only sent fragments of size f (where f < s).
452 * We only set DF/IPSKB_FRAG_PMTU if such DF fragment was the largest
453 * frag seen to avoid sending tiny DF-fragments in case skb was built
454 * from one very small df-fragment and one large non-df frag.
456 if (qp->max_df_size == qp->q.max_size) {
457 IPCB(skb)->flags |= IPSKB_FRAG_PMTU;
458 iph->frag_off = htons(IP_DF);
465 __IP_INC_STATS(net, IPSTATS_MIB_REASMOKS);
466 qp->q.rb_fragments = RB_ROOT;
467 qp->q.fragments_tail = NULL;
468 qp->q.last_run_head = NULL;
472 net_dbg_ratelimited("queue_glue: no memory for gluing queue %p\n", qp);
476 net_info_ratelimited("Oversized IP packet from %pI4\n", &qp->q.key.v4.saddr);
478 __IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
482 /* Process an incoming IP datagram fragment. */
483 int ip_defrag(struct net *net, struct sk_buff *skb, u32 user)
485 struct net_device *dev = skb->dev ? : skb_dst(skb)->dev;
486 int vif = l3mdev_master_ifindex_rcu(dev);
489 __IP_INC_STATS(net, IPSTATS_MIB_REASMREQDS);
492 /* Lookup (or create) queue header */
493 qp = ip_find(net, ip_hdr(skb), user, vif);
497 spin_lock(&qp->q.lock);
499 ret = ip_frag_queue(qp, skb);
501 spin_unlock(&qp->q.lock);
506 __IP_INC_STATS(net, IPSTATS_MIB_REASMFAILS);
510 EXPORT_SYMBOL(ip_defrag);
512 struct sk_buff *ip_check_defrag(struct net *net, struct sk_buff *skb, u32 user)
518 if (skb->protocol != htons(ETH_P_IP))
521 netoff = skb_network_offset(skb);
523 if (skb_copy_bits(skb, netoff, &iph, sizeof(iph)) < 0)
526 if (iph.ihl < 5 || iph.version != 4)
529 len = ntohs(iph.tot_len);
530 if (skb->len < netoff + len || len < (iph.ihl * 4))
533 if (ip_is_fragment(&iph)) {
534 skb = skb_share_check(skb, GFP_ATOMIC);
536 if (!pskb_may_pull(skb, netoff + iph.ihl * 4)) {
540 if (pskb_trim_rcsum(skb, netoff + len)) {
544 memset(IPCB(skb), 0, sizeof(struct inet_skb_parm));
545 if (ip_defrag(net, skb, user))
552 EXPORT_SYMBOL(ip_check_defrag);
556 static unsigned long ipfrag_low_thresh_unused;
557 static struct ctl_table ip4_frags_ns_ctl_table[] = {
559 .procname = "ipfrag_high_thresh",
560 .maxlen = sizeof(unsigned long),
562 .proc_handler = proc_doulongvec_minmax,
565 .procname = "ipfrag_low_thresh",
566 .maxlen = sizeof(unsigned long),
568 .proc_handler = proc_doulongvec_minmax,
571 .procname = "ipfrag_time",
572 .maxlen = sizeof(int),
574 .proc_handler = proc_dointvec_jiffies,
577 .procname = "ipfrag_max_dist",
578 .maxlen = sizeof(int),
580 .proc_handler = proc_dointvec_minmax,
586 /* secret interval has been deprecated */
587 static int ip4_frags_secret_interval_unused;
588 static struct ctl_table ip4_frags_ctl_table[] = {
590 .procname = "ipfrag_secret_interval",
591 .data = &ip4_frags_secret_interval_unused,
592 .maxlen = sizeof(int),
594 .proc_handler = proc_dointvec_jiffies,
599 static int __net_init ip4_frags_ns_ctl_register(struct net *net)
601 struct ctl_table *table;
602 struct ctl_table_header *hdr;
604 table = ip4_frags_ns_ctl_table;
605 if (!net_eq(net, &init_net)) {
606 table = kmemdup(table, sizeof(ip4_frags_ns_ctl_table), GFP_KERNEL);
611 table[0].data = &net->ipv4.fqdir->high_thresh;
612 table[0].extra1 = &ipfrag_low_thresh_unused;
613 table[1].data = &ipfrag_low_thresh_unused;
614 table[1].extra2 = &net->ipv4.fqdir->high_thresh;
615 table[2].data = &net->ipv4.fqdir->timeout;
616 table[3].data = &net->ipv4.fqdir->max_dist;
618 hdr = register_net_sysctl(net, "net/ipv4", table);
622 net->ipv4.frags_hdr = hdr;
626 if (!net_eq(net, &init_net))
632 static void __net_exit ip4_frags_ns_ctl_unregister(struct net *net)
634 struct ctl_table *table;
636 table = net->ipv4.frags_hdr->ctl_table_arg;
637 unregister_net_sysctl_table(net->ipv4.frags_hdr);
641 static void __init ip4_frags_ctl_register(void)
643 register_net_sysctl(&init_net, "net/ipv4", ip4_frags_ctl_table);
646 static int ip4_frags_ns_ctl_register(struct net *net)
651 static void ip4_frags_ns_ctl_unregister(struct net *net)
655 static void __init ip4_frags_ctl_register(void)
660 static int __net_init ipv4_frags_init_net(struct net *net)
664 res = fqdir_init(&net->ipv4.fqdir, &ip4_frags, net);
667 /* Fragment cache limits.
669 * The fragment memory accounting code, (tries to) account for
670 * the real memory usage, by measuring both the size of frag
671 * queue struct (inet_frag_queue (ipv4:ipq/ipv6:frag_queue))
672 * and the SKB's truesize.
674 * A 64K fragment consumes 129736 bytes (44*2944)+200
675 * (1500 truesize == 2944, sizeof(struct ipq) == 200)
677 * We will commit 4MB at one time. Should we cross that limit.
679 net->ipv4.fqdir->high_thresh = 4 * 1024 * 1024;
681 * Important NOTE! Fragment queue must be destroyed before MSL expires.
682 * RFC791 is wrong proposing to prolongate timer each fragment arrival
685 net->ipv4.fqdir->timeout = IP_FRAG_TIME;
687 net->ipv4.fqdir->max_dist = 64;
689 res = ip4_frags_ns_ctl_register(net);
691 fqdir_exit(net->ipv4.fqdir);
695 static void __net_exit ipv4_frags_pre_exit_net(struct net *net)
697 fqdir_pre_exit(net->ipv4.fqdir);
700 static void __net_exit ipv4_frags_exit_net(struct net *net)
702 ip4_frags_ns_ctl_unregister(net);
703 fqdir_exit(net->ipv4.fqdir);
706 static struct pernet_operations ip4_frags_ops = {
707 .init = ipv4_frags_init_net,
708 .pre_exit = ipv4_frags_pre_exit_net,
709 .exit = ipv4_frags_exit_net,
713 static u32 ip4_key_hashfn(const void *data, u32 len, u32 seed)
716 sizeof(struct frag_v4_compare_key) / sizeof(u32), seed);
719 static u32 ip4_obj_hashfn(const void *data, u32 len, u32 seed)
721 const struct inet_frag_queue *fq = data;
723 return jhash2((const u32 *)&fq->key.v4,
724 sizeof(struct frag_v4_compare_key) / sizeof(u32), seed);
727 static int ip4_obj_cmpfn(struct rhashtable_compare_arg *arg, const void *ptr)
729 const struct frag_v4_compare_key *key = arg->key;
730 const struct inet_frag_queue *fq = ptr;
732 return !!memcmp(&fq->key, key, sizeof(*key));
735 static const struct rhashtable_params ip4_rhash_params = {
736 .head_offset = offsetof(struct inet_frag_queue, node),
737 .key_offset = offsetof(struct inet_frag_queue, key),
738 .key_len = sizeof(struct frag_v4_compare_key),
739 .hashfn = ip4_key_hashfn,
740 .obj_hashfn = ip4_obj_hashfn,
741 .obj_cmpfn = ip4_obj_cmpfn,
742 .automatic_shrinking = true,
745 void __init ipfrag_init(void)
747 ip4_frags.constructor = ip4_frag_init;
748 ip4_frags.destructor = ip4_frag_free;
749 ip4_frags.qsize = sizeof(struct ipq);
750 ip4_frags.frag_expire = ip_expire;
751 ip4_frags.frags_cache_name = ip_frag_cache_name;
752 ip4_frags.rhash_params = ip4_rhash_params;
753 if (inet_frags_init(&ip4_frags))
754 panic("IP: failed to allocate ip4_frags cache\n");
755 ip4_frags_ctl_register();
756 register_pernet_subsys(&ip4_frags_ops);