1 // SPDX-License-Identifier: GPL-2.0-only
2 /* Connection state tracking for netfilter. This is separated from,
3 but required by, the NAT layer; it can also be used by an iptables
6 /* (C) 1999-2001 Paul `Rusty' Russell
7 * (C) 2002-2006 Netfilter Core Team <coreteam@netfilter.org>
8 * (C) 2003,2004 USAGI/WIDE Project <http://www.linux-ipv6.org>
9 * (C) 2005-2012 Patrick McHardy <kaber@trash.net>
12 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 #include <linux/types.h>
15 #include <linux/netfilter.h>
16 #include <linux/module.h>
17 #include <linux/sched.h>
18 #include <linux/skbuff.h>
19 #include <linux/proc_fs.h>
20 #include <linux/vmalloc.h>
21 #include <linux/stddef.h>
22 #include <linux/slab.h>
23 #include <linux/random.h>
24 #include <linux/siphash.h>
25 #include <linux/err.h>
26 #include <linux/percpu.h>
27 #include <linux/moduleparam.h>
28 #include <linux/notifier.h>
29 #include <linux/kernel.h>
30 #include <linux/netdevice.h>
31 #include <linux/socket.h>
33 #include <linux/nsproxy.h>
34 #include <linux/rculist_nulls.h>
36 #include <net/netfilter/nf_conntrack.h>
37 #include <net/netfilter/nf_conntrack_l4proto.h>
38 #include <net/netfilter/nf_conntrack_expect.h>
39 #include <net/netfilter/nf_conntrack_helper.h>
40 #include <net/netfilter/nf_conntrack_seqadj.h>
41 #include <net/netfilter/nf_conntrack_core.h>
42 #include <net/netfilter/nf_conntrack_extend.h>
43 #include <net/netfilter/nf_conntrack_acct.h>
44 #include <net/netfilter/nf_conntrack_ecache.h>
45 #include <net/netfilter/nf_conntrack_zones.h>
46 #include <net/netfilter/nf_conntrack_timestamp.h>
47 #include <net/netfilter/nf_conntrack_timeout.h>
48 #include <net/netfilter/nf_conntrack_labels.h>
49 #include <net/netfilter/nf_conntrack_synproxy.h>
50 #include <net/netfilter/nf_nat.h>
51 #include <net/netfilter/nf_nat_helper.h>
52 #include <net/netns/hash.h>
55 #include "nf_internals.h"
57 __cacheline_aligned_in_smp spinlock_t nf_conntrack_locks[CONNTRACK_LOCKS];
58 EXPORT_SYMBOL_GPL(nf_conntrack_locks);
60 __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
61 EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
63 struct hlist_nulls_head *nf_conntrack_hash __read_mostly;
64 EXPORT_SYMBOL_GPL(nf_conntrack_hash);
66 struct conntrack_gc_work {
67 struct delayed_work dwork;
73 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
74 static DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
75 static __read_mostly bool nf_conntrack_locks_all;
77 /* serialize hash resizes and nf_ct_iterate_cleanup */
78 static DEFINE_MUTEX(nf_conntrack_mutex);
80 #define GC_SCAN_INTERVAL (120u * HZ)
81 #define GC_SCAN_MAX_DURATION msecs_to_jiffies(10)
83 #define MIN_CHAINLEN 8u
84 #define MAX_CHAINLEN (32u - MIN_CHAINLEN)
86 static struct conntrack_gc_work conntrack_gc_work;
88 void nf_conntrack_lock(spinlock_t *lock) __acquires(lock)
90 /* 1) Acquire the lock */
93 /* 2) read nf_conntrack_locks_all, with ACQUIRE semantics
94 * It pairs with the smp_store_release() in nf_conntrack_all_unlock()
96 if (likely(smp_load_acquire(&nf_conntrack_locks_all) == false))
99 /* fast path failed, unlock */
102 /* Slow path 1) get global lock */
103 spin_lock(&nf_conntrack_locks_all_lock);
105 /* Slow path 2) get the lock we want */
108 /* Slow path 3) release the global lock */
109 spin_unlock(&nf_conntrack_locks_all_lock);
111 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
113 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
115 h1 %= CONNTRACK_LOCKS;
116 h2 %= CONNTRACK_LOCKS;
117 spin_unlock(&nf_conntrack_locks[h1]);
119 spin_unlock(&nf_conntrack_locks[h2]);
122 /* return true if we need to recompute hashes (in case hash table was resized) */
123 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
124 unsigned int h2, unsigned int sequence)
126 h1 %= CONNTRACK_LOCKS;
127 h2 %= CONNTRACK_LOCKS;
129 nf_conntrack_lock(&nf_conntrack_locks[h1]);
131 spin_lock_nested(&nf_conntrack_locks[h2],
132 SINGLE_DEPTH_NESTING);
134 nf_conntrack_lock(&nf_conntrack_locks[h2]);
135 spin_lock_nested(&nf_conntrack_locks[h1],
136 SINGLE_DEPTH_NESTING);
138 if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
139 nf_conntrack_double_unlock(h1, h2);
145 static void nf_conntrack_all_lock(void)
146 __acquires(&nf_conntrack_locks_all_lock)
150 spin_lock(&nf_conntrack_locks_all_lock);
152 /* For nf_contrack_locks_all, only the latest time when another
153 * CPU will see an update is controlled, by the "release" of the
155 * The earliest time is not controlled, an thus KCSAN could detect
156 * a race when nf_conntract_lock() reads the variable.
157 * WRITE_ONCE() is used to ensure the compiler will not
158 * optimize the write.
160 WRITE_ONCE(nf_conntrack_locks_all, true);
162 for (i = 0; i < CONNTRACK_LOCKS; i++) {
163 spin_lock(&nf_conntrack_locks[i]);
165 /* This spin_unlock provides the "release" to ensure that
166 * nf_conntrack_locks_all==true is visible to everyone that
167 * acquired spin_lock(&nf_conntrack_locks[]).
169 spin_unlock(&nf_conntrack_locks[i]);
173 static void nf_conntrack_all_unlock(void)
174 __releases(&nf_conntrack_locks_all_lock)
176 /* All prior stores must be complete before we clear
177 * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
178 * might observe the false value but not the entire
180 * It pairs with the smp_load_acquire() in nf_conntrack_lock()
182 smp_store_release(&nf_conntrack_locks_all, false);
183 spin_unlock(&nf_conntrack_locks_all_lock);
186 unsigned int nf_conntrack_htable_size __read_mostly;
187 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
189 unsigned int nf_conntrack_max __read_mostly;
190 EXPORT_SYMBOL_GPL(nf_conntrack_max);
191 seqcount_spinlock_t nf_conntrack_generation __read_mostly;
192 static siphash_key_t nf_conntrack_hash_rnd __read_mostly;
194 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
196 const struct net *net)
199 struct nf_conntrack_man src;
200 union nf_inet_addr dst_addr;
205 } __aligned(SIPHASH_ALIGNMENT) combined;
207 get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
209 memset(&combined, 0, sizeof(combined));
211 /* The direction must be ignored, so handle usable members manually. */
212 combined.src = tuple->src;
213 combined.dst_addr = tuple->dst.u3;
214 combined.zone = zoneid;
215 combined.net_mix = net_hash_mix(net);
216 combined.dport = (__force __u16)tuple->dst.u.all;
217 combined.proto = tuple->dst.protonum;
219 return (u32)siphash(&combined, sizeof(combined), &nf_conntrack_hash_rnd);
222 static u32 scale_hash(u32 hash)
224 return reciprocal_scale(hash, nf_conntrack_htable_size);
227 static u32 __hash_conntrack(const struct net *net,
228 const struct nf_conntrack_tuple *tuple,
232 return reciprocal_scale(hash_conntrack_raw(tuple, zoneid, net), size);
235 static u32 hash_conntrack(const struct net *net,
236 const struct nf_conntrack_tuple *tuple,
239 return scale_hash(hash_conntrack_raw(tuple, zoneid, net));
242 static bool nf_ct_get_tuple_ports(const struct sk_buff *skb,
243 unsigned int dataoff,
244 struct nf_conntrack_tuple *tuple)
248 } _inet_hdr, *inet_hdr;
250 /* Actually only need first 4 bytes to get ports. */
251 inet_hdr = skb_header_pointer(skb, dataoff, sizeof(_inet_hdr), &_inet_hdr);
255 tuple->src.u.udp.port = inet_hdr->sport;
256 tuple->dst.u.udp.port = inet_hdr->dport;
261 nf_ct_get_tuple(const struct sk_buff *skb,
263 unsigned int dataoff,
267 struct nf_conntrack_tuple *tuple)
273 memset(tuple, 0, sizeof(*tuple));
275 tuple->src.l3num = l3num;
278 nhoff += offsetof(struct iphdr, saddr);
279 size = 2 * sizeof(__be32);
282 nhoff += offsetof(struct ipv6hdr, saddr);
283 size = sizeof(_addrs);
289 ap = skb_header_pointer(skb, nhoff, size, _addrs);
295 tuple->src.u3.ip = ap[0];
296 tuple->dst.u3.ip = ap[1];
299 memcpy(tuple->src.u3.ip6, ap, sizeof(tuple->src.u3.ip6));
300 memcpy(tuple->dst.u3.ip6, ap + 4, sizeof(tuple->dst.u3.ip6));
304 tuple->dst.protonum = protonum;
305 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
308 #if IS_ENABLED(CONFIG_IPV6)
310 return icmpv6_pkt_to_tuple(skb, dataoff, net, tuple);
313 return icmp_pkt_to_tuple(skb, dataoff, net, tuple);
314 #ifdef CONFIG_NF_CT_PROTO_GRE
316 return gre_pkt_to_tuple(skb, dataoff, net, tuple);
319 case IPPROTO_UDP: /* fallthrough */
320 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
321 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
322 case IPPROTO_UDPLITE:
323 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
325 #ifdef CONFIG_NF_CT_PROTO_SCTP
327 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
329 #ifdef CONFIG_NF_CT_PROTO_DCCP
331 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
340 static int ipv4_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
344 const struct iphdr *iph;
347 iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
351 /* Conntrack defragments packets, we might still see fragments
352 * inside ICMP packets though.
354 if (iph->frag_off & htons(IP_OFFSET))
357 dataoff = nhoff + (iph->ihl << 2);
358 *protonum = iph->protocol;
360 /* Check bogus IP headers */
361 if (dataoff > skb->len) {
362 pr_debug("bogus IPv4 packet: nhoff %u, ihl %u, skblen %u\n",
363 nhoff, iph->ihl << 2, skb->len);
369 #if IS_ENABLED(CONFIG_IPV6)
370 static int ipv6_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
374 unsigned int extoff = nhoff + sizeof(struct ipv6hdr);
378 if (skb_copy_bits(skb, nhoff + offsetof(struct ipv6hdr, nexthdr),
379 &nexthdr, sizeof(nexthdr)) != 0) {
380 pr_debug("can't get nexthdr\n");
383 protoff = ipv6_skip_exthdr(skb, extoff, &nexthdr, &frag_off);
385 * (protoff == skb->len) means the packet has not data, just
386 * IPv6 and possibly extensions headers, but it is tracked anyway
388 if (protoff < 0 || (frag_off & htons(~0x7)) != 0) {
389 pr_debug("can't find proto in pkt\n");
398 static int get_l4proto(const struct sk_buff *skb,
399 unsigned int nhoff, u8 pf, u8 *l4num)
403 return ipv4_get_l4proto(skb, nhoff, l4num);
404 #if IS_ENABLED(CONFIG_IPV6)
406 return ipv6_get_l4proto(skb, nhoff, l4num);
415 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
417 struct net *net, struct nf_conntrack_tuple *tuple)
422 protoff = get_l4proto(skb, nhoff, l3num, &protonum);
426 return nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple);
428 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
431 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
432 const struct nf_conntrack_tuple *orig)
434 memset(inverse, 0, sizeof(*inverse));
436 inverse->src.l3num = orig->src.l3num;
438 switch (orig->src.l3num) {
440 inverse->src.u3.ip = orig->dst.u3.ip;
441 inverse->dst.u3.ip = orig->src.u3.ip;
444 inverse->src.u3.in6 = orig->dst.u3.in6;
445 inverse->dst.u3.in6 = orig->src.u3.in6;
451 inverse->dst.dir = !orig->dst.dir;
453 inverse->dst.protonum = orig->dst.protonum;
455 switch (orig->dst.protonum) {
457 return nf_conntrack_invert_icmp_tuple(inverse, orig);
458 #if IS_ENABLED(CONFIG_IPV6)
460 return nf_conntrack_invert_icmpv6_tuple(inverse, orig);
464 inverse->src.u.all = orig->dst.u.all;
465 inverse->dst.u.all = orig->src.u.all;
468 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
470 /* Generate a almost-unique pseudo-id for a given conntrack.
472 * intentionally doesn't re-use any of the seeds used for hash
473 * table location, we assume id gets exposed to userspace.
475 * Following nf_conn items do not change throughout lifetime
479 * 2. nf_conn->master address (normally NULL)
480 * 3. the associated net namespace
481 * 4. the original direction tuple
483 u32 nf_ct_get_id(const struct nf_conn *ct)
485 static __read_mostly siphash_key_t ct_id_seed;
486 unsigned long a, b, c, d;
488 net_get_random_once(&ct_id_seed, sizeof(ct_id_seed));
490 a = (unsigned long)ct;
491 b = (unsigned long)ct->master;
492 c = (unsigned long)nf_ct_net(ct);
493 d = (unsigned long)siphash(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
494 sizeof(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple),
497 return siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &ct_id_seed);
499 return siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &ct_id_seed);
502 EXPORT_SYMBOL_GPL(nf_ct_get_id);
505 clean_from_lists(struct nf_conn *ct)
507 pr_debug("clean_from_lists(%p)\n", ct);
508 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
509 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
511 /* Destroy all pending expectations */
512 nf_ct_remove_expectations(ct);
515 /* must be called with local_bh_disable */
516 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
518 struct ct_pcpu *pcpu;
520 /* add this conntrack to the (per cpu) dying list */
521 ct->cpu = smp_processor_id();
522 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
524 spin_lock(&pcpu->lock);
525 hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
527 spin_unlock(&pcpu->lock);
530 /* must be called with local_bh_disable */
531 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
533 struct ct_pcpu *pcpu;
535 /* add this conntrack to the (per cpu) unconfirmed list */
536 ct->cpu = smp_processor_id();
537 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
539 spin_lock(&pcpu->lock);
540 hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
542 spin_unlock(&pcpu->lock);
545 /* must be called with local_bh_disable */
546 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
548 struct ct_pcpu *pcpu;
550 /* We overload first tuple to link into unconfirmed or dying list.*/
551 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
553 spin_lock(&pcpu->lock);
554 BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
555 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
556 spin_unlock(&pcpu->lock);
559 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
561 /* Released via destroy_conntrack() */
562 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
563 const struct nf_conntrack_zone *zone,
566 struct nf_conn *tmpl, *p;
568 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
569 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
574 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
576 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
577 tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
580 tmpl = kzalloc(sizeof(*tmpl), flags);
585 tmpl->status = IPS_TEMPLATE;
586 write_pnet(&tmpl->ct_net, net);
587 nf_ct_zone_add(tmpl, zone);
588 atomic_set(&tmpl->ct_general.use, 0);
592 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
594 void nf_ct_tmpl_free(struct nf_conn *tmpl)
596 nf_ct_ext_destroy(tmpl);
598 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
599 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
603 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
605 static void destroy_gre_conntrack(struct nf_conn *ct)
607 #ifdef CONFIG_NF_CT_PROTO_GRE
608 struct nf_conn *master = ct->master;
611 nf_ct_gre_keymap_destroy(master);
616 destroy_conntrack(struct nf_conntrack *nfct)
618 struct nf_conn *ct = (struct nf_conn *)nfct;
620 pr_debug("destroy_conntrack(%p)\n", ct);
621 WARN_ON(atomic_read(&nfct->use) != 0);
623 if (unlikely(nf_ct_is_template(ct))) {
628 if (unlikely(nf_ct_protonum(ct) == IPPROTO_GRE))
629 destroy_gre_conntrack(ct);
632 /* Expectations will have been removed in clean_from_lists,
633 * except TFTP can create an expectation on the first packet,
634 * before connection is in the list, so we need to clean here,
637 nf_ct_remove_expectations(ct);
639 nf_ct_del_from_dying_or_unconfirmed_list(ct);
644 nf_ct_put(ct->master);
646 pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
647 nf_conntrack_free(ct);
650 static void nf_ct_delete_from_lists(struct nf_conn *ct)
652 struct net *net = nf_ct_net(ct);
653 unsigned int hash, reply_hash;
654 unsigned int sequence;
656 nf_ct_helper_destroy(ct);
660 sequence = read_seqcount_begin(&nf_conntrack_generation);
661 hash = hash_conntrack(net,
662 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
663 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_ORIGINAL));
664 reply_hash = hash_conntrack(net,
665 &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
666 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
667 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
669 clean_from_lists(ct);
670 nf_conntrack_double_unlock(hash, reply_hash);
672 nf_ct_add_to_dying_list(ct);
677 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
679 struct nf_conn_tstamp *tstamp;
682 if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
685 tstamp = nf_conn_tstamp_find(ct);
687 s32 timeout = ct->timeout - nfct_time_stamp;
689 tstamp->stop = ktime_get_real_ns();
691 tstamp->stop -= jiffies_to_nsecs(-timeout);
694 if (nf_conntrack_event_report(IPCT_DESTROY, ct,
695 portid, report) < 0) {
696 /* destroy event was not delivered. nf_ct_put will
697 * be done by event cache worker on redelivery.
699 nf_ct_delete_from_lists(ct);
700 nf_conntrack_ecache_work(nf_ct_net(ct), NFCT_ECACHE_DESTROY_FAIL);
705 if (nf_conntrack_ecache_dwork_pending(net))
706 nf_conntrack_ecache_work(net, NFCT_ECACHE_DESTROY_SENT);
707 nf_ct_delete_from_lists(ct);
711 EXPORT_SYMBOL_GPL(nf_ct_delete);
714 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
715 const struct nf_conntrack_tuple *tuple,
716 const struct nf_conntrack_zone *zone,
717 const struct net *net)
719 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
721 /* A conntrack can be recreated with the equal tuple,
722 * so we need to check that the conntrack is confirmed
724 return nf_ct_tuple_equal(tuple, &h->tuple) &&
725 nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
726 nf_ct_is_confirmed(ct) &&
727 net_eq(net, nf_ct_net(ct));
731 nf_ct_match(const struct nf_conn *ct1, const struct nf_conn *ct2)
733 return nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
734 &ct2->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
735 nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_REPLY].tuple,
736 &ct2->tuplehash[IP_CT_DIR_REPLY].tuple) &&
737 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_ORIGINAL) &&
738 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_REPLY) &&
739 net_eq(nf_ct_net(ct1), nf_ct_net(ct2));
742 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
743 static void nf_ct_gc_expired(struct nf_conn *ct)
745 if (!atomic_inc_not_zero(&ct->ct_general.use))
748 if (nf_ct_should_gc(ct))
756 * - Caller must take a reference on returned object
757 * and recheck nf_ct_tuple_equal(tuple, &h->tuple)
759 static struct nf_conntrack_tuple_hash *
760 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
761 const struct nf_conntrack_tuple *tuple, u32 hash)
763 struct nf_conntrack_tuple_hash *h;
764 struct hlist_nulls_head *ct_hash;
765 struct hlist_nulls_node *n;
766 unsigned int bucket, hsize;
769 nf_conntrack_get_ht(&ct_hash, &hsize);
770 bucket = reciprocal_scale(hash, hsize);
772 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
775 ct = nf_ct_tuplehash_to_ctrack(h);
776 if (nf_ct_is_expired(ct)) {
777 nf_ct_gc_expired(ct);
781 if (nf_ct_key_equal(h, tuple, zone, net))
785 * if the nulls value we got at the end of this lookup is
786 * not the expected one, we must restart lookup.
787 * We probably met an item that was moved to another chain.
789 if (get_nulls_value(n) != bucket) {
790 NF_CT_STAT_INC_ATOMIC(net, search_restart);
797 /* Find a connection corresponding to a tuple. */
798 static struct nf_conntrack_tuple_hash *
799 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
800 const struct nf_conntrack_tuple *tuple, u32 hash)
802 struct nf_conntrack_tuple_hash *h;
807 h = ____nf_conntrack_find(net, zone, tuple, hash);
809 /* We have a candidate that matches the tuple we're interested
810 * in, try to obtain a reference and re-check tuple
812 ct = nf_ct_tuplehash_to_ctrack(h);
813 if (likely(atomic_inc_not_zero(&ct->ct_general.use))) {
814 if (likely(nf_ct_key_equal(h, tuple, zone, net)))
817 /* TYPESAFE_BY_RCU recycled the candidate */
829 struct nf_conntrack_tuple_hash *
830 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
831 const struct nf_conntrack_tuple *tuple)
833 unsigned int rid, zone_id = nf_ct_zone_id(zone, IP_CT_DIR_ORIGINAL);
834 struct nf_conntrack_tuple_hash *thash;
836 thash = __nf_conntrack_find_get(net, zone, tuple,
837 hash_conntrack_raw(tuple, zone_id, net));
842 rid = nf_ct_zone_id(zone, IP_CT_DIR_REPLY);
844 return __nf_conntrack_find_get(net, zone, tuple,
845 hash_conntrack_raw(tuple, rid, net));
848 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
850 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
852 unsigned int reply_hash)
854 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
855 &nf_conntrack_hash[hash]);
856 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
857 &nf_conntrack_hash[reply_hash]);
861 nf_conntrack_hash_check_insert(struct nf_conn *ct)
863 const struct nf_conntrack_zone *zone;
864 struct net *net = nf_ct_net(ct);
865 unsigned int hash, reply_hash;
866 struct nf_conntrack_tuple_hash *h;
867 struct hlist_nulls_node *n;
868 unsigned int max_chainlen;
869 unsigned int chainlen = 0;
870 unsigned int sequence;
873 zone = nf_ct_zone(ct);
877 sequence = read_seqcount_begin(&nf_conntrack_generation);
878 hash = hash_conntrack(net,
879 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
880 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_ORIGINAL));
881 reply_hash = hash_conntrack(net,
882 &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
883 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
884 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
886 max_chainlen = MIN_CHAINLEN + prandom_u32_max(MAX_CHAINLEN);
888 /* See if there's one in the list already, including reverse */
889 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode) {
890 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
894 if (chainlen++ > max_chainlen)
900 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode) {
901 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
904 if (chainlen++ > max_chainlen)
909 /* The caller holds a reference to this object */
910 atomic_set(&ct->ct_general.use, 2);
911 __nf_conntrack_hash_insert(ct, hash, reply_hash);
912 nf_conntrack_double_unlock(hash, reply_hash);
913 NF_CT_STAT_INC(net, insert);
917 NF_CT_STAT_INC(net, chaintoolong);
920 nf_conntrack_double_unlock(hash, reply_hash);
924 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
926 void nf_ct_acct_add(struct nf_conn *ct, u32 dir, unsigned int packets,
929 struct nf_conn_acct *acct;
931 acct = nf_conn_acct_find(ct);
933 struct nf_conn_counter *counter = acct->counter;
935 atomic64_add(packets, &counter[dir].packets);
936 atomic64_add(bytes, &counter[dir].bytes);
939 EXPORT_SYMBOL_GPL(nf_ct_acct_add);
941 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
942 const struct nf_conn *loser_ct)
944 struct nf_conn_acct *acct;
946 acct = nf_conn_acct_find(loser_ct);
948 struct nf_conn_counter *counter = acct->counter;
951 /* u32 should be fine since we must have seen one packet. */
952 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
953 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), bytes);
957 static void __nf_conntrack_insert_prepare(struct nf_conn *ct)
959 struct nf_conn_tstamp *tstamp;
961 atomic_inc(&ct->ct_general.use);
962 ct->status |= IPS_CONFIRMED;
964 /* set conntrack timestamp, if enabled. */
965 tstamp = nf_conn_tstamp_find(ct);
967 tstamp->start = ktime_get_real_ns();
970 /* caller must hold locks to prevent concurrent changes */
971 static int __nf_ct_resolve_clash(struct sk_buff *skb,
972 struct nf_conntrack_tuple_hash *h)
974 /* This is the conntrack entry already in hashes that won race. */
975 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
976 enum ip_conntrack_info ctinfo;
977 struct nf_conn *loser_ct;
979 loser_ct = nf_ct_get(skb, &ctinfo);
981 if (nf_ct_is_dying(ct))
984 if (((ct->status & IPS_NAT_DONE_MASK) == 0) ||
985 nf_ct_match(ct, loser_ct)) {
986 struct net *net = nf_ct_net(ct);
988 nf_conntrack_get(&ct->ct_general);
990 nf_ct_acct_merge(ct, ctinfo, loser_ct);
991 nf_ct_add_to_dying_list(loser_ct);
992 nf_conntrack_put(&loser_ct->ct_general);
993 nf_ct_set(skb, ct, ctinfo);
995 NF_CT_STAT_INC(net, clash_resolve);
1003 * nf_ct_resolve_clash_harder - attempt to insert clashing conntrack entry
1005 * @skb: skb that causes the collision
1006 * @repl_idx: hash slot for reply direction
1008 * Called when origin or reply direction had a clash.
1009 * The skb can be handled without packet drop provided the reply direction
1010 * is unique or there the existing entry has the identical tuple in both
1013 * Caller must hold conntrack table locks to prevent concurrent updates.
1015 * Returns NF_DROP if the clash could not be handled.
1017 static int nf_ct_resolve_clash_harder(struct sk_buff *skb, u32 repl_idx)
1019 struct nf_conn *loser_ct = (struct nf_conn *)skb_nfct(skb);
1020 const struct nf_conntrack_zone *zone;
1021 struct nf_conntrack_tuple_hash *h;
1022 struct hlist_nulls_node *n;
1025 zone = nf_ct_zone(loser_ct);
1026 net = nf_ct_net(loser_ct);
1028 /* Reply direction must never result in a clash, unless both origin
1029 * and reply tuples are identical.
1031 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[repl_idx], hnnode) {
1032 if (nf_ct_key_equal(h,
1033 &loser_ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1035 return __nf_ct_resolve_clash(skb, h);
1038 /* We want the clashing entry to go away real soon: 1 second timeout. */
1039 loser_ct->timeout = nfct_time_stamp + HZ;
1041 /* IPS_NAT_CLASH removes the entry automatically on the first
1042 * reply. Also prevents UDP tracker from moving the entry to
1043 * ASSURED state, i.e. the entry can always be evicted under
1046 loser_ct->status |= IPS_FIXED_TIMEOUT | IPS_NAT_CLASH;
1048 __nf_conntrack_insert_prepare(loser_ct);
1050 /* fake add for ORIGINAL dir: we want lookups to only find the entry
1051 * already in the table. This also hides the clashing entry from
1052 * ctnetlink iteration, i.e. conntrack -L won't show them.
1054 hlist_nulls_add_fake(&loser_ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
1056 hlist_nulls_add_head_rcu(&loser_ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
1057 &nf_conntrack_hash[repl_idx]);
1059 NF_CT_STAT_INC(net, clash_resolve);
1064 * nf_ct_resolve_clash - attempt to handle clash without packet drop
1066 * @skb: skb that causes the clash
1067 * @h: tuplehash of the clashing entry already in table
1068 * @reply_hash: hash slot for reply direction
1070 * A conntrack entry can be inserted to the connection tracking table
1071 * if there is no existing entry with an identical tuple.
1073 * If there is one, @skb (and the assocated, unconfirmed conntrack) has
1074 * to be dropped. In case @skb is retransmitted, next conntrack lookup
1075 * will find the already-existing entry.
1077 * The major problem with such packet drop is the extra delay added by
1078 * the packet loss -- it will take some time for a retransmit to occur
1079 * (or the sender to time out when waiting for a reply).
1081 * This function attempts to handle the situation without packet drop.
1083 * If @skb has no NAT transformation or if the colliding entries are
1084 * exactly the same, only the to-be-confirmed conntrack entry is discarded
1085 * and @skb is associated with the conntrack entry already in the table.
1087 * Failing that, the new, unconfirmed conntrack is still added to the table
1088 * provided that the collision only occurs in the ORIGINAL direction.
1089 * The new entry will be added only in the non-clashing REPLY direction,
1090 * so packets in the ORIGINAL direction will continue to match the existing
1091 * entry. The new entry will also have a fixed timeout so it expires --
1092 * due to the collision, it will only see reply traffic.
1094 * Returns NF_DROP if the clash could not be resolved.
1096 static __cold noinline int
1097 nf_ct_resolve_clash(struct sk_buff *skb, struct nf_conntrack_tuple_hash *h,
1100 /* This is the conntrack entry already in hashes that won race. */
1101 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1102 const struct nf_conntrack_l4proto *l4proto;
1103 enum ip_conntrack_info ctinfo;
1104 struct nf_conn *loser_ct;
1108 loser_ct = nf_ct_get(skb, &ctinfo);
1109 net = nf_ct_net(loser_ct);
1111 l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1112 if (!l4proto->allow_clash)
1115 ret = __nf_ct_resolve_clash(skb, h);
1116 if (ret == NF_ACCEPT)
1119 ret = nf_ct_resolve_clash_harder(skb, reply_hash);
1120 if (ret == NF_ACCEPT)
1124 nf_ct_add_to_dying_list(loser_ct);
1125 NF_CT_STAT_INC(net, drop);
1126 NF_CT_STAT_INC(net, insert_failed);
1130 /* Confirm a connection given skb; places it in hash table */
1132 __nf_conntrack_confirm(struct sk_buff *skb)
1134 unsigned int chainlen = 0, sequence, max_chainlen;
1135 const struct nf_conntrack_zone *zone;
1136 unsigned int hash, reply_hash;
1137 struct nf_conntrack_tuple_hash *h;
1139 struct nf_conn_help *help;
1140 struct hlist_nulls_node *n;
1141 enum ip_conntrack_info ctinfo;
1145 ct = nf_ct_get(skb, &ctinfo);
1146 net = nf_ct_net(ct);
1148 /* ipt_REJECT uses nf_conntrack_attach to attach related
1149 ICMP/TCP RST packets in other direction. Actual packet
1150 which created connection will be IP_CT_NEW or for an
1151 expected connection, IP_CT_RELATED. */
1152 if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
1155 zone = nf_ct_zone(ct);
1159 sequence = read_seqcount_begin(&nf_conntrack_generation);
1160 /* reuse the hash saved before */
1161 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
1162 hash = scale_hash(hash);
1163 reply_hash = hash_conntrack(net,
1164 &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1165 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
1166 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
1168 /* We're not in hash table, and we refuse to set up related
1169 * connections for unconfirmed conns. But packet copies and
1170 * REJECT will give spurious warnings here.
1173 /* Another skb with the same unconfirmed conntrack may
1174 * win the race. This may happen for bridge(br_flood)
1175 * or broadcast/multicast packets do skb_clone with
1176 * unconfirmed conntrack.
1178 if (unlikely(nf_ct_is_confirmed(ct))) {
1180 nf_conntrack_double_unlock(hash, reply_hash);
1185 pr_debug("Confirming conntrack %p\n", ct);
1186 /* We have to check the DYING flag after unlink to prevent
1187 * a race against nf_ct_get_next_corpse() possibly called from
1188 * user context, else we insert an already 'dead' hash, blocking
1189 * further use of that particular connection -JM.
1191 nf_ct_del_from_dying_or_unconfirmed_list(ct);
1193 if (unlikely(nf_ct_is_dying(ct))) {
1194 nf_ct_add_to_dying_list(ct);
1195 NF_CT_STAT_INC(net, insert_failed);
1199 max_chainlen = MIN_CHAINLEN + prandom_u32_max(MAX_CHAINLEN);
1200 /* See if there's one in the list already, including reverse:
1201 NAT could have grabbed it without realizing, since we're
1202 not in the hash. If there is, we lost race. */
1203 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode) {
1204 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1207 if (chainlen++ > max_chainlen)
1212 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode) {
1213 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1216 if (chainlen++ > max_chainlen) {
1218 nf_ct_add_to_dying_list(ct);
1219 NF_CT_STAT_INC(net, chaintoolong);
1220 NF_CT_STAT_INC(net, insert_failed);
1226 /* Timer relative to confirmation time, not original
1227 setting time, otherwise we'd get timer wrap in
1228 weird delay cases. */
1229 ct->timeout += nfct_time_stamp;
1231 __nf_conntrack_insert_prepare(ct);
1233 /* Since the lookup is lockless, hash insertion must be done after
1234 * starting the timer and setting the CONFIRMED bit. The RCU barriers
1235 * guarantee that no other CPU can find the conntrack before the above
1236 * stores are visible.
1238 __nf_conntrack_hash_insert(ct, hash, reply_hash);
1239 nf_conntrack_double_unlock(hash, reply_hash);
1242 help = nfct_help(ct);
1243 if (help && help->helper)
1244 nf_conntrack_event_cache(IPCT_HELPER, ct);
1246 nf_conntrack_event_cache(master_ct(ct) ?
1247 IPCT_RELATED : IPCT_NEW, ct);
1251 ret = nf_ct_resolve_clash(skb, h, reply_hash);
1253 nf_conntrack_double_unlock(hash, reply_hash);
1257 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
1259 /* Returns true if a connection correspondings to the tuple (required
1262 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
1263 const struct nf_conn *ignored_conntrack)
1265 struct net *net = nf_ct_net(ignored_conntrack);
1266 const struct nf_conntrack_zone *zone;
1267 struct nf_conntrack_tuple_hash *h;
1268 struct hlist_nulls_head *ct_hash;
1269 unsigned int hash, hsize;
1270 struct hlist_nulls_node *n;
1273 zone = nf_ct_zone(ignored_conntrack);
1277 nf_conntrack_get_ht(&ct_hash, &hsize);
1278 hash = __hash_conntrack(net, tuple, nf_ct_zone_id(zone, IP_CT_DIR_REPLY), hsize);
1280 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
1281 ct = nf_ct_tuplehash_to_ctrack(h);
1283 if (ct == ignored_conntrack)
1286 if (nf_ct_is_expired(ct)) {
1287 nf_ct_gc_expired(ct);
1291 if (nf_ct_key_equal(h, tuple, zone, net)) {
1292 /* Tuple is taken already, so caller will need to find
1293 * a new source port to use.
1296 * If the *original tuples* are identical, then both
1297 * conntracks refer to the same flow.
1298 * This is a rare situation, it can occur e.g. when
1299 * more than one UDP packet is sent from same socket
1300 * in different threads.
1302 * Let nf_ct_resolve_clash() deal with this later.
1304 if (nf_ct_tuple_equal(&ignored_conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1305 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
1306 nf_ct_zone_equal(ct, zone, IP_CT_DIR_ORIGINAL))
1309 NF_CT_STAT_INC_ATOMIC(net, found);
1315 if (get_nulls_value(n) != hash) {
1316 NF_CT_STAT_INC_ATOMIC(net, search_restart);
1324 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
1326 #define NF_CT_EVICTION_RANGE 8
1328 /* There's a small race here where we may free a just-assured
1329 connection. Too bad: we're in trouble anyway. */
1330 static unsigned int early_drop_list(struct net *net,
1331 struct hlist_nulls_head *head)
1333 struct nf_conntrack_tuple_hash *h;
1334 struct hlist_nulls_node *n;
1335 unsigned int drops = 0;
1336 struct nf_conn *tmp;
1338 hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
1339 tmp = nf_ct_tuplehash_to_ctrack(h);
1341 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status))
1344 if (nf_ct_is_expired(tmp)) {
1345 nf_ct_gc_expired(tmp);
1349 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
1350 !net_eq(nf_ct_net(tmp), net) ||
1351 nf_ct_is_dying(tmp))
1354 if (!atomic_inc_not_zero(&tmp->ct_general.use))
1357 /* kill only if still in same netns -- might have moved due to
1358 * SLAB_TYPESAFE_BY_RCU rules.
1360 * We steal the timer reference. If that fails timer has
1361 * already fired or someone else deleted it. Just drop ref
1362 * and move to next entry.
1364 if (net_eq(nf_ct_net(tmp), net) &&
1365 nf_ct_is_confirmed(tmp) &&
1366 nf_ct_delete(tmp, 0, 0))
1375 static noinline int early_drop(struct net *net, unsigned int hash)
1377 unsigned int i, bucket;
1379 for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
1380 struct hlist_nulls_head *ct_hash;
1381 unsigned int hsize, drops;
1384 nf_conntrack_get_ht(&ct_hash, &hsize);
1386 bucket = reciprocal_scale(hash, hsize);
1388 bucket = (bucket + 1) % hsize;
1390 drops = early_drop_list(net, &ct_hash[bucket]);
1394 NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
1402 static bool gc_worker_skip_ct(const struct nf_conn *ct)
1404 return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
1407 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
1409 const struct nf_conntrack_l4proto *l4proto;
1411 if (!test_bit(IPS_ASSURED_BIT, &ct->status))
1414 l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1415 if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
1421 static void gc_worker(struct work_struct *work)
1423 unsigned long end_time = jiffies + GC_SCAN_MAX_DURATION;
1424 unsigned int i, hashsz, nf_conntrack_max95 = 0;
1425 unsigned long next_run = GC_SCAN_INTERVAL;
1426 struct conntrack_gc_work *gc_work;
1427 gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
1429 i = gc_work->next_bucket;
1430 if (gc_work->early_drop)
1431 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
1434 struct nf_conntrack_tuple_hash *h;
1435 struct hlist_nulls_head *ct_hash;
1436 struct hlist_nulls_node *n;
1437 struct nf_conn *tmp;
1441 nf_conntrack_get_ht(&ct_hash, &hashsz);
1447 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1448 struct nf_conntrack_net *cnet;
1451 tmp = nf_ct_tuplehash_to_ctrack(h);
1453 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status)) {
1454 nf_ct_offload_timeout(tmp);
1458 if (nf_ct_is_expired(tmp)) {
1459 nf_ct_gc_expired(tmp);
1463 if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1466 net = nf_ct_net(tmp);
1467 cnet = nf_ct_pernet(net);
1468 if (atomic_read(&cnet->count) < nf_conntrack_max95)
1471 /* need to take reference to avoid possible races */
1472 if (!atomic_inc_not_zero(&tmp->ct_general.use))
1475 if (gc_worker_skip_ct(tmp)) {
1480 if (gc_worker_can_early_drop(tmp))
1486 /* could check get_nulls_value() here and restart if ct
1487 * was moved to another chain. But given gc is best-effort
1488 * we will just continue with next hash slot.
1494 if (time_after(jiffies, end_time) && i < hashsz) {
1495 gc_work->next_bucket = i;
1499 } while (i < hashsz);
1501 if (gc_work->exiting)
1505 * Eviction will normally happen from the packet path, and not
1506 * from this gc worker.
1508 * This worker is only here to reap expired entries when system went
1509 * idle after a busy period.
1512 gc_work->early_drop = false;
1513 gc_work->next_bucket = 0;
1515 queue_delayed_work(system_power_efficient_wq, &gc_work->dwork, next_run);
1518 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1520 INIT_DEFERRABLE_WORK(&gc_work->dwork, gc_worker);
1521 gc_work->exiting = false;
1524 static struct nf_conn *
1525 __nf_conntrack_alloc(struct net *net,
1526 const struct nf_conntrack_zone *zone,
1527 const struct nf_conntrack_tuple *orig,
1528 const struct nf_conntrack_tuple *repl,
1529 gfp_t gfp, u32 hash)
1531 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
1532 unsigned int ct_count;
1535 /* We don't want any race condition at early drop stage */
1536 ct_count = atomic_inc_return(&cnet->count);
1538 if (nf_conntrack_max && unlikely(ct_count > nf_conntrack_max)) {
1539 if (!early_drop(net, hash)) {
1540 if (!conntrack_gc_work.early_drop)
1541 conntrack_gc_work.early_drop = true;
1542 atomic_dec(&cnet->count);
1543 net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1544 return ERR_PTR(-ENOMEM);
1549 * Do not use kmem_cache_zalloc(), as this cache uses
1550 * SLAB_TYPESAFE_BY_RCU.
1552 ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1556 spin_lock_init(&ct->lock);
1557 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1558 ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1559 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1560 /* save hash for reusing when confirming */
1561 *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1564 write_pnet(&ct->ct_net, net);
1565 memset(&ct->__nfct_init_offset, 0,
1566 offsetof(struct nf_conn, proto) -
1567 offsetof(struct nf_conn, __nfct_init_offset));
1569 nf_ct_zone_add(ct, zone);
1571 /* Because we use RCU lookups, we set ct_general.use to zero before
1572 * this is inserted in any list.
1574 atomic_set(&ct->ct_general.use, 0);
1577 atomic_dec(&cnet->count);
1578 return ERR_PTR(-ENOMEM);
1581 struct nf_conn *nf_conntrack_alloc(struct net *net,
1582 const struct nf_conntrack_zone *zone,
1583 const struct nf_conntrack_tuple *orig,
1584 const struct nf_conntrack_tuple *repl,
1587 return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1589 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1591 void nf_conntrack_free(struct nf_conn *ct)
1593 struct net *net = nf_ct_net(ct);
1594 struct nf_conntrack_net *cnet;
1596 /* A freed object has refcnt == 0, that's
1597 * the golden rule for SLAB_TYPESAFE_BY_RCU
1599 WARN_ON(atomic_read(&ct->ct_general.use) != 0);
1601 nf_ct_ext_destroy(ct);
1602 kmem_cache_free(nf_conntrack_cachep, ct);
1603 cnet = nf_ct_pernet(net);
1605 smp_mb__before_atomic();
1606 atomic_dec(&cnet->count);
1608 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1611 /* Allocate a new conntrack: we return -ENOMEM if classification
1612 failed due to stress. Otherwise it really is unclassifiable. */
1613 static noinline struct nf_conntrack_tuple_hash *
1614 init_conntrack(struct net *net, struct nf_conn *tmpl,
1615 const struct nf_conntrack_tuple *tuple,
1616 struct sk_buff *skb,
1617 unsigned int dataoff, u32 hash)
1620 struct nf_conn_help *help;
1621 struct nf_conntrack_tuple repl_tuple;
1622 struct nf_conntrack_ecache *ecache;
1623 struct nf_conntrack_expect *exp = NULL;
1624 const struct nf_conntrack_zone *zone;
1625 struct nf_conn_timeout *timeout_ext;
1626 struct nf_conntrack_zone tmp;
1627 struct nf_conntrack_net *cnet;
1629 if (!nf_ct_invert_tuple(&repl_tuple, tuple)) {
1630 pr_debug("Can't invert tuple.\n");
1634 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1635 ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1638 return (struct nf_conntrack_tuple_hash *)ct;
1640 if (!nf_ct_add_synproxy(ct, tmpl)) {
1641 nf_conntrack_free(ct);
1642 return ERR_PTR(-ENOMEM);
1645 timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1648 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1651 nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1652 nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1653 nf_ct_labels_ext_add(ct);
1655 ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1656 nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1657 ecache ? ecache->expmask : 0,
1661 cnet = nf_ct_pernet(net);
1662 if (cnet->expect_count) {
1663 spin_lock(&nf_conntrack_expect_lock);
1664 exp = nf_ct_find_expectation(net, zone, tuple);
1666 pr_debug("expectation arrives ct=%p exp=%p\n",
1668 /* Welcome, Mr. Bond. We've been expecting you... */
1669 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1670 /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1671 ct->master = exp->master;
1673 help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1675 rcu_assign_pointer(help->helper, exp->helper);
1678 #ifdef CONFIG_NF_CONNTRACK_MARK
1679 ct->mark = exp->master->mark;
1681 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1682 ct->secmark = exp->master->secmark;
1684 NF_CT_STAT_INC(net, expect_new);
1686 spin_unlock(&nf_conntrack_expect_lock);
1689 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1691 /* Now it is inserted into the unconfirmed list, bump refcount */
1692 nf_conntrack_get(&ct->ct_general);
1693 nf_ct_add_to_unconfirmed_list(ct);
1699 exp->expectfn(ct, exp);
1700 nf_ct_expect_put(exp);
1703 return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1706 /* On success, returns 0, sets skb->_nfct | ctinfo */
1708 resolve_normal_ct(struct nf_conn *tmpl,
1709 struct sk_buff *skb,
1710 unsigned int dataoff,
1712 const struct nf_hook_state *state)
1714 const struct nf_conntrack_zone *zone;
1715 struct nf_conntrack_tuple tuple;
1716 struct nf_conntrack_tuple_hash *h;
1717 enum ip_conntrack_info ctinfo;
1718 struct nf_conntrack_zone tmp;
1719 u32 hash, zone_id, rid;
1722 if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1723 dataoff, state->pf, protonum, state->net,
1725 pr_debug("Can't get tuple\n");
1729 /* look for tuple match */
1730 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1732 zone_id = nf_ct_zone_id(zone, IP_CT_DIR_ORIGINAL);
1733 hash = hash_conntrack_raw(&tuple, zone_id, state->net);
1734 h = __nf_conntrack_find_get(state->net, zone, &tuple, hash);
1737 rid = nf_ct_zone_id(zone, IP_CT_DIR_REPLY);
1738 if (zone_id != rid) {
1739 u32 tmp = hash_conntrack_raw(&tuple, rid, state->net);
1741 h = __nf_conntrack_find_get(state->net, zone, &tuple, tmp);
1746 h = init_conntrack(state->net, tmpl, &tuple,
1747 skb, dataoff, hash);
1753 ct = nf_ct_tuplehash_to_ctrack(h);
1755 /* It exists; we have (non-exclusive) reference. */
1756 if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1757 ctinfo = IP_CT_ESTABLISHED_REPLY;
1759 /* Once we've had two way comms, always ESTABLISHED. */
1760 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1761 pr_debug("normal packet for %p\n", ct);
1762 ctinfo = IP_CT_ESTABLISHED;
1763 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1764 pr_debug("related packet for %p\n", ct);
1765 ctinfo = IP_CT_RELATED;
1767 pr_debug("new packet for %p\n", ct);
1771 nf_ct_set(skb, ct, ctinfo);
1776 * icmp packets need special treatment to handle error messages that are
1777 * related to a connection.
1779 * Callers need to check if skb has a conntrack assigned when this
1780 * helper returns; in such case skb belongs to an already known connection.
1782 static unsigned int __cold
1783 nf_conntrack_handle_icmp(struct nf_conn *tmpl,
1784 struct sk_buff *skb,
1785 unsigned int dataoff,
1787 const struct nf_hook_state *state)
1791 if (state->pf == NFPROTO_IPV4 && protonum == IPPROTO_ICMP)
1792 ret = nf_conntrack_icmpv4_error(tmpl, skb, dataoff, state);
1793 #if IS_ENABLED(CONFIG_IPV6)
1794 else if (state->pf == NFPROTO_IPV6 && protonum == IPPROTO_ICMPV6)
1795 ret = nf_conntrack_icmpv6_error(tmpl, skb, dataoff, state);
1801 NF_CT_STAT_INC_ATOMIC(state->net, error);
1806 static int generic_packet(struct nf_conn *ct, struct sk_buff *skb,
1807 enum ip_conntrack_info ctinfo)
1809 const unsigned int *timeout = nf_ct_timeout_lookup(ct);
1812 timeout = &nf_generic_pernet(nf_ct_net(ct))->timeout;
1814 nf_ct_refresh_acct(ct, ctinfo, skb, *timeout);
1818 /* Returns verdict for packet, or -1 for invalid. */
1819 static int nf_conntrack_handle_packet(struct nf_conn *ct,
1820 struct sk_buff *skb,
1821 unsigned int dataoff,
1822 enum ip_conntrack_info ctinfo,
1823 const struct nf_hook_state *state)
1825 switch (nf_ct_protonum(ct)) {
1827 return nf_conntrack_tcp_packet(ct, skb, dataoff,
1830 return nf_conntrack_udp_packet(ct, skb, dataoff,
1833 return nf_conntrack_icmp_packet(ct, skb, ctinfo, state);
1834 #if IS_ENABLED(CONFIG_IPV6)
1835 case IPPROTO_ICMPV6:
1836 return nf_conntrack_icmpv6_packet(ct, skb, ctinfo, state);
1838 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
1839 case IPPROTO_UDPLITE:
1840 return nf_conntrack_udplite_packet(ct, skb, dataoff,
1843 #ifdef CONFIG_NF_CT_PROTO_SCTP
1845 return nf_conntrack_sctp_packet(ct, skb, dataoff,
1848 #ifdef CONFIG_NF_CT_PROTO_DCCP
1850 return nf_conntrack_dccp_packet(ct, skb, dataoff,
1853 #ifdef CONFIG_NF_CT_PROTO_GRE
1855 return nf_conntrack_gre_packet(ct, skb, dataoff,
1860 return generic_packet(ct, skb, ctinfo);
1864 nf_conntrack_in(struct sk_buff *skb, const struct nf_hook_state *state)
1866 enum ip_conntrack_info ctinfo;
1867 struct nf_conn *ct, *tmpl;
1871 tmpl = nf_ct_get(skb, &ctinfo);
1872 if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1873 /* Previously seen (loopback or untracked)? Ignore. */
1874 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1875 ctinfo == IP_CT_UNTRACKED)
1880 /* rcu_read_lock()ed by nf_hook_thresh */
1881 dataoff = get_l4proto(skb, skb_network_offset(skb), state->pf, &protonum);
1883 pr_debug("not prepared to track yet or error occurred\n");
1884 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1889 if (protonum == IPPROTO_ICMP || protonum == IPPROTO_ICMPV6) {
1890 ret = nf_conntrack_handle_icmp(tmpl, skb, dataoff,
1896 /* ICMP[v6] protocol trackers may assign one conntrack. */
1901 ret = resolve_normal_ct(tmpl, skb, dataoff,
1904 /* Too stressed to deal. */
1905 NF_CT_STAT_INC_ATOMIC(state->net, drop);
1910 ct = nf_ct_get(skb, &ctinfo);
1912 /* Not valid part of a connection */
1913 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1918 ret = nf_conntrack_handle_packet(ct, skb, dataoff, ctinfo, state);
1920 /* Invalid: inverse of the return code tells
1921 * the netfilter core what to do */
1922 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1923 nf_conntrack_put(&ct->ct_general);
1925 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1926 if (ret == -NF_DROP)
1927 NF_CT_STAT_INC_ATOMIC(state->net, drop);
1928 /* Special case: TCP tracker reports an attempt to reopen a
1929 * closed/aborted connection. We have to go back and create a
1932 if (ret == -NF_REPEAT)
1938 if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
1939 !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1940 nf_conntrack_event_cache(IPCT_REPLY, ct);
1947 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1949 /* Alter reply tuple (maybe alter helper). This is for NAT, and is
1950 implicitly racy: see __nf_conntrack_confirm */
1951 void nf_conntrack_alter_reply(struct nf_conn *ct,
1952 const struct nf_conntrack_tuple *newreply)
1954 struct nf_conn_help *help = nfct_help(ct);
1956 /* Should be unconfirmed, so not in hash table yet */
1957 WARN_ON(nf_ct_is_confirmed(ct));
1959 pr_debug("Altering reply tuple of %p to ", ct);
1960 nf_ct_dump_tuple(newreply);
1962 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1963 if (ct->master || (help && !hlist_empty(&help->expectations)))
1967 __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1970 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1972 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1973 void __nf_ct_refresh_acct(struct nf_conn *ct,
1974 enum ip_conntrack_info ctinfo,
1975 const struct sk_buff *skb,
1979 /* Only update if this is not a fixed timeout */
1980 if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1983 /* If not in hash table, timer will not be active yet */
1984 if (nf_ct_is_confirmed(ct))
1985 extra_jiffies += nfct_time_stamp;
1987 if (READ_ONCE(ct->timeout) != extra_jiffies)
1988 WRITE_ONCE(ct->timeout, extra_jiffies);
1991 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
1993 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1995 bool nf_ct_kill_acct(struct nf_conn *ct,
1996 enum ip_conntrack_info ctinfo,
1997 const struct sk_buff *skb)
1999 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
2001 return nf_ct_delete(ct, 0, 0);
2003 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
2005 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
2007 #include <linux/netfilter/nfnetlink.h>
2008 #include <linux/netfilter/nfnetlink_conntrack.h>
2009 #include <linux/mutex.h>
2011 /* Generic function for tcp/udp/sctp/dccp and alike. */
2012 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
2013 const struct nf_conntrack_tuple *tuple)
2015 if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
2016 nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
2017 goto nla_put_failure;
2023 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
2025 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
2026 [CTA_PROTO_SRC_PORT] = { .type = NLA_U16 },
2027 [CTA_PROTO_DST_PORT] = { .type = NLA_U16 },
2029 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
2031 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
2032 struct nf_conntrack_tuple *t,
2035 if (flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) {
2036 if (!tb[CTA_PROTO_SRC_PORT])
2039 t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
2042 if (flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) {
2043 if (!tb[CTA_PROTO_DST_PORT])
2046 t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
2051 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
2053 unsigned int nf_ct_port_nlattr_tuple_size(void)
2055 static unsigned int size __read_mostly;
2058 size = nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
2062 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
2065 /* Used by ipt_REJECT and ip6t_REJECT. */
2066 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
2069 enum ip_conntrack_info ctinfo;
2071 /* This ICMP is in reverse direction to the packet which caused it */
2072 ct = nf_ct_get(skb, &ctinfo);
2073 if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
2074 ctinfo = IP_CT_RELATED_REPLY;
2076 ctinfo = IP_CT_RELATED;
2078 /* Attach to new skbuff, and increment count */
2079 nf_ct_set(nskb, ct, ctinfo);
2080 nf_conntrack_get(skb_nfct(nskb));
2083 static int __nf_conntrack_update(struct net *net, struct sk_buff *skb,
2085 enum ip_conntrack_info ctinfo)
2087 struct nf_conntrack_tuple_hash *h;
2088 struct nf_conntrack_tuple tuple;
2089 struct nf_nat_hook *nat_hook;
2090 unsigned int status;
2095 l3num = nf_ct_l3num(ct);
2097 dataoff = get_l4proto(skb, skb_network_offset(skb), l3num, &l4num);
2101 if (!nf_ct_get_tuple(skb, skb_network_offset(skb), dataoff, l3num,
2102 l4num, net, &tuple))
2105 if (ct->status & IPS_SRC_NAT) {
2106 memcpy(tuple.src.u3.all,
2107 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.all,
2108 sizeof(tuple.src.u3.all));
2110 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u.all;
2113 if (ct->status & IPS_DST_NAT) {
2114 memcpy(tuple.dst.u3.all,
2115 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.all,
2116 sizeof(tuple.dst.u3.all));
2118 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u.all;
2121 h = nf_conntrack_find_get(net, nf_ct_zone(ct), &tuple);
2125 /* Store status bits of the conntrack that is clashing to re-do NAT
2126 * mangling according to what it has been done already to this packet.
2128 status = ct->status;
2131 ct = nf_ct_tuplehash_to_ctrack(h);
2132 nf_ct_set(skb, ct, ctinfo);
2134 nat_hook = rcu_dereference(nf_nat_hook);
2138 if (status & IPS_SRC_NAT &&
2139 nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_SRC,
2140 IP_CT_DIR_ORIGINAL) == NF_DROP)
2143 if (status & IPS_DST_NAT &&
2144 nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_DST,
2145 IP_CT_DIR_ORIGINAL) == NF_DROP)
2151 /* This packet is coming from userspace via nf_queue, complete the packet
2152 * processing after the helper invocation in nf_confirm().
2154 static int nf_confirm_cthelper(struct sk_buff *skb, struct nf_conn *ct,
2155 enum ip_conntrack_info ctinfo)
2157 const struct nf_conntrack_helper *helper;
2158 const struct nf_conn_help *help;
2161 help = nfct_help(ct);
2165 helper = rcu_dereference(help->helper);
2166 if (!(helper->flags & NF_CT_HELPER_F_USERSPACE))
2169 switch (nf_ct_l3num(ct)) {
2171 protoff = skb_network_offset(skb) + ip_hdrlen(skb);
2173 #if IS_ENABLED(CONFIG_IPV6)
2174 case NFPROTO_IPV6: {
2178 pnum = ipv6_hdr(skb)->nexthdr;
2179 protoff = ipv6_skip_exthdr(skb, sizeof(struct ipv6hdr), &pnum,
2181 if (protoff < 0 || (frag_off & htons(~0x7)) != 0)
2190 if (test_bit(IPS_SEQ_ADJUST_BIT, &ct->status) &&
2191 !nf_is_loopback_packet(skb)) {
2192 if (!nf_ct_seq_adjust(skb, ct, ctinfo, protoff)) {
2193 NF_CT_STAT_INC_ATOMIC(nf_ct_net(ct), drop);
2198 /* We've seen it coming out the other side: confirm it */
2199 return nf_conntrack_confirm(skb) == NF_DROP ? - 1 : 0;
2202 static int nf_conntrack_update(struct net *net, struct sk_buff *skb)
2204 enum ip_conntrack_info ctinfo;
2208 ct = nf_ct_get(skb, &ctinfo);
2212 if (!nf_ct_is_confirmed(ct)) {
2213 err = __nf_conntrack_update(net, skb, ct, ctinfo);
2217 ct = nf_ct_get(skb, &ctinfo);
2220 return nf_confirm_cthelper(skb, ct, ctinfo);
2223 static bool nf_conntrack_get_tuple_skb(struct nf_conntrack_tuple *dst_tuple,
2224 const struct sk_buff *skb)
2226 const struct nf_conntrack_tuple *src_tuple;
2227 const struct nf_conntrack_tuple_hash *hash;
2228 struct nf_conntrack_tuple srctuple;
2229 enum ip_conntrack_info ctinfo;
2232 ct = nf_ct_get(skb, &ctinfo);
2234 src_tuple = nf_ct_tuple(ct, CTINFO2DIR(ctinfo));
2235 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2239 if (!nf_ct_get_tuplepr(skb, skb_network_offset(skb),
2240 NFPROTO_IPV4, dev_net(skb->dev),
2244 hash = nf_conntrack_find_get(dev_net(skb->dev),
2250 ct = nf_ct_tuplehash_to_ctrack(hash);
2251 src_tuple = nf_ct_tuple(ct, !hash->tuple.dst.dir);
2252 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2258 /* Bring out ya dead! */
2259 static struct nf_conn *
2260 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
2261 void *data, unsigned int *bucket)
2263 struct nf_conntrack_tuple_hash *h;
2265 struct hlist_nulls_node *n;
2268 for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
2269 struct hlist_nulls_head *hslot = &nf_conntrack_hash[*bucket];
2271 if (hlist_nulls_empty(hslot))
2274 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
2276 nf_conntrack_lock(lockp);
2277 hlist_nulls_for_each_entry(h, n, hslot, hnnode) {
2278 if (NF_CT_DIRECTION(h) != IP_CT_DIR_REPLY)
2280 /* All nf_conn objects are added to hash table twice, one
2281 * for original direction tuple, once for the reply tuple.
2283 * Exception: In the IPS_NAT_CLASH case, only the reply
2284 * tuple is added (the original tuple already existed for
2285 * a different object).
2287 * We only need to call the iterator once for each
2288 * conntrack, so we just use the 'reply' direction
2289 * tuple while iterating.
2291 ct = nf_ct_tuplehash_to_ctrack(h);
2302 atomic_inc(&ct->ct_general.use);
2308 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
2309 void *data, u32 portid, int report)
2311 unsigned int bucket = 0;
2316 mutex_lock(&nf_conntrack_mutex);
2317 while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
2318 /* Time to push up daises... */
2320 nf_ct_delete(ct, portid, report);
2324 mutex_unlock(&nf_conntrack_mutex);
2328 int (*iter)(struct nf_conn *i, void *data);
2333 static int iter_net_only(struct nf_conn *i, void *data)
2335 struct iter_data *d = data;
2337 if (!net_eq(d->net, nf_ct_net(i)))
2340 return d->iter(i, d->data);
2344 __nf_ct_unconfirmed_destroy(struct net *net)
2348 for_each_possible_cpu(cpu) {
2349 struct nf_conntrack_tuple_hash *h;
2350 struct hlist_nulls_node *n;
2351 struct ct_pcpu *pcpu;
2353 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2355 spin_lock_bh(&pcpu->lock);
2356 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
2359 ct = nf_ct_tuplehash_to_ctrack(h);
2361 /* we cannot call iter() on unconfirmed list, the
2362 * owning cpu can reallocate ct->ext at any time.
2364 set_bit(IPS_DYING_BIT, &ct->status);
2366 spin_unlock_bh(&pcpu->lock);
2371 void nf_ct_unconfirmed_destroy(struct net *net)
2373 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2377 if (atomic_read(&cnet->count) > 0) {
2378 __nf_ct_unconfirmed_destroy(net);
2379 nf_queue_nf_hook_drop(net);
2383 EXPORT_SYMBOL_GPL(nf_ct_unconfirmed_destroy);
2385 void nf_ct_iterate_cleanup_net(struct net *net,
2386 int (*iter)(struct nf_conn *i, void *data),
2387 void *data, u32 portid, int report)
2389 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2394 if (atomic_read(&cnet->count) == 0)
2401 nf_ct_iterate_cleanup(iter_net_only, &d, portid, report);
2403 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
2406 * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
2407 * @iter: callback to invoke for each conntrack
2408 * @data: data to pass to @iter
2410 * Like nf_ct_iterate_cleanup, but first marks conntracks on the
2411 * unconfirmed list as dying (so they will not be inserted into
2414 * Can only be called in module exit path.
2417 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
2421 down_read(&net_rwsem);
2423 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2425 if (atomic_read(&cnet->count) == 0)
2427 __nf_ct_unconfirmed_destroy(net);
2428 nf_queue_nf_hook_drop(net);
2430 up_read(&net_rwsem);
2432 /* Need to wait for netns cleanup worker to finish, if its
2433 * running -- it might have deleted a net namespace from
2434 * the global list, so our __nf_ct_unconfirmed_destroy() might
2435 * not have affected all namespaces.
2439 /* a conntrack could have been unlinked from unconfirmed list
2440 * before we grabbed pcpu lock in __nf_ct_unconfirmed_destroy().
2441 * This makes sure its inserted into conntrack table.
2445 nf_ct_iterate_cleanup(iter, data, 0, 0);
2447 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
2449 static int kill_all(struct nf_conn *i, void *data)
2451 return net_eq(nf_ct_net(i), data);
2454 void nf_conntrack_cleanup_start(void)
2456 conntrack_gc_work.exiting = true;
2457 RCU_INIT_POINTER(ip_ct_attach, NULL);
2460 void nf_conntrack_cleanup_end(void)
2462 RCU_INIT_POINTER(nf_ct_hook, NULL);
2463 cancel_delayed_work_sync(&conntrack_gc_work.dwork);
2464 kvfree(nf_conntrack_hash);
2466 nf_conntrack_proto_fini();
2467 nf_conntrack_seqadj_fini();
2468 nf_conntrack_labels_fini();
2469 nf_conntrack_helper_fini();
2470 nf_conntrack_timeout_fini();
2471 nf_conntrack_ecache_fini();
2472 nf_conntrack_tstamp_fini();
2473 nf_conntrack_acct_fini();
2474 nf_conntrack_expect_fini();
2476 kmem_cache_destroy(nf_conntrack_cachep);
2480 * Mishearing the voices in his head, our hero wonders how he's
2481 * supposed to kill the mall.
2483 void nf_conntrack_cleanup_net(struct net *net)
2487 list_add(&net->exit_list, &single);
2488 nf_conntrack_cleanup_net_list(&single);
2491 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
2497 * This makes sure all current packets have passed through
2498 * netfilter framework. Roll on, two-stage module
2504 list_for_each_entry(net, net_exit_list, exit_list) {
2505 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2507 nf_ct_iterate_cleanup(kill_all, net, 0, 0);
2508 if (atomic_read(&cnet->count) != 0)
2513 goto i_see_dead_people;
2516 list_for_each_entry(net, net_exit_list, exit_list) {
2517 nf_conntrack_ecache_pernet_fini(net);
2518 nf_conntrack_expect_pernet_fini(net);
2519 free_percpu(net->ct.stat);
2520 free_percpu(net->ct.pcpu_lists);
2524 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
2526 struct hlist_nulls_head *hash;
2527 unsigned int nr_slots, i;
2529 if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
2532 BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
2533 nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
2535 hash = kvcalloc(nr_slots, sizeof(struct hlist_nulls_head), GFP_KERNEL);
2538 for (i = 0; i < nr_slots; i++)
2539 INIT_HLIST_NULLS_HEAD(&hash[i], i);
2543 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
2545 int nf_conntrack_hash_resize(unsigned int hashsize)
2548 unsigned int old_size;
2549 struct hlist_nulls_head *hash, *old_hash;
2550 struct nf_conntrack_tuple_hash *h;
2556 hash = nf_ct_alloc_hashtable(&hashsize, 1);
2560 mutex_lock(&nf_conntrack_mutex);
2561 old_size = nf_conntrack_htable_size;
2562 if (old_size == hashsize) {
2563 mutex_unlock(&nf_conntrack_mutex);
2569 nf_conntrack_all_lock();
2570 write_seqcount_begin(&nf_conntrack_generation);
2572 /* Lookups in the old hash might happen in parallel, which means we
2573 * might get false negatives during connection lookup. New connections
2574 * created because of a false negative won't make it into the hash
2575 * though since that required taking the locks.
2578 for (i = 0; i < nf_conntrack_htable_size; i++) {
2579 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
2580 unsigned int zone_id;
2582 h = hlist_nulls_entry(nf_conntrack_hash[i].first,
2583 struct nf_conntrack_tuple_hash, hnnode);
2584 ct = nf_ct_tuplehash_to_ctrack(h);
2585 hlist_nulls_del_rcu(&h->hnnode);
2587 zone_id = nf_ct_zone_id(nf_ct_zone(ct), NF_CT_DIRECTION(h));
2588 bucket = __hash_conntrack(nf_ct_net(ct),
2589 &h->tuple, zone_id, hashsize);
2590 hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
2593 old_size = nf_conntrack_htable_size;
2594 old_hash = nf_conntrack_hash;
2596 nf_conntrack_hash = hash;
2597 nf_conntrack_htable_size = hashsize;
2599 write_seqcount_end(&nf_conntrack_generation);
2600 nf_conntrack_all_unlock();
2603 mutex_unlock(&nf_conntrack_mutex);
2610 int nf_conntrack_set_hashsize(const char *val, const struct kernel_param *kp)
2612 unsigned int hashsize;
2615 if (current->nsproxy->net_ns != &init_net)
2618 /* On boot, we can set this without any fancy locking. */
2619 if (!nf_conntrack_hash)
2620 return param_set_uint(val, kp);
2622 rc = kstrtouint(val, 0, &hashsize);
2626 return nf_conntrack_hash_resize(hashsize);
2629 static __always_inline unsigned int total_extension_size(void)
2631 /* remember to add new extensions below */
2632 BUILD_BUG_ON(NF_CT_EXT_NUM > 9);
2634 return sizeof(struct nf_ct_ext) +
2635 sizeof(struct nf_conn_help)
2636 #if IS_ENABLED(CONFIG_NF_NAT)
2637 + sizeof(struct nf_conn_nat)
2639 + sizeof(struct nf_conn_seqadj)
2640 + sizeof(struct nf_conn_acct)
2641 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2642 + sizeof(struct nf_conntrack_ecache)
2644 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
2645 + sizeof(struct nf_conn_tstamp)
2647 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
2648 + sizeof(struct nf_conn_timeout)
2650 #ifdef CONFIG_NF_CONNTRACK_LABELS
2651 + sizeof(struct nf_conn_labels)
2653 #if IS_ENABLED(CONFIG_NETFILTER_SYNPROXY)
2654 + sizeof(struct nf_conn_synproxy)
2659 int nf_conntrack_init_start(void)
2661 unsigned long nr_pages = totalram_pages();
2666 /* struct nf_ct_ext uses u8 to store offsets/size */
2667 BUILD_BUG_ON(total_extension_size() > 255u);
2669 seqcount_spinlock_init(&nf_conntrack_generation,
2670 &nf_conntrack_locks_all_lock);
2672 for (i = 0; i < CONNTRACK_LOCKS; i++)
2673 spin_lock_init(&nf_conntrack_locks[i]);
2675 if (!nf_conntrack_htable_size) {
2676 nf_conntrack_htable_size
2677 = (((nr_pages << PAGE_SHIFT) / 16384)
2678 / sizeof(struct hlist_head));
2679 if (BITS_PER_LONG >= 64 &&
2680 nr_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2681 nf_conntrack_htable_size = 262144;
2682 else if (nr_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2683 nf_conntrack_htable_size = 65536;
2685 if (nf_conntrack_htable_size < 1024)
2686 nf_conntrack_htable_size = 1024;
2687 /* Use a max. factor of one by default to keep the average
2688 * hash chain length at 2 entries. Each entry has to be added
2689 * twice (once for original direction, once for reply).
2690 * When a table size is given we use the old value of 8 to
2691 * avoid implicit reduction of the max entries setting.
2696 nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2697 if (!nf_conntrack_hash)
2700 nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2702 nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2703 sizeof(struct nf_conn),
2705 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2706 if (!nf_conntrack_cachep)
2709 ret = nf_conntrack_expect_init();
2713 ret = nf_conntrack_acct_init();
2717 ret = nf_conntrack_tstamp_init();
2721 ret = nf_conntrack_ecache_init();
2725 ret = nf_conntrack_timeout_init();
2729 ret = nf_conntrack_helper_init();
2733 ret = nf_conntrack_labels_init();
2737 ret = nf_conntrack_seqadj_init();
2741 ret = nf_conntrack_proto_init();
2745 conntrack_gc_work_init(&conntrack_gc_work);
2746 queue_delayed_work(system_power_efficient_wq, &conntrack_gc_work.dwork, HZ);
2751 nf_conntrack_seqadj_fini();
2753 nf_conntrack_labels_fini();
2755 nf_conntrack_helper_fini();
2757 nf_conntrack_timeout_fini();
2759 nf_conntrack_ecache_fini();
2761 nf_conntrack_tstamp_fini();
2763 nf_conntrack_acct_fini();
2765 nf_conntrack_expect_fini();
2767 kmem_cache_destroy(nf_conntrack_cachep);
2769 kvfree(nf_conntrack_hash);
2773 static struct nf_ct_hook nf_conntrack_hook = {
2774 .update = nf_conntrack_update,
2775 .destroy = destroy_conntrack,
2776 .get_tuple_skb = nf_conntrack_get_tuple_skb,
2779 void nf_conntrack_init_end(void)
2781 /* For use by REJECT target */
2782 RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
2783 RCU_INIT_POINTER(nf_ct_hook, &nf_conntrack_hook);
2787 * We need to use special "null" values, not used in hash table
2789 #define UNCONFIRMED_NULLS_VAL ((1<<30)+0)
2790 #define DYING_NULLS_VAL ((1<<30)+1)
2792 int nf_conntrack_init_net(struct net *net)
2794 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2798 BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2799 BUILD_BUG_ON_NOT_POWER_OF_2(CONNTRACK_LOCKS);
2800 atomic_set(&cnet->count, 0);
2802 net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
2803 if (!net->ct.pcpu_lists)
2806 for_each_possible_cpu(cpu) {
2807 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2809 spin_lock_init(&pcpu->lock);
2810 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
2811 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
2814 net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2816 goto err_pcpu_lists;
2818 ret = nf_conntrack_expect_pernet_init(net);
2822 nf_conntrack_acct_pernet_init(net);
2823 nf_conntrack_tstamp_pernet_init(net);
2824 nf_conntrack_ecache_pernet_init(net);
2825 nf_conntrack_helper_pernet_init(net);
2826 nf_conntrack_proto_pernet_init(net);
2831 free_percpu(net->ct.stat);
2833 free_percpu(net->ct.pcpu_lists);