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_bpf.h>
38 #include <net/netfilter/nf_conntrack_l4proto.h>
39 #include <net/netfilter/nf_conntrack_expect.h>
40 #include <net/netfilter/nf_conntrack_helper.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;
76 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
77 static DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
78 static __read_mostly bool nf_conntrack_locks_all;
80 /* serialize hash resizes and nf_ct_iterate_cleanup */
81 static DEFINE_MUTEX(nf_conntrack_mutex);
83 #define GC_SCAN_INTERVAL_MAX (60ul * HZ)
84 #define GC_SCAN_INTERVAL_MIN (1ul * HZ)
86 /* clamp timeouts to this value (TCP unacked) */
87 #define GC_SCAN_INTERVAL_CLAMP (300ul * HZ)
89 /* Initial bias pretending we have 100 entries at the upper bound so we don't
90 * wakeup often just because we have three entries with a 1s timeout while still
91 * allowing non-idle machines to wakeup more often when needed.
93 #define GC_SCAN_INITIAL_COUNT 100
94 #define GC_SCAN_INTERVAL_INIT GC_SCAN_INTERVAL_MAX
96 #define GC_SCAN_MAX_DURATION msecs_to_jiffies(10)
97 #define GC_SCAN_EXPIRED_MAX (64000u / HZ)
99 #define MIN_CHAINLEN 8u
100 #define MAX_CHAINLEN (32u - MIN_CHAINLEN)
102 static struct conntrack_gc_work conntrack_gc_work;
104 void nf_conntrack_lock(spinlock_t *lock) __acquires(lock)
106 /* 1) Acquire the lock */
109 /* 2) read nf_conntrack_locks_all, with ACQUIRE semantics
110 * It pairs with the smp_store_release() in nf_conntrack_all_unlock()
112 if (likely(smp_load_acquire(&nf_conntrack_locks_all) == false))
115 /* fast path failed, unlock */
118 /* Slow path 1) get global lock */
119 spin_lock(&nf_conntrack_locks_all_lock);
121 /* Slow path 2) get the lock we want */
124 /* Slow path 3) release the global lock */
125 spin_unlock(&nf_conntrack_locks_all_lock);
127 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
129 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
131 h1 %= CONNTRACK_LOCKS;
132 h2 %= CONNTRACK_LOCKS;
133 spin_unlock(&nf_conntrack_locks[h1]);
135 spin_unlock(&nf_conntrack_locks[h2]);
138 /* return true if we need to recompute hashes (in case hash table was resized) */
139 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
140 unsigned int h2, unsigned int sequence)
142 h1 %= CONNTRACK_LOCKS;
143 h2 %= CONNTRACK_LOCKS;
145 nf_conntrack_lock(&nf_conntrack_locks[h1]);
147 spin_lock_nested(&nf_conntrack_locks[h2],
148 SINGLE_DEPTH_NESTING);
150 nf_conntrack_lock(&nf_conntrack_locks[h2]);
151 spin_lock_nested(&nf_conntrack_locks[h1],
152 SINGLE_DEPTH_NESTING);
154 if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
155 nf_conntrack_double_unlock(h1, h2);
161 static void nf_conntrack_all_lock(void)
162 __acquires(&nf_conntrack_locks_all_lock)
166 spin_lock(&nf_conntrack_locks_all_lock);
168 /* For nf_contrack_locks_all, only the latest time when another
169 * CPU will see an update is controlled, by the "release" of the
171 * The earliest time is not controlled, an thus KCSAN could detect
172 * a race when nf_conntract_lock() reads the variable.
173 * WRITE_ONCE() is used to ensure the compiler will not
174 * optimize the write.
176 WRITE_ONCE(nf_conntrack_locks_all, true);
178 for (i = 0; i < CONNTRACK_LOCKS; i++) {
179 spin_lock(&nf_conntrack_locks[i]);
181 /* This spin_unlock provides the "release" to ensure that
182 * nf_conntrack_locks_all==true is visible to everyone that
183 * acquired spin_lock(&nf_conntrack_locks[]).
185 spin_unlock(&nf_conntrack_locks[i]);
189 static void nf_conntrack_all_unlock(void)
190 __releases(&nf_conntrack_locks_all_lock)
192 /* All prior stores must be complete before we clear
193 * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
194 * might observe the false value but not the entire
196 * It pairs with the smp_load_acquire() in nf_conntrack_lock()
198 smp_store_release(&nf_conntrack_locks_all, false);
199 spin_unlock(&nf_conntrack_locks_all_lock);
202 unsigned int nf_conntrack_htable_size __read_mostly;
203 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
205 unsigned int nf_conntrack_max __read_mostly;
206 EXPORT_SYMBOL_GPL(nf_conntrack_max);
207 seqcount_spinlock_t nf_conntrack_generation __read_mostly;
208 static siphash_aligned_key_t nf_conntrack_hash_rnd;
210 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
212 const struct net *net)
216 get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
218 /* The direction must be ignored, handle usable tuplehash members manually */
219 a = (u64)tuple->src.u3.all[0] << 32 | tuple->src.u3.all[3];
220 b = (u64)tuple->dst.u3.all[0] << 32 | tuple->dst.u3.all[3];
222 c = (__force u64)tuple->src.u.all << 32 | (__force u64)tuple->dst.u.all << 16;
223 c |= tuple->dst.protonum;
225 d = (u64)zoneid << 32 | net_hash_mix(net);
227 /* IPv4: u3.all[1,2,3] == 0 */
228 c ^= (u64)tuple->src.u3.all[1] << 32 | tuple->src.u3.all[2];
229 d += (u64)tuple->dst.u3.all[1] << 32 | tuple->dst.u3.all[2];
231 return (u32)siphash_4u64(a, b, c, d, &nf_conntrack_hash_rnd);
234 static u32 scale_hash(u32 hash)
236 return reciprocal_scale(hash, nf_conntrack_htable_size);
239 static u32 __hash_conntrack(const struct net *net,
240 const struct nf_conntrack_tuple *tuple,
244 return reciprocal_scale(hash_conntrack_raw(tuple, zoneid, net), size);
247 static u32 hash_conntrack(const struct net *net,
248 const struct nf_conntrack_tuple *tuple,
251 return scale_hash(hash_conntrack_raw(tuple, zoneid, net));
254 static bool nf_ct_get_tuple_ports(const struct sk_buff *skb,
255 unsigned int dataoff,
256 struct nf_conntrack_tuple *tuple)
260 } _inet_hdr, *inet_hdr;
262 /* Actually only need first 4 bytes to get ports. */
263 inet_hdr = skb_header_pointer(skb, dataoff, sizeof(_inet_hdr), &_inet_hdr);
267 tuple->src.u.udp.port = inet_hdr->sport;
268 tuple->dst.u.udp.port = inet_hdr->dport;
273 nf_ct_get_tuple(const struct sk_buff *skb,
275 unsigned int dataoff,
279 struct nf_conntrack_tuple *tuple)
285 memset(tuple, 0, sizeof(*tuple));
287 tuple->src.l3num = l3num;
290 nhoff += offsetof(struct iphdr, saddr);
291 size = 2 * sizeof(__be32);
294 nhoff += offsetof(struct ipv6hdr, saddr);
295 size = sizeof(_addrs);
301 ap = skb_header_pointer(skb, nhoff, size, _addrs);
307 tuple->src.u3.ip = ap[0];
308 tuple->dst.u3.ip = ap[1];
311 memcpy(tuple->src.u3.ip6, ap, sizeof(tuple->src.u3.ip6));
312 memcpy(tuple->dst.u3.ip6, ap + 4, sizeof(tuple->dst.u3.ip6));
316 tuple->dst.protonum = protonum;
317 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
320 #if IS_ENABLED(CONFIG_IPV6)
322 return icmpv6_pkt_to_tuple(skb, dataoff, net, tuple);
325 return icmp_pkt_to_tuple(skb, dataoff, net, tuple);
326 #ifdef CONFIG_NF_CT_PROTO_GRE
328 return gre_pkt_to_tuple(skb, dataoff, net, tuple);
332 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
333 case IPPROTO_UDPLITE:
335 #ifdef CONFIG_NF_CT_PROTO_SCTP
338 #ifdef CONFIG_NF_CT_PROTO_DCCP
342 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
350 static int ipv4_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
354 const struct iphdr *iph;
357 iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
361 /* Conntrack defragments packets, we might still see fragments
362 * inside ICMP packets though.
364 if (iph->frag_off & htons(IP_OFFSET))
367 dataoff = nhoff + (iph->ihl << 2);
368 *protonum = iph->protocol;
370 /* Check bogus IP headers */
371 if (dataoff > skb->len) {
372 pr_debug("bogus IPv4 packet: nhoff %u, ihl %u, skblen %u\n",
373 nhoff, iph->ihl << 2, skb->len);
379 #if IS_ENABLED(CONFIG_IPV6)
380 static int ipv6_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
384 unsigned int extoff = nhoff + sizeof(struct ipv6hdr);
388 if (skb_copy_bits(skb, nhoff + offsetof(struct ipv6hdr, nexthdr),
389 &nexthdr, sizeof(nexthdr)) != 0) {
390 pr_debug("can't get nexthdr\n");
393 protoff = ipv6_skip_exthdr(skb, extoff, &nexthdr, &frag_off);
395 * (protoff == skb->len) means the packet has not data, just
396 * IPv6 and possibly extensions headers, but it is tracked anyway
398 if (protoff < 0 || (frag_off & htons(~0x7)) != 0) {
399 pr_debug("can't find proto in pkt\n");
408 static int get_l4proto(const struct sk_buff *skb,
409 unsigned int nhoff, u8 pf, u8 *l4num)
413 return ipv4_get_l4proto(skb, nhoff, l4num);
414 #if IS_ENABLED(CONFIG_IPV6)
416 return ipv6_get_l4proto(skb, nhoff, l4num);
425 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
427 struct net *net, struct nf_conntrack_tuple *tuple)
432 protoff = get_l4proto(skb, nhoff, l3num, &protonum);
436 return nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple);
438 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
441 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
442 const struct nf_conntrack_tuple *orig)
444 memset(inverse, 0, sizeof(*inverse));
446 inverse->src.l3num = orig->src.l3num;
448 switch (orig->src.l3num) {
450 inverse->src.u3.ip = orig->dst.u3.ip;
451 inverse->dst.u3.ip = orig->src.u3.ip;
454 inverse->src.u3.in6 = orig->dst.u3.in6;
455 inverse->dst.u3.in6 = orig->src.u3.in6;
461 inverse->dst.dir = !orig->dst.dir;
463 inverse->dst.protonum = orig->dst.protonum;
465 switch (orig->dst.protonum) {
467 return nf_conntrack_invert_icmp_tuple(inverse, orig);
468 #if IS_ENABLED(CONFIG_IPV6)
470 return nf_conntrack_invert_icmpv6_tuple(inverse, orig);
474 inverse->src.u.all = orig->dst.u.all;
475 inverse->dst.u.all = orig->src.u.all;
478 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
480 /* Generate a almost-unique pseudo-id for a given conntrack.
482 * intentionally doesn't re-use any of the seeds used for hash
483 * table location, we assume id gets exposed to userspace.
485 * Following nf_conn items do not change throughout lifetime
489 * 2. nf_conn->master address (normally NULL)
490 * 3. the associated net namespace
491 * 4. the original direction tuple
493 u32 nf_ct_get_id(const struct nf_conn *ct)
495 static siphash_aligned_key_t ct_id_seed;
496 unsigned long a, b, c, d;
498 net_get_random_once(&ct_id_seed, sizeof(ct_id_seed));
500 a = (unsigned long)ct;
501 b = (unsigned long)ct->master;
502 c = (unsigned long)nf_ct_net(ct);
503 d = (unsigned long)siphash(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
504 sizeof(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple),
507 return siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &ct_id_seed);
509 return siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &ct_id_seed);
512 EXPORT_SYMBOL_GPL(nf_ct_get_id);
515 clean_from_lists(struct nf_conn *ct)
517 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
518 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
520 /* Destroy all pending expectations */
521 nf_ct_remove_expectations(ct);
524 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
526 /* Released via nf_ct_destroy() */
527 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
528 const struct nf_conntrack_zone *zone,
531 struct nf_conn *tmpl, *p;
533 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
534 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
539 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
541 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
542 tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
545 tmpl = kzalloc(sizeof(*tmpl), flags);
550 tmpl->status = IPS_TEMPLATE;
551 write_pnet(&tmpl->ct_net, net);
552 nf_ct_zone_add(tmpl, zone);
553 refcount_set(&tmpl->ct_general.use, 1);
557 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
559 void nf_ct_tmpl_free(struct nf_conn *tmpl)
563 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
564 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
568 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
570 static void destroy_gre_conntrack(struct nf_conn *ct)
572 #ifdef CONFIG_NF_CT_PROTO_GRE
573 struct nf_conn *master = ct->master;
576 nf_ct_gre_keymap_destroy(master);
580 void nf_ct_destroy(struct nf_conntrack *nfct)
582 struct nf_conn *ct = (struct nf_conn *)nfct;
584 WARN_ON(refcount_read(&nfct->use) != 0);
586 if (unlikely(nf_ct_is_template(ct))) {
591 if (unlikely(nf_ct_protonum(ct) == IPPROTO_GRE))
592 destroy_gre_conntrack(ct);
594 /* Expectations will have been removed in clean_from_lists,
595 * except TFTP can create an expectation on the first packet,
596 * before connection is in the list, so we need to clean here,
599 nf_ct_remove_expectations(ct);
602 nf_ct_put(ct->master);
604 nf_conntrack_free(ct);
606 EXPORT_SYMBOL(nf_ct_destroy);
608 static void __nf_ct_delete_from_lists(struct nf_conn *ct)
610 struct net *net = nf_ct_net(ct);
611 unsigned int hash, reply_hash;
612 unsigned int sequence;
615 sequence = read_seqcount_begin(&nf_conntrack_generation);
616 hash = hash_conntrack(net,
617 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
618 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_ORIGINAL));
619 reply_hash = hash_conntrack(net,
620 &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
621 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
622 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
624 clean_from_lists(ct);
625 nf_conntrack_double_unlock(hash, reply_hash);
628 static void nf_ct_delete_from_lists(struct nf_conn *ct)
630 nf_ct_helper_destroy(ct);
633 __nf_ct_delete_from_lists(ct);
638 static void nf_ct_add_to_ecache_list(struct nf_conn *ct)
640 #ifdef CONFIG_NF_CONNTRACK_EVENTS
641 struct nf_conntrack_net *cnet = nf_ct_pernet(nf_ct_net(ct));
643 spin_lock(&cnet->ecache.dying_lock);
644 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
645 &cnet->ecache.dying_list);
646 spin_unlock(&cnet->ecache.dying_lock);
650 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
652 struct nf_conn_tstamp *tstamp;
655 if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
658 tstamp = nf_conn_tstamp_find(ct);
660 s32 timeout = READ_ONCE(ct->timeout) - nfct_time_stamp;
662 tstamp->stop = ktime_get_real_ns();
664 tstamp->stop -= jiffies_to_nsecs(-timeout);
667 if (nf_conntrack_event_report(IPCT_DESTROY, ct,
668 portid, report) < 0) {
669 /* destroy event was not delivered. nf_ct_put will
670 * be done by event cache worker on redelivery.
672 nf_ct_helper_destroy(ct);
674 __nf_ct_delete_from_lists(ct);
675 nf_ct_add_to_ecache_list(ct);
678 nf_conntrack_ecache_work(nf_ct_net(ct), NFCT_ECACHE_DESTROY_FAIL);
683 if (nf_conntrack_ecache_dwork_pending(net))
684 nf_conntrack_ecache_work(net, NFCT_ECACHE_DESTROY_SENT);
685 nf_ct_delete_from_lists(ct);
689 EXPORT_SYMBOL_GPL(nf_ct_delete);
692 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
693 const struct nf_conntrack_tuple *tuple,
694 const struct nf_conntrack_zone *zone,
695 const struct net *net)
697 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
699 /* A conntrack can be recreated with the equal tuple,
700 * so we need to check that the conntrack is confirmed
702 return nf_ct_tuple_equal(tuple, &h->tuple) &&
703 nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
704 nf_ct_is_confirmed(ct) &&
705 net_eq(net, nf_ct_net(ct));
709 nf_ct_match(const struct nf_conn *ct1, const struct nf_conn *ct2)
711 return nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
712 &ct2->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
713 nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_REPLY].tuple,
714 &ct2->tuplehash[IP_CT_DIR_REPLY].tuple) &&
715 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_ORIGINAL) &&
716 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_REPLY) &&
717 net_eq(nf_ct_net(ct1), nf_ct_net(ct2));
720 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
721 static void nf_ct_gc_expired(struct nf_conn *ct)
723 if (!refcount_inc_not_zero(&ct->ct_general.use))
726 /* load ->status after refcount increase */
727 smp_acquire__after_ctrl_dep();
729 if (nf_ct_should_gc(ct))
737 * - Caller must take a reference on returned object
738 * and recheck nf_ct_tuple_equal(tuple, &h->tuple)
740 static struct nf_conntrack_tuple_hash *
741 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
742 const struct nf_conntrack_tuple *tuple, u32 hash)
744 struct nf_conntrack_tuple_hash *h;
745 struct hlist_nulls_head *ct_hash;
746 struct hlist_nulls_node *n;
747 unsigned int bucket, hsize;
750 nf_conntrack_get_ht(&ct_hash, &hsize);
751 bucket = reciprocal_scale(hash, hsize);
753 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
756 ct = nf_ct_tuplehash_to_ctrack(h);
757 if (nf_ct_is_expired(ct)) {
758 nf_ct_gc_expired(ct);
762 if (nf_ct_key_equal(h, tuple, zone, net))
766 * if the nulls value we got at the end of this lookup is
767 * not the expected one, we must restart lookup.
768 * We probably met an item that was moved to another chain.
770 if (get_nulls_value(n) != bucket) {
771 NF_CT_STAT_INC_ATOMIC(net, search_restart);
778 /* Find a connection corresponding to a tuple. */
779 static struct nf_conntrack_tuple_hash *
780 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
781 const struct nf_conntrack_tuple *tuple, u32 hash)
783 struct nf_conntrack_tuple_hash *h;
788 h = ____nf_conntrack_find(net, zone, tuple, hash);
790 /* We have a candidate that matches the tuple we're interested
791 * in, try to obtain a reference and re-check tuple
793 ct = nf_ct_tuplehash_to_ctrack(h);
794 if (likely(refcount_inc_not_zero(&ct->ct_general.use))) {
795 /* re-check key after refcount */
796 smp_acquire__after_ctrl_dep();
798 if (likely(nf_ct_key_equal(h, tuple, zone, net)))
801 /* TYPESAFE_BY_RCU recycled the candidate */
813 struct nf_conntrack_tuple_hash *
814 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
815 const struct nf_conntrack_tuple *tuple)
817 unsigned int rid, zone_id = nf_ct_zone_id(zone, IP_CT_DIR_ORIGINAL);
818 struct nf_conntrack_tuple_hash *thash;
820 thash = __nf_conntrack_find_get(net, zone, tuple,
821 hash_conntrack_raw(tuple, zone_id, net));
826 rid = nf_ct_zone_id(zone, IP_CT_DIR_REPLY);
828 return __nf_conntrack_find_get(net, zone, tuple,
829 hash_conntrack_raw(tuple, rid, net));
832 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
834 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
836 unsigned int reply_hash)
838 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
839 &nf_conntrack_hash[hash]);
840 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
841 &nf_conntrack_hash[reply_hash]);
844 static bool nf_ct_ext_valid_pre(const struct nf_ct_ext *ext)
846 /* if ext->gen_id is not equal to nf_conntrack_ext_genid, some extensions
847 * may contain stale pointers to e.g. helper that has been removed.
849 * The helper can't clear this because the nf_conn object isn't in
850 * any hash and synchronize_rcu() isn't enough because associated skb
851 * might sit in a queue.
853 return !ext || ext->gen_id == atomic_read(&nf_conntrack_ext_genid);
856 static bool nf_ct_ext_valid_post(struct nf_ct_ext *ext)
861 if (ext->gen_id != atomic_read(&nf_conntrack_ext_genid))
864 /* inserted into conntrack table, nf_ct_iterate_cleanup()
865 * will find it. Disable nf_ct_ext_find() id check.
867 WRITE_ONCE(ext->gen_id, 0);
872 nf_conntrack_hash_check_insert(struct nf_conn *ct)
874 const struct nf_conntrack_zone *zone;
875 struct net *net = nf_ct_net(ct);
876 unsigned int hash, reply_hash;
877 struct nf_conntrack_tuple_hash *h;
878 struct hlist_nulls_node *n;
879 unsigned int max_chainlen;
880 unsigned int chainlen = 0;
881 unsigned int sequence;
884 zone = nf_ct_zone(ct);
886 if (!nf_ct_ext_valid_pre(ct->ext)) {
887 NF_CT_STAT_INC_ATOMIC(net, insert_failed);
893 sequence = read_seqcount_begin(&nf_conntrack_generation);
894 hash = hash_conntrack(net,
895 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
896 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_ORIGINAL));
897 reply_hash = hash_conntrack(net,
898 &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
899 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
900 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
902 max_chainlen = MIN_CHAINLEN + get_random_u32_below(MAX_CHAINLEN);
904 /* See if there's one in the list already, including reverse */
905 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode) {
906 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
910 if (chainlen++ > max_chainlen)
916 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode) {
917 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
920 if (chainlen++ > max_chainlen)
925 /* The caller holds a reference to this object */
926 refcount_set(&ct->ct_general.use, 2);
927 __nf_conntrack_hash_insert(ct, hash, reply_hash);
928 nf_conntrack_double_unlock(hash, reply_hash);
929 NF_CT_STAT_INC(net, insert);
932 if (!nf_ct_ext_valid_post(ct->ext)) {
934 NF_CT_STAT_INC_ATOMIC(net, drop);
940 NF_CT_STAT_INC(net, chaintoolong);
943 nf_conntrack_double_unlock(hash, reply_hash);
947 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
949 void nf_ct_acct_add(struct nf_conn *ct, u32 dir, unsigned int packets,
952 struct nf_conn_acct *acct;
954 acct = nf_conn_acct_find(ct);
956 struct nf_conn_counter *counter = acct->counter;
958 atomic64_add(packets, &counter[dir].packets);
959 atomic64_add(bytes, &counter[dir].bytes);
962 EXPORT_SYMBOL_GPL(nf_ct_acct_add);
964 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
965 const struct nf_conn *loser_ct)
967 struct nf_conn_acct *acct;
969 acct = nf_conn_acct_find(loser_ct);
971 struct nf_conn_counter *counter = acct->counter;
974 /* u32 should be fine since we must have seen one packet. */
975 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
976 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), bytes);
980 static void __nf_conntrack_insert_prepare(struct nf_conn *ct)
982 struct nf_conn_tstamp *tstamp;
984 refcount_inc(&ct->ct_general.use);
986 /* set conntrack timestamp, if enabled. */
987 tstamp = nf_conn_tstamp_find(ct);
989 tstamp->start = ktime_get_real_ns();
992 /* caller must hold locks to prevent concurrent changes */
993 static int __nf_ct_resolve_clash(struct sk_buff *skb,
994 struct nf_conntrack_tuple_hash *h)
996 /* This is the conntrack entry already in hashes that won race. */
997 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
998 enum ip_conntrack_info ctinfo;
999 struct nf_conn *loser_ct;
1001 loser_ct = nf_ct_get(skb, &ctinfo);
1003 if (nf_ct_is_dying(ct))
1006 if (((ct->status & IPS_NAT_DONE_MASK) == 0) ||
1007 nf_ct_match(ct, loser_ct)) {
1008 struct net *net = nf_ct_net(ct);
1010 nf_conntrack_get(&ct->ct_general);
1012 nf_ct_acct_merge(ct, ctinfo, loser_ct);
1013 nf_ct_put(loser_ct);
1014 nf_ct_set(skb, ct, ctinfo);
1016 NF_CT_STAT_INC(net, clash_resolve);
1024 * nf_ct_resolve_clash_harder - attempt to insert clashing conntrack entry
1026 * @skb: skb that causes the collision
1027 * @repl_idx: hash slot for reply direction
1029 * Called when origin or reply direction had a clash.
1030 * The skb can be handled without packet drop provided the reply direction
1031 * is unique or there the existing entry has the identical tuple in both
1034 * Caller must hold conntrack table locks to prevent concurrent updates.
1036 * Returns NF_DROP if the clash could not be handled.
1038 static int nf_ct_resolve_clash_harder(struct sk_buff *skb, u32 repl_idx)
1040 struct nf_conn *loser_ct = (struct nf_conn *)skb_nfct(skb);
1041 const struct nf_conntrack_zone *zone;
1042 struct nf_conntrack_tuple_hash *h;
1043 struct hlist_nulls_node *n;
1046 zone = nf_ct_zone(loser_ct);
1047 net = nf_ct_net(loser_ct);
1049 /* Reply direction must never result in a clash, unless both origin
1050 * and reply tuples are identical.
1052 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[repl_idx], hnnode) {
1053 if (nf_ct_key_equal(h,
1054 &loser_ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1056 return __nf_ct_resolve_clash(skb, h);
1059 /* We want the clashing entry to go away real soon: 1 second timeout. */
1060 WRITE_ONCE(loser_ct->timeout, nfct_time_stamp + HZ);
1062 /* IPS_NAT_CLASH removes the entry automatically on the first
1063 * reply. Also prevents UDP tracker from moving the entry to
1064 * ASSURED state, i.e. the entry can always be evicted under
1067 loser_ct->status |= IPS_FIXED_TIMEOUT | IPS_NAT_CLASH;
1069 __nf_conntrack_insert_prepare(loser_ct);
1071 /* fake add for ORIGINAL dir: we want lookups to only find the entry
1072 * already in the table. This also hides the clashing entry from
1073 * ctnetlink iteration, i.e. conntrack -L won't show them.
1075 hlist_nulls_add_fake(&loser_ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
1077 hlist_nulls_add_head_rcu(&loser_ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
1078 &nf_conntrack_hash[repl_idx]);
1080 NF_CT_STAT_INC(net, clash_resolve);
1085 * nf_ct_resolve_clash - attempt to handle clash without packet drop
1087 * @skb: skb that causes the clash
1088 * @h: tuplehash of the clashing entry already in table
1089 * @reply_hash: hash slot for reply direction
1091 * A conntrack entry can be inserted to the connection tracking table
1092 * if there is no existing entry with an identical tuple.
1094 * If there is one, @skb (and the assocated, unconfirmed conntrack) has
1095 * to be dropped. In case @skb is retransmitted, next conntrack lookup
1096 * will find the already-existing entry.
1098 * The major problem with such packet drop is the extra delay added by
1099 * the packet loss -- it will take some time for a retransmit to occur
1100 * (or the sender to time out when waiting for a reply).
1102 * This function attempts to handle the situation without packet drop.
1104 * If @skb has no NAT transformation or if the colliding entries are
1105 * exactly the same, only the to-be-confirmed conntrack entry is discarded
1106 * and @skb is associated with the conntrack entry already in the table.
1108 * Failing that, the new, unconfirmed conntrack is still added to the table
1109 * provided that the collision only occurs in the ORIGINAL direction.
1110 * The new entry will be added only in the non-clashing REPLY direction,
1111 * so packets in the ORIGINAL direction will continue to match the existing
1112 * entry. The new entry will also have a fixed timeout so it expires --
1113 * due to the collision, it will only see reply traffic.
1115 * Returns NF_DROP if the clash could not be resolved.
1117 static __cold noinline int
1118 nf_ct_resolve_clash(struct sk_buff *skb, struct nf_conntrack_tuple_hash *h,
1121 /* This is the conntrack entry already in hashes that won race. */
1122 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1123 const struct nf_conntrack_l4proto *l4proto;
1124 enum ip_conntrack_info ctinfo;
1125 struct nf_conn *loser_ct;
1129 loser_ct = nf_ct_get(skb, &ctinfo);
1130 net = nf_ct_net(loser_ct);
1132 l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1133 if (!l4proto->allow_clash)
1136 ret = __nf_ct_resolve_clash(skb, h);
1137 if (ret == NF_ACCEPT)
1140 ret = nf_ct_resolve_clash_harder(skb, reply_hash);
1141 if (ret == NF_ACCEPT)
1145 NF_CT_STAT_INC(net, drop);
1146 NF_CT_STAT_INC(net, insert_failed);
1150 /* Confirm a connection given skb; places it in hash table */
1152 __nf_conntrack_confirm(struct sk_buff *skb)
1154 unsigned int chainlen = 0, sequence, max_chainlen;
1155 const struct nf_conntrack_zone *zone;
1156 unsigned int hash, reply_hash;
1157 struct nf_conntrack_tuple_hash *h;
1159 struct nf_conn_help *help;
1160 struct hlist_nulls_node *n;
1161 enum ip_conntrack_info ctinfo;
1165 ct = nf_ct_get(skb, &ctinfo);
1166 net = nf_ct_net(ct);
1168 /* ipt_REJECT uses nf_conntrack_attach to attach related
1169 ICMP/TCP RST packets in other direction. Actual packet
1170 which created connection will be IP_CT_NEW or for an
1171 expected connection, IP_CT_RELATED. */
1172 if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
1175 zone = nf_ct_zone(ct);
1179 sequence = read_seqcount_begin(&nf_conntrack_generation);
1180 /* reuse the hash saved before */
1181 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
1182 hash = scale_hash(hash);
1183 reply_hash = hash_conntrack(net,
1184 &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1185 nf_ct_zone_id(nf_ct_zone(ct), IP_CT_DIR_REPLY));
1186 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
1188 /* We're not in hash table, and we refuse to set up related
1189 * connections for unconfirmed conns. But packet copies and
1190 * REJECT will give spurious warnings here.
1193 /* Another skb with the same unconfirmed conntrack may
1194 * win the race. This may happen for bridge(br_flood)
1195 * or broadcast/multicast packets do skb_clone with
1196 * unconfirmed conntrack.
1198 if (unlikely(nf_ct_is_confirmed(ct))) {
1200 nf_conntrack_double_unlock(hash, reply_hash);
1205 if (!nf_ct_ext_valid_pre(ct->ext)) {
1206 NF_CT_STAT_INC(net, insert_failed);
1210 /* We have to check the DYING flag after unlink to prevent
1211 * a race against nf_ct_get_next_corpse() possibly called from
1212 * user context, else we insert an already 'dead' hash, blocking
1213 * further use of that particular connection -JM.
1215 ct->status |= IPS_CONFIRMED;
1217 if (unlikely(nf_ct_is_dying(ct))) {
1218 NF_CT_STAT_INC(net, insert_failed);
1222 max_chainlen = MIN_CHAINLEN + get_random_u32_below(MAX_CHAINLEN);
1223 /* See if there's one in the list already, including reverse:
1224 NAT could have grabbed it without realizing, since we're
1225 not in the hash. If there is, we lost race. */
1226 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode) {
1227 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1230 if (chainlen++ > max_chainlen)
1235 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode) {
1236 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1239 if (chainlen++ > max_chainlen) {
1241 NF_CT_STAT_INC(net, chaintoolong);
1242 NF_CT_STAT_INC(net, insert_failed);
1248 /* Timer relative to confirmation time, not original
1249 setting time, otherwise we'd get timer wrap in
1250 weird delay cases. */
1251 ct->timeout += nfct_time_stamp;
1253 __nf_conntrack_insert_prepare(ct);
1255 /* Since the lookup is lockless, hash insertion must be done after
1256 * starting the timer and setting the CONFIRMED bit. The RCU barriers
1257 * guarantee that no other CPU can find the conntrack before the above
1258 * stores are visible.
1260 __nf_conntrack_hash_insert(ct, hash, reply_hash);
1261 nf_conntrack_double_unlock(hash, reply_hash);
1264 /* ext area is still valid (rcu read lock is held,
1265 * but will go out of scope soon, we need to remove
1266 * this conntrack again.
1268 if (!nf_ct_ext_valid_post(ct->ext)) {
1270 NF_CT_STAT_INC_ATOMIC(net, drop);
1274 help = nfct_help(ct);
1275 if (help && help->helper)
1276 nf_conntrack_event_cache(IPCT_HELPER, ct);
1278 nf_conntrack_event_cache(master_ct(ct) ?
1279 IPCT_RELATED : IPCT_NEW, ct);
1283 ret = nf_ct_resolve_clash(skb, h, reply_hash);
1285 nf_conntrack_double_unlock(hash, reply_hash);
1289 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
1291 /* Returns true if a connection correspondings to the tuple (required
1294 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
1295 const struct nf_conn *ignored_conntrack)
1297 struct net *net = nf_ct_net(ignored_conntrack);
1298 const struct nf_conntrack_zone *zone;
1299 struct nf_conntrack_tuple_hash *h;
1300 struct hlist_nulls_head *ct_hash;
1301 unsigned int hash, hsize;
1302 struct hlist_nulls_node *n;
1305 zone = nf_ct_zone(ignored_conntrack);
1309 nf_conntrack_get_ht(&ct_hash, &hsize);
1310 hash = __hash_conntrack(net, tuple, nf_ct_zone_id(zone, IP_CT_DIR_REPLY), hsize);
1312 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
1313 ct = nf_ct_tuplehash_to_ctrack(h);
1315 if (ct == ignored_conntrack)
1318 if (nf_ct_is_expired(ct)) {
1319 nf_ct_gc_expired(ct);
1323 if (nf_ct_key_equal(h, tuple, zone, net)) {
1324 /* Tuple is taken already, so caller will need to find
1325 * a new source port to use.
1328 * If the *original tuples* are identical, then both
1329 * conntracks refer to the same flow.
1330 * This is a rare situation, it can occur e.g. when
1331 * more than one UDP packet is sent from same socket
1332 * in different threads.
1334 * Let nf_ct_resolve_clash() deal with this later.
1336 if (nf_ct_tuple_equal(&ignored_conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1337 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
1338 nf_ct_zone_equal(ct, zone, IP_CT_DIR_ORIGINAL))
1341 NF_CT_STAT_INC_ATOMIC(net, found);
1347 if (get_nulls_value(n) != hash) {
1348 NF_CT_STAT_INC_ATOMIC(net, search_restart);
1356 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
1358 #define NF_CT_EVICTION_RANGE 8
1360 /* There's a small race here where we may free a just-assured
1361 connection. Too bad: we're in trouble anyway. */
1362 static unsigned int early_drop_list(struct net *net,
1363 struct hlist_nulls_head *head)
1365 struct nf_conntrack_tuple_hash *h;
1366 struct hlist_nulls_node *n;
1367 unsigned int drops = 0;
1368 struct nf_conn *tmp;
1370 hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
1371 tmp = nf_ct_tuplehash_to_ctrack(h);
1373 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status))
1376 if (nf_ct_is_expired(tmp)) {
1377 nf_ct_gc_expired(tmp);
1381 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
1382 !net_eq(nf_ct_net(tmp), net) ||
1383 nf_ct_is_dying(tmp))
1386 if (!refcount_inc_not_zero(&tmp->ct_general.use))
1389 /* load ->ct_net and ->status after refcount increase */
1390 smp_acquire__after_ctrl_dep();
1392 /* kill only if still in same netns -- might have moved due to
1393 * SLAB_TYPESAFE_BY_RCU rules.
1395 * We steal the timer reference. If that fails timer has
1396 * already fired or someone else deleted it. Just drop ref
1397 * and move to next entry.
1399 if (net_eq(nf_ct_net(tmp), net) &&
1400 nf_ct_is_confirmed(tmp) &&
1401 nf_ct_delete(tmp, 0, 0))
1410 static noinline int early_drop(struct net *net, unsigned int hash)
1412 unsigned int i, bucket;
1414 for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
1415 struct hlist_nulls_head *ct_hash;
1416 unsigned int hsize, drops;
1419 nf_conntrack_get_ht(&ct_hash, &hsize);
1421 bucket = reciprocal_scale(hash, hsize);
1423 bucket = (bucket + 1) % hsize;
1425 drops = early_drop_list(net, &ct_hash[bucket]);
1429 NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
1437 static bool gc_worker_skip_ct(const struct nf_conn *ct)
1439 return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
1442 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
1444 const struct nf_conntrack_l4proto *l4proto;
1446 if (!test_bit(IPS_ASSURED_BIT, &ct->status))
1449 l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1450 if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
1456 static void gc_worker(struct work_struct *work)
1458 unsigned int i, hashsz, nf_conntrack_max95 = 0;
1459 u32 end_time, start_time = nfct_time_stamp;
1460 struct conntrack_gc_work *gc_work;
1461 unsigned int expired_count = 0;
1462 unsigned long next_run;
1466 gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
1468 i = gc_work->next_bucket;
1469 if (gc_work->early_drop)
1470 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
1473 gc_work->avg_timeout = GC_SCAN_INTERVAL_INIT;
1474 gc_work->count = GC_SCAN_INITIAL_COUNT;
1475 gc_work->start_time = start_time;
1478 next_run = gc_work->avg_timeout;
1479 count = gc_work->count;
1481 end_time = start_time + GC_SCAN_MAX_DURATION;
1484 struct nf_conntrack_tuple_hash *h;
1485 struct hlist_nulls_head *ct_hash;
1486 struct hlist_nulls_node *n;
1487 struct nf_conn *tmp;
1491 nf_conntrack_get_ht(&ct_hash, &hashsz);
1497 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1498 struct nf_conntrack_net *cnet;
1502 tmp = nf_ct_tuplehash_to_ctrack(h);
1504 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status)) {
1505 nf_ct_offload_timeout(tmp);
1509 if (expired_count > GC_SCAN_EXPIRED_MAX) {
1512 gc_work->next_bucket = i;
1513 gc_work->avg_timeout = next_run;
1514 gc_work->count = count;
1516 delta_time = nfct_time_stamp - gc_work->start_time;
1518 /* re-sched immediately if total cycle time is exceeded */
1519 next_run = delta_time < (s32)GC_SCAN_INTERVAL_MAX;
1523 if (nf_ct_is_expired(tmp)) {
1524 nf_ct_gc_expired(tmp);
1529 expires = clamp(nf_ct_expires(tmp), GC_SCAN_INTERVAL_MIN, GC_SCAN_INTERVAL_CLAMP);
1530 expires = (expires - (long)next_run) / ++count;
1531 next_run += expires;
1533 if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1536 net = nf_ct_net(tmp);
1537 cnet = nf_ct_pernet(net);
1538 if (atomic_read(&cnet->count) < nf_conntrack_max95)
1541 /* need to take reference to avoid possible races */
1542 if (!refcount_inc_not_zero(&tmp->ct_general.use))
1545 /* load ->status after refcount increase */
1546 smp_acquire__after_ctrl_dep();
1548 if (gc_worker_skip_ct(tmp)) {
1553 if (gc_worker_can_early_drop(tmp)) {
1561 /* could check get_nulls_value() here and restart if ct
1562 * was moved to another chain. But given gc is best-effort
1563 * we will just continue with next hash slot.
1569 delta_time = nfct_time_stamp - end_time;
1570 if (delta_time > 0 && i < hashsz) {
1571 gc_work->avg_timeout = next_run;
1572 gc_work->count = count;
1573 gc_work->next_bucket = i;
1577 } while (i < hashsz);
1579 gc_work->next_bucket = 0;
1581 next_run = clamp(next_run, GC_SCAN_INTERVAL_MIN, GC_SCAN_INTERVAL_MAX);
1583 delta_time = max_t(s32, nfct_time_stamp - gc_work->start_time, 1);
1584 if (next_run > (unsigned long)delta_time)
1585 next_run -= delta_time;
1590 if (gc_work->exiting)
1594 gc_work->early_drop = false;
1596 queue_delayed_work(system_power_efficient_wq, &gc_work->dwork, next_run);
1599 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1601 INIT_DELAYED_WORK(&gc_work->dwork, gc_worker);
1602 gc_work->exiting = false;
1605 static struct nf_conn *
1606 __nf_conntrack_alloc(struct net *net,
1607 const struct nf_conntrack_zone *zone,
1608 const struct nf_conntrack_tuple *orig,
1609 const struct nf_conntrack_tuple *repl,
1610 gfp_t gfp, u32 hash)
1612 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
1613 unsigned int ct_count;
1616 /* We don't want any race condition at early drop stage */
1617 ct_count = atomic_inc_return(&cnet->count);
1619 if (nf_conntrack_max && unlikely(ct_count > nf_conntrack_max)) {
1620 if (!early_drop(net, hash)) {
1621 if (!conntrack_gc_work.early_drop)
1622 conntrack_gc_work.early_drop = true;
1623 atomic_dec(&cnet->count);
1624 net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1625 return ERR_PTR(-ENOMEM);
1630 * Do not use kmem_cache_zalloc(), as this cache uses
1631 * SLAB_TYPESAFE_BY_RCU.
1633 ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1637 spin_lock_init(&ct->lock);
1638 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1639 ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1640 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1641 /* save hash for reusing when confirming */
1642 *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1644 WRITE_ONCE(ct->timeout, 0);
1645 write_pnet(&ct->ct_net, net);
1646 memset_after(ct, 0, __nfct_init_offset);
1648 nf_ct_zone_add(ct, zone);
1650 /* Because we use RCU lookups, we set ct_general.use to zero before
1651 * this is inserted in any list.
1653 refcount_set(&ct->ct_general.use, 0);
1656 atomic_dec(&cnet->count);
1657 return ERR_PTR(-ENOMEM);
1660 struct nf_conn *nf_conntrack_alloc(struct net *net,
1661 const struct nf_conntrack_zone *zone,
1662 const struct nf_conntrack_tuple *orig,
1663 const struct nf_conntrack_tuple *repl,
1666 return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1668 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1670 void nf_conntrack_free(struct nf_conn *ct)
1672 struct net *net = nf_ct_net(ct);
1673 struct nf_conntrack_net *cnet;
1675 /* A freed object has refcnt == 0, that's
1676 * the golden rule for SLAB_TYPESAFE_BY_RCU
1678 WARN_ON(refcount_read(&ct->ct_general.use) != 0);
1680 if (ct->status & IPS_SRC_NAT_DONE) {
1681 const struct nf_nat_hook *nat_hook;
1684 nat_hook = rcu_dereference(nf_nat_hook);
1686 nat_hook->remove_nat_bysrc(ct);
1691 kmem_cache_free(nf_conntrack_cachep, ct);
1692 cnet = nf_ct_pernet(net);
1694 smp_mb__before_atomic();
1695 atomic_dec(&cnet->count);
1697 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1700 /* Allocate a new conntrack: we return -ENOMEM if classification
1701 failed due to stress. Otherwise it really is unclassifiable. */
1702 static noinline struct nf_conntrack_tuple_hash *
1703 init_conntrack(struct net *net, struct nf_conn *tmpl,
1704 const struct nf_conntrack_tuple *tuple,
1705 struct sk_buff *skb,
1706 unsigned int dataoff, u32 hash)
1709 struct nf_conn_help *help;
1710 struct nf_conntrack_tuple repl_tuple;
1711 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1712 struct nf_conntrack_ecache *ecache;
1714 struct nf_conntrack_expect *exp = NULL;
1715 const struct nf_conntrack_zone *zone;
1716 struct nf_conn_timeout *timeout_ext;
1717 struct nf_conntrack_zone tmp;
1718 struct nf_conntrack_net *cnet;
1720 if (!nf_ct_invert_tuple(&repl_tuple, tuple))
1723 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1724 ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1727 return (struct nf_conntrack_tuple_hash *)ct;
1729 if (!nf_ct_add_synproxy(ct, tmpl)) {
1730 nf_conntrack_free(ct);
1731 return ERR_PTR(-ENOMEM);
1734 timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1737 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1740 nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1741 nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1742 nf_ct_labels_ext_add(ct);
1744 #ifdef CONFIG_NF_CONNTRACK_EVENTS
1745 ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1747 if ((ecache || net->ct.sysctl_events) &&
1748 !nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1749 ecache ? ecache->expmask : 0,
1751 nf_conntrack_free(ct);
1752 return ERR_PTR(-ENOMEM);
1756 cnet = nf_ct_pernet(net);
1757 if (cnet->expect_count) {
1758 spin_lock_bh(&nf_conntrack_expect_lock);
1759 exp = nf_ct_find_expectation(net, zone, tuple);
1761 /* Welcome, Mr. Bond. We've been expecting you... */
1762 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1763 /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1764 ct->master = exp->master;
1766 help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1768 rcu_assign_pointer(help->helper, exp->helper);
1771 #ifdef CONFIG_NF_CONNTRACK_MARK
1772 ct->mark = READ_ONCE(exp->master->mark);
1774 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1775 ct->secmark = exp->master->secmark;
1777 NF_CT_STAT_INC(net, expect_new);
1779 spin_unlock_bh(&nf_conntrack_expect_lock);
1782 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1784 /* Other CPU might have obtained a pointer to this object before it was
1785 * released. Because refcount is 0, refcount_inc_not_zero() will fail.
1787 * After refcount_set(1) it will succeed; ensure that zeroing of
1788 * ct->status and the correct ct->net pointer are visible; else other
1789 * core might observe CONFIRMED bit which means the entry is valid and
1790 * in the hash table, but its not (anymore).
1794 /* Now it is going to be associated with an sk_buff, set refcount to 1. */
1795 refcount_set(&ct->ct_general.use, 1);
1799 exp->expectfn(ct, exp);
1800 nf_ct_expect_put(exp);
1803 return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1806 /* On success, returns 0, sets skb->_nfct | ctinfo */
1808 resolve_normal_ct(struct nf_conn *tmpl,
1809 struct sk_buff *skb,
1810 unsigned int dataoff,
1812 const struct nf_hook_state *state)
1814 const struct nf_conntrack_zone *zone;
1815 struct nf_conntrack_tuple tuple;
1816 struct nf_conntrack_tuple_hash *h;
1817 enum ip_conntrack_info ctinfo;
1818 struct nf_conntrack_zone tmp;
1819 u32 hash, zone_id, rid;
1822 if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1823 dataoff, state->pf, protonum, state->net,
1827 /* look for tuple match */
1828 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1830 zone_id = nf_ct_zone_id(zone, IP_CT_DIR_ORIGINAL);
1831 hash = hash_conntrack_raw(&tuple, zone_id, state->net);
1832 h = __nf_conntrack_find_get(state->net, zone, &tuple, hash);
1835 rid = nf_ct_zone_id(zone, IP_CT_DIR_REPLY);
1836 if (zone_id != rid) {
1837 u32 tmp = hash_conntrack_raw(&tuple, rid, state->net);
1839 h = __nf_conntrack_find_get(state->net, zone, &tuple, tmp);
1844 h = init_conntrack(state->net, tmpl, &tuple,
1845 skb, dataoff, hash);
1851 ct = nf_ct_tuplehash_to_ctrack(h);
1853 /* It exists; we have (non-exclusive) reference. */
1854 if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1855 ctinfo = IP_CT_ESTABLISHED_REPLY;
1857 unsigned long status = READ_ONCE(ct->status);
1859 /* Once we've had two way comms, always ESTABLISHED. */
1860 if (likely(status & IPS_SEEN_REPLY))
1861 ctinfo = IP_CT_ESTABLISHED;
1862 else if (status & IPS_EXPECTED)
1863 ctinfo = IP_CT_RELATED;
1867 nf_ct_set(skb, ct, ctinfo);
1872 * icmp packets need special treatment to handle error messages that are
1873 * related to a connection.
1875 * Callers need to check if skb has a conntrack assigned when this
1876 * helper returns; in such case skb belongs to an already known connection.
1878 static unsigned int __cold
1879 nf_conntrack_handle_icmp(struct nf_conn *tmpl,
1880 struct sk_buff *skb,
1881 unsigned int dataoff,
1883 const struct nf_hook_state *state)
1887 if (state->pf == NFPROTO_IPV4 && protonum == IPPROTO_ICMP)
1888 ret = nf_conntrack_icmpv4_error(tmpl, skb, dataoff, state);
1889 #if IS_ENABLED(CONFIG_IPV6)
1890 else if (state->pf == NFPROTO_IPV6 && protonum == IPPROTO_ICMPV6)
1891 ret = nf_conntrack_icmpv6_error(tmpl, skb, dataoff, state);
1897 NF_CT_STAT_INC_ATOMIC(state->net, error);
1902 static int generic_packet(struct nf_conn *ct, struct sk_buff *skb,
1903 enum ip_conntrack_info ctinfo)
1905 const unsigned int *timeout = nf_ct_timeout_lookup(ct);
1908 timeout = &nf_generic_pernet(nf_ct_net(ct))->timeout;
1910 nf_ct_refresh_acct(ct, ctinfo, skb, *timeout);
1914 /* Returns verdict for packet, or -1 for invalid. */
1915 static int nf_conntrack_handle_packet(struct nf_conn *ct,
1916 struct sk_buff *skb,
1917 unsigned int dataoff,
1918 enum ip_conntrack_info ctinfo,
1919 const struct nf_hook_state *state)
1921 switch (nf_ct_protonum(ct)) {
1923 return nf_conntrack_tcp_packet(ct, skb, dataoff,
1926 return nf_conntrack_udp_packet(ct, skb, dataoff,
1929 return nf_conntrack_icmp_packet(ct, skb, ctinfo, state);
1930 #if IS_ENABLED(CONFIG_IPV6)
1931 case IPPROTO_ICMPV6:
1932 return nf_conntrack_icmpv6_packet(ct, skb, ctinfo, state);
1934 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
1935 case IPPROTO_UDPLITE:
1936 return nf_conntrack_udplite_packet(ct, skb, dataoff,
1939 #ifdef CONFIG_NF_CT_PROTO_SCTP
1941 return nf_conntrack_sctp_packet(ct, skb, dataoff,
1944 #ifdef CONFIG_NF_CT_PROTO_DCCP
1946 return nf_conntrack_dccp_packet(ct, skb, dataoff,
1949 #ifdef CONFIG_NF_CT_PROTO_GRE
1951 return nf_conntrack_gre_packet(ct, skb, dataoff,
1956 return generic_packet(ct, skb, ctinfo);
1960 nf_conntrack_in(struct sk_buff *skb, const struct nf_hook_state *state)
1962 enum ip_conntrack_info ctinfo;
1963 struct nf_conn *ct, *tmpl;
1967 tmpl = nf_ct_get(skb, &ctinfo);
1968 if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1969 /* Previously seen (loopback or untracked)? Ignore. */
1970 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1971 ctinfo == IP_CT_UNTRACKED)
1976 /* rcu_read_lock()ed by nf_hook_thresh */
1977 dataoff = get_l4proto(skb, skb_network_offset(skb), state->pf, &protonum);
1979 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1984 if (protonum == IPPROTO_ICMP || protonum == IPPROTO_ICMPV6) {
1985 ret = nf_conntrack_handle_icmp(tmpl, skb, dataoff,
1991 /* ICMP[v6] protocol trackers may assign one conntrack. */
1996 ret = resolve_normal_ct(tmpl, skb, dataoff,
1999 /* Too stressed to deal. */
2000 NF_CT_STAT_INC_ATOMIC(state->net, drop);
2005 ct = nf_ct_get(skb, &ctinfo);
2007 /* Not valid part of a connection */
2008 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
2013 ret = nf_conntrack_handle_packet(ct, skb, dataoff, ctinfo, state);
2015 /* Invalid: inverse of the return code tells
2016 * the netfilter core what to do */
2019 /* Special case: TCP tracker reports an attempt to reopen a
2020 * closed/aborted connection. We have to go back and create a
2023 if (ret == -NF_REPEAT)
2026 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
2027 if (ret == -NF_DROP)
2028 NF_CT_STAT_INC_ATOMIC(state->net, drop);
2034 if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
2035 !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
2036 nf_conntrack_event_cache(IPCT_REPLY, ct);
2043 EXPORT_SYMBOL_GPL(nf_conntrack_in);
2045 /* Alter reply tuple (maybe alter helper). This is for NAT, and is
2046 implicitly racy: see __nf_conntrack_confirm */
2047 void nf_conntrack_alter_reply(struct nf_conn *ct,
2048 const struct nf_conntrack_tuple *newreply)
2050 struct nf_conn_help *help = nfct_help(ct);
2052 /* Should be unconfirmed, so not in hash table yet */
2053 WARN_ON(nf_ct_is_confirmed(ct));
2055 nf_ct_dump_tuple(newreply);
2057 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
2058 if (ct->master || (help && !hlist_empty(&help->expectations)))
2061 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
2063 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
2064 void __nf_ct_refresh_acct(struct nf_conn *ct,
2065 enum ip_conntrack_info ctinfo,
2066 const struct sk_buff *skb,
2070 /* Only update if this is not a fixed timeout */
2071 if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
2074 /* If not in hash table, timer will not be active yet */
2075 if (nf_ct_is_confirmed(ct))
2076 extra_jiffies += nfct_time_stamp;
2078 if (READ_ONCE(ct->timeout) != extra_jiffies)
2079 WRITE_ONCE(ct->timeout, extra_jiffies);
2082 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
2084 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
2086 bool nf_ct_kill_acct(struct nf_conn *ct,
2087 enum ip_conntrack_info ctinfo,
2088 const struct sk_buff *skb)
2090 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
2092 return nf_ct_delete(ct, 0, 0);
2094 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
2096 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
2098 #include <linux/netfilter/nfnetlink.h>
2099 #include <linux/netfilter/nfnetlink_conntrack.h>
2100 #include <linux/mutex.h>
2102 /* Generic function for tcp/udp/sctp/dccp and alike. */
2103 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
2104 const struct nf_conntrack_tuple *tuple)
2106 if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
2107 nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
2108 goto nla_put_failure;
2114 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
2116 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
2117 [CTA_PROTO_SRC_PORT] = { .type = NLA_U16 },
2118 [CTA_PROTO_DST_PORT] = { .type = NLA_U16 },
2120 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
2122 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
2123 struct nf_conntrack_tuple *t,
2126 if (flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) {
2127 if (!tb[CTA_PROTO_SRC_PORT])
2130 t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
2133 if (flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) {
2134 if (!tb[CTA_PROTO_DST_PORT])
2137 t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
2142 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
2144 unsigned int nf_ct_port_nlattr_tuple_size(void)
2146 static unsigned int size __read_mostly;
2149 size = nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
2153 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
2156 /* Used by ipt_REJECT and ip6t_REJECT. */
2157 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
2160 enum ip_conntrack_info ctinfo;
2162 /* This ICMP is in reverse direction to the packet which caused it */
2163 ct = nf_ct_get(skb, &ctinfo);
2164 if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
2165 ctinfo = IP_CT_RELATED_REPLY;
2167 ctinfo = IP_CT_RELATED;
2169 /* Attach to new skbuff, and increment count */
2170 nf_ct_set(nskb, ct, ctinfo);
2171 nf_conntrack_get(skb_nfct(nskb));
2174 static int __nf_conntrack_update(struct net *net, struct sk_buff *skb,
2176 enum ip_conntrack_info ctinfo)
2178 const struct nf_nat_hook *nat_hook;
2179 struct nf_conntrack_tuple_hash *h;
2180 struct nf_conntrack_tuple tuple;
2181 unsigned int status;
2186 l3num = nf_ct_l3num(ct);
2188 dataoff = get_l4proto(skb, skb_network_offset(skb), l3num, &l4num);
2192 if (!nf_ct_get_tuple(skb, skb_network_offset(skb), dataoff, l3num,
2193 l4num, net, &tuple))
2196 if (ct->status & IPS_SRC_NAT) {
2197 memcpy(tuple.src.u3.all,
2198 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.all,
2199 sizeof(tuple.src.u3.all));
2201 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u.all;
2204 if (ct->status & IPS_DST_NAT) {
2205 memcpy(tuple.dst.u3.all,
2206 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.all,
2207 sizeof(tuple.dst.u3.all));
2209 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u.all;
2212 h = nf_conntrack_find_get(net, nf_ct_zone(ct), &tuple);
2216 /* Store status bits of the conntrack that is clashing to re-do NAT
2217 * mangling according to what it has been done already to this packet.
2219 status = ct->status;
2222 ct = nf_ct_tuplehash_to_ctrack(h);
2223 nf_ct_set(skb, ct, ctinfo);
2225 nat_hook = rcu_dereference(nf_nat_hook);
2229 if (status & IPS_SRC_NAT &&
2230 nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_SRC,
2231 IP_CT_DIR_ORIGINAL) == NF_DROP)
2234 if (status & IPS_DST_NAT &&
2235 nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_DST,
2236 IP_CT_DIR_ORIGINAL) == NF_DROP)
2242 /* This packet is coming from userspace via nf_queue, complete the packet
2243 * processing after the helper invocation in nf_confirm().
2245 static int nf_confirm_cthelper(struct sk_buff *skb, struct nf_conn *ct,
2246 enum ip_conntrack_info ctinfo)
2248 const struct nf_conntrack_helper *helper;
2249 const struct nf_conn_help *help;
2252 help = nfct_help(ct);
2256 helper = rcu_dereference(help->helper);
2257 if (!(helper->flags & NF_CT_HELPER_F_USERSPACE))
2260 switch (nf_ct_l3num(ct)) {
2262 protoff = skb_network_offset(skb) + ip_hdrlen(skb);
2264 #if IS_ENABLED(CONFIG_IPV6)
2265 case NFPROTO_IPV6: {
2269 pnum = ipv6_hdr(skb)->nexthdr;
2270 protoff = ipv6_skip_exthdr(skb, sizeof(struct ipv6hdr), &pnum,
2272 if (protoff < 0 || (frag_off & htons(~0x7)) != 0)
2281 if (test_bit(IPS_SEQ_ADJUST_BIT, &ct->status) &&
2282 !nf_is_loopback_packet(skb)) {
2283 if (!nf_ct_seq_adjust(skb, ct, ctinfo, protoff)) {
2284 NF_CT_STAT_INC_ATOMIC(nf_ct_net(ct), drop);
2289 /* We've seen it coming out the other side: confirm it */
2290 return nf_conntrack_confirm(skb) == NF_DROP ? - 1 : 0;
2293 static int nf_conntrack_update(struct net *net, struct sk_buff *skb)
2295 enum ip_conntrack_info ctinfo;
2299 ct = nf_ct_get(skb, &ctinfo);
2303 if (!nf_ct_is_confirmed(ct)) {
2304 err = __nf_conntrack_update(net, skb, ct, ctinfo);
2308 ct = nf_ct_get(skb, &ctinfo);
2311 return nf_confirm_cthelper(skb, ct, ctinfo);
2314 static bool nf_conntrack_get_tuple_skb(struct nf_conntrack_tuple *dst_tuple,
2315 const struct sk_buff *skb)
2317 const struct nf_conntrack_tuple *src_tuple;
2318 const struct nf_conntrack_tuple_hash *hash;
2319 struct nf_conntrack_tuple srctuple;
2320 enum ip_conntrack_info ctinfo;
2323 ct = nf_ct_get(skb, &ctinfo);
2325 src_tuple = nf_ct_tuple(ct, CTINFO2DIR(ctinfo));
2326 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2330 if (!nf_ct_get_tuplepr(skb, skb_network_offset(skb),
2331 NFPROTO_IPV4, dev_net(skb->dev),
2335 hash = nf_conntrack_find_get(dev_net(skb->dev),
2341 ct = nf_ct_tuplehash_to_ctrack(hash);
2342 src_tuple = nf_ct_tuple(ct, !hash->tuple.dst.dir);
2343 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2349 /* Bring out ya dead! */
2350 static struct nf_conn *
2351 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
2352 const struct nf_ct_iter_data *iter_data, unsigned int *bucket)
2354 struct nf_conntrack_tuple_hash *h;
2356 struct hlist_nulls_node *n;
2359 for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
2360 struct hlist_nulls_head *hslot = &nf_conntrack_hash[*bucket];
2362 if (hlist_nulls_empty(hslot))
2365 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
2367 nf_conntrack_lock(lockp);
2368 hlist_nulls_for_each_entry(h, n, hslot, hnnode) {
2369 if (NF_CT_DIRECTION(h) != IP_CT_DIR_REPLY)
2371 /* All nf_conn objects are added to hash table twice, one
2372 * for original direction tuple, once for the reply tuple.
2374 * Exception: In the IPS_NAT_CLASH case, only the reply
2375 * tuple is added (the original tuple already existed for
2376 * a different object).
2378 * We only need to call the iterator once for each
2379 * conntrack, so we just use the 'reply' direction
2380 * tuple while iterating.
2382 ct = nf_ct_tuplehash_to_ctrack(h);
2384 if (iter_data->net &&
2385 !net_eq(iter_data->net, nf_ct_net(ct)))
2388 if (iter(ct, iter_data->data))
2398 refcount_inc(&ct->ct_general.use);
2404 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
2405 const struct nf_ct_iter_data *iter_data)
2407 unsigned int bucket = 0;
2412 mutex_lock(&nf_conntrack_mutex);
2413 while ((ct = get_next_corpse(iter, iter_data, &bucket)) != NULL) {
2414 /* Time to push up daises... */
2416 nf_ct_delete(ct, iter_data->portid, iter_data->report);
2420 mutex_unlock(&nf_conntrack_mutex);
2423 void nf_ct_iterate_cleanup_net(int (*iter)(struct nf_conn *i, void *data),
2424 const struct nf_ct_iter_data *iter_data)
2426 struct net *net = iter_data->net;
2427 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2431 if (atomic_read(&cnet->count) == 0)
2434 nf_ct_iterate_cleanup(iter, iter_data);
2436 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
2439 * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
2440 * @iter: callback to invoke for each conntrack
2441 * @data: data to pass to @iter
2443 * Like nf_ct_iterate_cleanup, but first marks conntracks on the
2444 * unconfirmed list as dying (so they will not be inserted into
2447 * Can only be called in module exit path.
2450 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
2452 struct nf_ct_iter_data iter_data = {};
2455 down_read(&net_rwsem);
2457 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2459 if (atomic_read(&cnet->count) == 0)
2461 nf_queue_nf_hook_drop(net);
2463 up_read(&net_rwsem);
2465 /* Need to wait for netns cleanup worker to finish, if its
2466 * running -- it might have deleted a net namespace from
2467 * the global list, so hook drop above might not have
2468 * affected all namespaces.
2472 /* a skb w. unconfirmed conntrack could have been reinjected just
2473 * before we called nf_queue_nf_hook_drop().
2475 * This makes sure its inserted into conntrack table.
2479 nf_ct_ext_bump_genid();
2480 iter_data.data = data;
2481 nf_ct_iterate_cleanup(iter, &iter_data);
2483 /* Another cpu might be in a rcu read section with
2484 * rcu protected pointer cleared in iter callback
2485 * or hidden via nf_ct_ext_bump_genid() above.
2487 * Wait until those are done.
2491 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
2493 static int kill_all(struct nf_conn *i, void *data)
2498 void nf_conntrack_cleanup_start(void)
2500 cleanup_nf_conntrack_bpf();
2501 conntrack_gc_work.exiting = true;
2504 void nf_conntrack_cleanup_end(void)
2506 RCU_INIT_POINTER(nf_ct_hook, NULL);
2507 cancel_delayed_work_sync(&conntrack_gc_work.dwork);
2508 kvfree(nf_conntrack_hash);
2510 nf_conntrack_proto_fini();
2511 nf_conntrack_helper_fini();
2512 nf_conntrack_expect_fini();
2514 kmem_cache_destroy(nf_conntrack_cachep);
2518 * Mishearing the voices in his head, our hero wonders how he's
2519 * supposed to kill the mall.
2521 void nf_conntrack_cleanup_net(struct net *net)
2525 list_add(&net->exit_list, &single);
2526 nf_conntrack_cleanup_net_list(&single);
2529 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
2531 struct nf_ct_iter_data iter_data = {};
2536 * This makes sure all current packets have passed through
2537 * netfilter framework. Roll on, two-stage module
2543 list_for_each_entry(net, net_exit_list, exit_list) {
2544 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2546 iter_data.net = net;
2547 nf_ct_iterate_cleanup_net(kill_all, &iter_data);
2548 if (atomic_read(&cnet->count) != 0)
2553 goto i_see_dead_people;
2556 list_for_each_entry(net, net_exit_list, exit_list) {
2557 nf_conntrack_ecache_pernet_fini(net);
2558 nf_conntrack_expect_pernet_fini(net);
2559 free_percpu(net->ct.stat);
2563 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
2565 struct hlist_nulls_head *hash;
2566 unsigned int nr_slots, i;
2568 if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
2571 BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
2572 nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
2574 hash = kvcalloc(nr_slots, sizeof(struct hlist_nulls_head), GFP_KERNEL);
2577 for (i = 0; i < nr_slots; i++)
2578 INIT_HLIST_NULLS_HEAD(&hash[i], i);
2582 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
2584 int nf_conntrack_hash_resize(unsigned int hashsize)
2587 unsigned int old_size;
2588 struct hlist_nulls_head *hash, *old_hash;
2589 struct nf_conntrack_tuple_hash *h;
2595 hash = nf_ct_alloc_hashtable(&hashsize, 1);
2599 mutex_lock(&nf_conntrack_mutex);
2600 old_size = nf_conntrack_htable_size;
2601 if (old_size == hashsize) {
2602 mutex_unlock(&nf_conntrack_mutex);
2608 nf_conntrack_all_lock();
2609 write_seqcount_begin(&nf_conntrack_generation);
2611 /* Lookups in the old hash might happen in parallel, which means we
2612 * might get false negatives during connection lookup. New connections
2613 * created because of a false negative won't make it into the hash
2614 * though since that required taking the locks.
2617 for (i = 0; i < nf_conntrack_htable_size; i++) {
2618 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
2619 unsigned int zone_id;
2621 h = hlist_nulls_entry(nf_conntrack_hash[i].first,
2622 struct nf_conntrack_tuple_hash, hnnode);
2623 ct = nf_ct_tuplehash_to_ctrack(h);
2624 hlist_nulls_del_rcu(&h->hnnode);
2626 zone_id = nf_ct_zone_id(nf_ct_zone(ct), NF_CT_DIRECTION(h));
2627 bucket = __hash_conntrack(nf_ct_net(ct),
2628 &h->tuple, zone_id, hashsize);
2629 hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
2632 old_hash = nf_conntrack_hash;
2634 nf_conntrack_hash = hash;
2635 nf_conntrack_htable_size = hashsize;
2637 write_seqcount_end(&nf_conntrack_generation);
2638 nf_conntrack_all_unlock();
2641 mutex_unlock(&nf_conntrack_mutex);
2648 int nf_conntrack_set_hashsize(const char *val, const struct kernel_param *kp)
2650 unsigned int hashsize;
2653 if (current->nsproxy->net_ns != &init_net)
2656 /* On boot, we can set this without any fancy locking. */
2657 if (!nf_conntrack_hash)
2658 return param_set_uint(val, kp);
2660 rc = kstrtouint(val, 0, &hashsize);
2664 return nf_conntrack_hash_resize(hashsize);
2667 int nf_conntrack_init_start(void)
2669 unsigned long nr_pages = totalram_pages();
2674 seqcount_spinlock_init(&nf_conntrack_generation,
2675 &nf_conntrack_locks_all_lock);
2677 for (i = 0; i < CONNTRACK_LOCKS; i++)
2678 spin_lock_init(&nf_conntrack_locks[i]);
2680 if (!nf_conntrack_htable_size) {
2681 nf_conntrack_htable_size
2682 = (((nr_pages << PAGE_SHIFT) / 16384)
2683 / sizeof(struct hlist_head));
2684 if (BITS_PER_LONG >= 64 &&
2685 nr_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2686 nf_conntrack_htable_size = 262144;
2687 else if (nr_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2688 nf_conntrack_htable_size = 65536;
2690 if (nf_conntrack_htable_size < 1024)
2691 nf_conntrack_htable_size = 1024;
2692 /* Use a max. factor of one by default to keep the average
2693 * hash chain length at 2 entries. Each entry has to be added
2694 * twice (once for original direction, once for reply).
2695 * When a table size is given we use the old value of 8 to
2696 * avoid implicit reduction of the max entries setting.
2701 nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2702 if (!nf_conntrack_hash)
2705 nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2707 nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2708 sizeof(struct nf_conn),
2710 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2711 if (!nf_conntrack_cachep)
2714 ret = nf_conntrack_expect_init();
2718 ret = nf_conntrack_helper_init();
2722 ret = nf_conntrack_proto_init();
2726 conntrack_gc_work_init(&conntrack_gc_work);
2727 queue_delayed_work(system_power_efficient_wq, &conntrack_gc_work.dwork, HZ);
2729 ret = register_nf_conntrack_bpf();
2736 cancel_delayed_work_sync(&conntrack_gc_work.dwork);
2737 nf_conntrack_proto_fini();
2739 nf_conntrack_helper_fini();
2741 nf_conntrack_expect_fini();
2743 kmem_cache_destroy(nf_conntrack_cachep);
2745 kvfree(nf_conntrack_hash);
2749 static const struct nf_ct_hook nf_conntrack_hook = {
2750 .update = nf_conntrack_update,
2751 .destroy = nf_ct_destroy,
2752 .get_tuple_skb = nf_conntrack_get_tuple_skb,
2753 .attach = nf_conntrack_attach,
2756 void nf_conntrack_init_end(void)
2758 RCU_INIT_POINTER(nf_ct_hook, &nf_conntrack_hook);
2762 * We need to use special "null" values, not used in hash table
2764 #define UNCONFIRMED_NULLS_VAL ((1<<30)+0)
2766 int nf_conntrack_init_net(struct net *net)
2768 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2771 BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2772 BUILD_BUG_ON_NOT_POWER_OF_2(CONNTRACK_LOCKS);
2773 atomic_set(&cnet->count, 0);
2775 net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2779 ret = nf_conntrack_expect_pernet_init(net);
2783 nf_conntrack_acct_pernet_init(net);
2784 nf_conntrack_tstamp_pernet_init(net);
2785 nf_conntrack_ecache_pernet_init(net);
2786 nf_conntrack_proto_pernet_init(net);
2791 free_percpu(net->ct.stat);
2795 /* ctnetlink code shared by both ctnetlink and nf_conntrack_bpf */
2797 int __nf_ct_change_timeout(struct nf_conn *ct, u64 timeout)
2799 if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
2802 __nf_ct_set_timeout(ct, timeout);
2804 if (test_bit(IPS_DYING_BIT, &ct->status))
2809 EXPORT_SYMBOL_GPL(__nf_ct_change_timeout);
2811 void __nf_ct_change_status(struct nf_conn *ct, unsigned long on, unsigned long off)
2815 /* Ignore these unchangable bits */
2816 on &= ~IPS_UNCHANGEABLE_MASK;
2817 off &= ~IPS_UNCHANGEABLE_MASK;
2819 for (bit = 0; bit < __IPS_MAX_BIT; bit++) {
2820 if (on & (1 << bit))
2821 set_bit(bit, &ct->status);
2822 else if (off & (1 << bit))
2823 clear_bit(bit, &ct->status);
2826 EXPORT_SYMBOL_GPL(__nf_ct_change_status);
2828 int nf_ct_change_status_common(struct nf_conn *ct, unsigned int status)
2832 d = ct->status ^ status;
2834 if (d & (IPS_EXPECTED|IPS_CONFIRMED|IPS_DYING))
2838 if (d & IPS_SEEN_REPLY && !(status & IPS_SEEN_REPLY))
2839 /* SEEN_REPLY bit can only be set */
2842 if (d & IPS_ASSURED && !(status & IPS_ASSURED))
2843 /* ASSURED bit can only be set */
2846 __nf_ct_change_status(ct, status, 0);
2849 EXPORT_SYMBOL_GPL(nf_ct_change_status_common);