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/jhash.h>
25 #include <linux/siphash.h>
26 #include <linux/err.h>
27 #include <linux/percpu.h>
28 #include <linux/moduleparam.h>
29 #include <linux/notifier.h>
30 #include <linux/kernel.h>
31 #include <linux/netdevice.h>
32 #include <linux/socket.h>
34 #include <linux/nsproxy.h>
35 #include <linux/rculist_nulls.h>
37 #include <net/netfilter/nf_conntrack.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_seqadj.h>
42 #include <net/netfilter/nf_conntrack_core.h>
43 #include <net/netfilter/nf_conntrack_extend.h>
44 #include <net/netfilter/nf_conntrack_acct.h>
45 #include <net/netfilter/nf_conntrack_ecache.h>
46 #include <net/netfilter/nf_conntrack_zones.h>
47 #include <net/netfilter/nf_conntrack_timestamp.h>
48 #include <net/netfilter/nf_conntrack_timeout.h>
49 #include <net/netfilter/nf_conntrack_labels.h>
50 #include <net/netfilter/nf_conntrack_synproxy.h>
51 #include <net/netfilter/nf_nat.h>
52 #include <net/netfilter/nf_nat_helper.h>
53 #include <net/netns/hash.h>
56 #include "nf_internals.h"
58 __cacheline_aligned_in_smp spinlock_t nf_conntrack_locks[CONNTRACK_LOCKS];
59 EXPORT_SYMBOL_GPL(nf_conntrack_locks);
61 __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
62 EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
64 struct hlist_nulls_head *nf_conntrack_hash __read_mostly;
65 EXPORT_SYMBOL_GPL(nf_conntrack_hash);
67 struct conntrack_gc_work {
68 struct delayed_work dwork;
75 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
76 static DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
77 static __read_mostly bool nf_conntrack_locks_all;
79 /* every gc cycle scans at most 1/GC_MAX_BUCKETS_DIV part of table */
80 #define GC_MAX_BUCKETS_DIV 128u
81 /* upper bound of full table scan */
82 #define GC_MAX_SCAN_JIFFIES (16u * HZ)
83 /* desired ratio of entries found to be expired */
84 #define GC_EVICT_RATIO 50u
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 unsigned int nf_conntrack_hash_rnd __read_mostly;
194 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
195 const struct net *net)
200 get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
202 /* The direction must be ignored, so we hash everything up to the
203 * destination ports (which is a multiple of 4) and treat the last
204 * three bytes manually.
206 seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
207 n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
208 return jhash2((u32 *)tuple, n, seed ^
209 (((__force __u16)tuple->dst.u.all << 16) |
210 tuple->dst.protonum));
213 static u32 scale_hash(u32 hash)
215 return reciprocal_scale(hash, nf_conntrack_htable_size);
218 static u32 __hash_conntrack(const struct net *net,
219 const struct nf_conntrack_tuple *tuple,
222 return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
225 static u32 hash_conntrack(const struct net *net,
226 const struct nf_conntrack_tuple *tuple)
228 return scale_hash(hash_conntrack_raw(tuple, net));
231 static bool nf_ct_get_tuple_ports(const struct sk_buff *skb,
232 unsigned int dataoff,
233 struct nf_conntrack_tuple *tuple)
237 } _inet_hdr, *inet_hdr;
239 /* Actually only need first 4 bytes to get ports. */
240 inet_hdr = skb_header_pointer(skb, dataoff, sizeof(_inet_hdr), &_inet_hdr);
244 tuple->src.u.udp.port = inet_hdr->sport;
245 tuple->dst.u.udp.port = inet_hdr->dport;
250 nf_ct_get_tuple(const struct sk_buff *skb,
252 unsigned int dataoff,
256 struct nf_conntrack_tuple *tuple)
262 memset(tuple, 0, sizeof(*tuple));
264 tuple->src.l3num = l3num;
267 nhoff += offsetof(struct iphdr, saddr);
268 size = 2 * sizeof(__be32);
271 nhoff += offsetof(struct ipv6hdr, saddr);
272 size = sizeof(_addrs);
278 ap = skb_header_pointer(skb, nhoff, size, _addrs);
284 tuple->src.u3.ip = ap[0];
285 tuple->dst.u3.ip = ap[1];
288 memcpy(tuple->src.u3.ip6, ap, sizeof(tuple->src.u3.ip6));
289 memcpy(tuple->dst.u3.ip6, ap + 4, sizeof(tuple->dst.u3.ip6));
293 tuple->dst.protonum = protonum;
294 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
297 #if IS_ENABLED(CONFIG_IPV6)
299 return icmpv6_pkt_to_tuple(skb, dataoff, net, tuple);
302 return icmp_pkt_to_tuple(skb, dataoff, net, tuple);
303 #ifdef CONFIG_NF_CT_PROTO_GRE
305 return gre_pkt_to_tuple(skb, dataoff, net, tuple);
308 case IPPROTO_UDP: /* fallthrough */
309 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
310 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
311 case IPPROTO_UDPLITE:
312 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
314 #ifdef CONFIG_NF_CT_PROTO_SCTP
316 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
318 #ifdef CONFIG_NF_CT_PROTO_DCCP
320 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
329 static int ipv4_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
333 const struct iphdr *iph;
336 iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
340 /* Conntrack defragments packets, we might still see fragments
341 * inside ICMP packets though.
343 if (iph->frag_off & htons(IP_OFFSET))
346 dataoff = nhoff + (iph->ihl << 2);
347 *protonum = iph->protocol;
349 /* Check bogus IP headers */
350 if (dataoff > skb->len) {
351 pr_debug("bogus IPv4 packet: nhoff %u, ihl %u, skblen %u\n",
352 nhoff, iph->ihl << 2, skb->len);
358 #if IS_ENABLED(CONFIG_IPV6)
359 static int ipv6_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
363 unsigned int extoff = nhoff + sizeof(struct ipv6hdr);
367 if (skb_copy_bits(skb, nhoff + offsetof(struct ipv6hdr, nexthdr),
368 &nexthdr, sizeof(nexthdr)) != 0) {
369 pr_debug("can't get nexthdr\n");
372 protoff = ipv6_skip_exthdr(skb, extoff, &nexthdr, &frag_off);
374 * (protoff == skb->len) means the packet has not data, just
375 * IPv6 and possibly extensions headers, but it is tracked anyway
377 if (protoff < 0 || (frag_off & htons(~0x7)) != 0) {
378 pr_debug("can't find proto in pkt\n");
387 static int get_l4proto(const struct sk_buff *skb,
388 unsigned int nhoff, u8 pf, u8 *l4num)
392 return ipv4_get_l4proto(skb, nhoff, l4num);
393 #if IS_ENABLED(CONFIG_IPV6)
395 return ipv6_get_l4proto(skb, nhoff, l4num);
404 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
406 struct net *net, struct nf_conntrack_tuple *tuple)
411 protoff = get_l4proto(skb, nhoff, l3num, &protonum);
415 return nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple);
417 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
420 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
421 const struct nf_conntrack_tuple *orig)
423 memset(inverse, 0, sizeof(*inverse));
425 inverse->src.l3num = orig->src.l3num;
427 switch (orig->src.l3num) {
429 inverse->src.u3.ip = orig->dst.u3.ip;
430 inverse->dst.u3.ip = orig->src.u3.ip;
433 inverse->src.u3.in6 = orig->dst.u3.in6;
434 inverse->dst.u3.in6 = orig->src.u3.in6;
440 inverse->dst.dir = !orig->dst.dir;
442 inverse->dst.protonum = orig->dst.protonum;
444 switch (orig->dst.protonum) {
446 return nf_conntrack_invert_icmp_tuple(inverse, orig);
447 #if IS_ENABLED(CONFIG_IPV6)
449 return nf_conntrack_invert_icmpv6_tuple(inverse, orig);
453 inverse->src.u.all = orig->dst.u.all;
454 inverse->dst.u.all = orig->src.u.all;
457 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
459 /* Generate a almost-unique pseudo-id for a given conntrack.
461 * intentionally doesn't re-use any of the seeds used for hash
462 * table location, we assume id gets exposed to userspace.
464 * Following nf_conn items do not change throughout lifetime
468 * 2. nf_conn->master address (normally NULL)
469 * 3. the associated net namespace
470 * 4. the original direction tuple
472 u32 nf_ct_get_id(const struct nf_conn *ct)
474 static __read_mostly siphash_key_t ct_id_seed;
475 unsigned long a, b, c, d;
477 net_get_random_once(&ct_id_seed, sizeof(ct_id_seed));
479 a = (unsigned long)ct;
480 b = (unsigned long)ct->master;
481 c = (unsigned long)nf_ct_net(ct);
482 d = (unsigned long)siphash(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
483 sizeof(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple),
486 return siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &ct_id_seed);
488 return siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &ct_id_seed);
491 EXPORT_SYMBOL_GPL(nf_ct_get_id);
494 clean_from_lists(struct nf_conn *ct)
496 pr_debug("clean_from_lists(%p)\n", ct);
497 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
498 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
500 /* Destroy all pending expectations */
501 nf_ct_remove_expectations(ct);
504 /* must be called with local_bh_disable */
505 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
507 struct ct_pcpu *pcpu;
509 /* add this conntrack to the (per cpu) dying list */
510 ct->cpu = smp_processor_id();
511 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
513 spin_lock(&pcpu->lock);
514 hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
516 spin_unlock(&pcpu->lock);
519 /* must be called with local_bh_disable */
520 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
522 struct ct_pcpu *pcpu;
524 /* add this conntrack to the (per cpu) unconfirmed list */
525 ct->cpu = smp_processor_id();
526 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
528 spin_lock(&pcpu->lock);
529 hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
531 spin_unlock(&pcpu->lock);
534 /* must be called with local_bh_disable */
535 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
537 struct ct_pcpu *pcpu;
539 /* We overload first tuple to link into unconfirmed or dying list.*/
540 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
542 spin_lock(&pcpu->lock);
543 BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
544 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
545 spin_unlock(&pcpu->lock);
548 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
550 /* Released via destroy_conntrack() */
551 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
552 const struct nf_conntrack_zone *zone,
555 struct nf_conn *tmpl, *p;
557 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
558 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
563 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
565 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
566 tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
569 tmpl = kzalloc(sizeof(*tmpl), flags);
574 tmpl->status = IPS_TEMPLATE;
575 write_pnet(&tmpl->ct_net, net);
576 nf_ct_zone_add(tmpl, zone);
577 atomic_set(&tmpl->ct_general.use, 0);
581 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
583 void nf_ct_tmpl_free(struct nf_conn *tmpl)
585 nf_ct_ext_destroy(tmpl);
587 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
588 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
592 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
594 static void destroy_gre_conntrack(struct nf_conn *ct)
596 #ifdef CONFIG_NF_CT_PROTO_GRE
597 struct nf_conn *master = ct->master;
600 nf_ct_gre_keymap_destroy(master);
605 destroy_conntrack(struct nf_conntrack *nfct)
607 struct nf_conn *ct = (struct nf_conn *)nfct;
609 pr_debug("destroy_conntrack(%p)\n", ct);
610 WARN_ON(atomic_read(&nfct->use) != 0);
612 if (unlikely(nf_ct_is_template(ct))) {
617 if (unlikely(nf_ct_protonum(ct) == IPPROTO_GRE))
618 destroy_gre_conntrack(ct);
621 /* Expectations will have been removed in clean_from_lists,
622 * except TFTP can create an expectation on the first packet,
623 * before connection is in the list, so we need to clean here,
626 nf_ct_remove_expectations(ct);
628 nf_ct_del_from_dying_or_unconfirmed_list(ct);
633 nf_ct_put(ct->master);
635 pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
636 nf_conntrack_free(ct);
639 static void nf_ct_delete_from_lists(struct nf_conn *ct)
641 struct net *net = nf_ct_net(ct);
642 unsigned int hash, reply_hash;
643 unsigned int sequence;
645 nf_ct_helper_destroy(ct);
649 sequence = read_seqcount_begin(&nf_conntrack_generation);
650 hash = hash_conntrack(net,
651 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
652 reply_hash = hash_conntrack(net,
653 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
654 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
656 clean_from_lists(ct);
657 nf_conntrack_double_unlock(hash, reply_hash);
659 nf_ct_add_to_dying_list(ct);
664 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
666 struct nf_conn_tstamp *tstamp;
669 if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
672 tstamp = nf_conn_tstamp_find(ct);
673 if (tstamp && tstamp->stop == 0)
674 tstamp->stop = ktime_get_real_ns();
676 if (nf_conntrack_event_report(IPCT_DESTROY, ct,
677 portid, report) < 0) {
678 /* destroy event was not delivered. nf_ct_put will
679 * be done by event cache worker on redelivery.
681 nf_ct_delete_from_lists(ct);
682 nf_conntrack_ecache_work(nf_ct_net(ct), NFCT_ECACHE_DESTROY_FAIL);
687 if (nf_conntrack_ecache_dwork_pending(net))
688 nf_conntrack_ecache_work(net, NFCT_ECACHE_DESTROY_SENT);
689 nf_ct_delete_from_lists(ct);
693 EXPORT_SYMBOL_GPL(nf_ct_delete);
696 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
697 const struct nf_conntrack_tuple *tuple,
698 const struct nf_conntrack_zone *zone,
699 const struct net *net)
701 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
703 /* A conntrack can be recreated with the equal tuple,
704 * so we need to check that the conntrack is confirmed
706 return nf_ct_tuple_equal(tuple, &h->tuple) &&
707 nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
708 nf_ct_is_confirmed(ct) &&
709 net_eq(net, nf_ct_net(ct));
713 nf_ct_match(const struct nf_conn *ct1, const struct nf_conn *ct2)
715 return nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
716 &ct2->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
717 nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_REPLY].tuple,
718 &ct2->tuplehash[IP_CT_DIR_REPLY].tuple) &&
719 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_ORIGINAL) &&
720 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_REPLY) &&
721 net_eq(nf_ct_net(ct1), nf_ct_net(ct2));
724 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
725 static void nf_ct_gc_expired(struct nf_conn *ct)
727 if (!atomic_inc_not_zero(&ct->ct_general.use))
730 if (nf_ct_should_gc(ct))
738 * - Caller must take a reference on returned object
739 * and recheck nf_ct_tuple_equal(tuple, &h->tuple)
741 static struct nf_conntrack_tuple_hash *
742 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
743 const struct nf_conntrack_tuple *tuple, u32 hash)
745 struct nf_conntrack_tuple_hash *h;
746 struct hlist_nulls_head *ct_hash;
747 struct hlist_nulls_node *n;
748 unsigned int bucket, hsize;
751 nf_conntrack_get_ht(&ct_hash, &hsize);
752 bucket = reciprocal_scale(hash, hsize);
754 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
757 ct = nf_ct_tuplehash_to_ctrack(h);
758 if (nf_ct_is_expired(ct)) {
759 nf_ct_gc_expired(ct);
763 if (nf_ct_key_equal(h, tuple, zone, net))
767 * if the nulls value we got at the end of this lookup is
768 * not the expected one, we must restart lookup.
769 * We probably met an item that was moved to another chain.
771 if (get_nulls_value(n) != bucket) {
772 NF_CT_STAT_INC_ATOMIC(net, search_restart);
779 /* Find a connection corresponding to a tuple. */
780 static struct nf_conntrack_tuple_hash *
781 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
782 const struct nf_conntrack_tuple *tuple, u32 hash)
784 struct nf_conntrack_tuple_hash *h;
789 h = ____nf_conntrack_find(net, zone, tuple, hash);
791 /* We have a candidate that matches the tuple we're interested
792 * in, try to obtain a reference and re-check tuple
794 ct = nf_ct_tuplehash_to_ctrack(h);
795 if (likely(atomic_inc_not_zero(&ct->ct_general.use))) {
796 if (likely(nf_ct_key_equal(h, tuple, zone, net)))
799 /* TYPESAFE_BY_RCU recycled the candidate */
811 struct nf_conntrack_tuple_hash *
812 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
813 const struct nf_conntrack_tuple *tuple)
815 return __nf_conntrack_find_get(net, zone, tuple,
816 hash_conntrack_raw(tuple, net));
818 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
820 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
822 unsigned int reply_hash)
824 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
825 &nf_conntrack_hash[hash]);
826 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
827 &nf_conntrack_hash[reply_hash]);
831 nf_conntrack_hash_check_insert(struct nf_conn *ct)
833 const struct nf_conntrack_zone *zone;
834 struct net *net = nf_ct_net(ct);
835 unsigned int hash, reply_hash;
836 struct nf_conntrack_tuple_hash *h;
837 struct hlist_nulls_node *n;
838 unsigned int sequence;
840 zone = nf_ct_zone(ct);
844 sequence = read_seqcount_begin(&nf_conntrack_generation);
845 hash = hash_conntrack(net,
846 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
847 reply_hash = hash_conntrack(net,
848 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
849 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
851 /* See if there's one in the list already, including reverse */
852 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
853 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
857 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
858 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
863 /* The caller holds a reference to this object */
864 atomic_set(&ct->ct_general.use, 2);
865 __nf_conntrack_hash_insert(ct, hash, reply_hash);
866 nf_conntrack_double_unlock(hash, reply_hash);
867 NF_CT_STAT_INC(net, insert);
872 nf_conntrack_double_unlock(hash, reply_hash);
876 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
878 void nf_ct_acct_add(struct nf_conn *ct, u32 dir, unsigned int packets,
881 struct nf_conn_acct *acct;
883 acct = nf_conn_acct_find(ct);
885 struct nf_conn_counter *counter = acct->counter;
887 atomic64_add(packets, &counter[dir].packets);
888 atomic64_add(bytes, &counter[dir].bytes);
891 EXPORT_SYMBOL_GPL(nf_ct_acct_add);
893 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
894 const struct nf_conn *loser_ct)
896 struct nf_conn_acct *acct;
898 acct = nf_conn_acct_find(loser_ct);
900 struct nf_conn_counter *counter = acct->counter;
903 /* u32 should be fine since we must have seen one packet. */
904 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
905 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), bytes);
909 static void __nf_conntrack_insert_prepare(struct nf_conn *ct)
911 struct nf_conn_tstamp *tstamp;
913 atomic_inc(&ct->ct_general.use);
914 ct->status |= IPS_CONFIRMED;
916 /* set conntrack timestamp, if enabled. */
917 tstamp = nf_conn_tstamp_find(ct);
919 tstamp->start = ktime_get_real_ns();
922 /* caller must hold locks to prevent concurrent changes */
923 static int __nf_ct_resolve_clash(struct sk_buff *skb,
924 struct nf_conntrack_tuple_hash *h)
926 /* This is the conntrack entry already in hashes that won race. */
927 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
928 enum ip_conntrack_info ctinfo;
929 struct nf_conn *loser_ct;
931 loser_ct = nf_ct_get(skb, &ctinfo);
933 if (nf_ct_is_dying(ct))
936 if (((ct->status & IPS_NAT_DONE_MASK) == 0) ||
937 nf_ct_match(ct, loser_ct)) {
938 struct net *net = nf_ct_net(ct);
940 nf_conntrack_get(&ct->ct_general);
942 nf_ct_acct_merge(ct, ctinfo, loser_ct);
943 nf_ct_add_to_dying_list(loser_ct);
944 nf_conntrack_put(&loser_ct->ct_general);
945 nf_ct_set(skb, ct, ctinfo);
947 NF_CT_STAT_INC(net, clash_resolve);
955 * nf_ct_resolve_clash_harder - attempt to insert clashing conntrack entry
957 * @skb: skb that causes the collision
958 * @repl_idx: hash slot for reply direction
960 * Called when origin or reply direction had a clash.
961 * The skb can be handled without packet drop provided the reply direction
962 * is unique or there the existing entry has the identical tuple in both
965 * Caller must hold conntrack table locks to prevent concurrent updates.
967 * Returns NF_DROP if the clash could not be handled.
969 static int nf_ct_resolve_clash_harder(struct sk_buff *skb, u32 repl_idx)
971 struct nf_conn *loser_ct = (struct nf_conn *)skb_nfct(skb);
972 const struct nf_conntrack_zone *zone;
973 struct nf_conntrack_tuple_hash *h;
974 struct hlist_nulls_node *n;
977 zone = nf_ct_zone(loser_ct);
978 net = nf_ct_net(loser_ct);
980 /* Reply direction must never result in a clash, unless both origin
981 * and reply tuples are identical.
983 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[repl_idx], hnnode) {
984 if (nf_ct_key_equal(h,
985 &loser_ct->tuplehash[IP_CT_DIR_REPLY].tuple,
987 return __nf_ct_resolve_clash(skb, h);
990 /* We want the clashing entry to go away real soon: 1 second timeout. */
991 loser_ct->timeout = nfct_time_stamp + HZ;
993 /* IPS_NAT_CLASH removes the entry automatically on the first
994 * reply. Also prevents UDP tracker from moving the entry to
995 * ASSURED state, i.e. the entry can always be evicted under
998 loser_ct->status |= IPS_FIXED_TIMEOUT | IPS_NAT_CLASH;
1000 __nf_conntrack_insert_prepare(loser_ct);
1002 /* fake add for ORIGINAL dir: we want lookups to only find the entry
1003 * already in the table. This also hides the clashing entry from
1004 * ctnetlink iteration, i.e. conntrack -L won't show them.
1006 hlist_nulls_add_fake(&loser_ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
1008 hlist_nulls_add_head_rcu(&loser_ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
1009 &nf_conntrack_hash[repl_idx]);
1011 NF_CT_STAT_INC(net, clash_resolve);
1016 * nf_ct_resolve_clash - attempt to handle clash without packet drop
1018 * @skb: skb that causes the clash
1019 * @h: tuplehash of the clashing entry already in table
1020 * @reply_hash: hash slot for reply direction
1022 * A conntrack entry can be inserted to the connection tracking table
1023 * if there is no existing entry with an identical tuple.
1025 * If there is one, @skb (and the assocated, unconfirmed conntrack) has
1026 * to be dropped. In case @skb is retransmitted, next conntrack lookup
1027 * will find the already-existing entry.
1029 * The major problem with such packet drop is the extra delay added by
1030 * the packet loss -- it will take some time for a retransmit to occur
1031 * (or the sender to time out when waiting for a reply).
1033 * This function attempts to handle the situation without packet drop.
1035 * If @skb has no NAT transformation or if the colliding entries are
1036 * exactly the same, only the to-be-confirmed conntrack entry is discarded
1037 * and @skb is associated with the conntrack entry already in the table.
1039 * Failing that, the new, unconfirmed conntrack is still added to the table
1040 * provided that the collision only occurs in the ORIGINAL direction.
1041 * The new entry will be added only in the non-clashing REPLY direction,
1042 * so packets in the ORIGINAL direction will continue to match the existing
1043 * entry. The new entry will also have a fixed timeout so it expires --
1044 * due to the collision, it will only see reply traffic.
1046 * Returns NF_DROP if the clash could not be resolved.
1048 static __cold noinline int
1049 nf_ct_resolve_clash(struct sk_buff *skb, struct nf_conntrack_tuple_hash *h,
1052 /* This is the conntrack entry already in hashes that won race. */
1053 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1054 const struct nf_conntrack_l4proto *l4proto;
1055 enum ip_conntrack_info ctinfo;
1056 struct nf_conn *loser_ct;
1060 loser_ct = nf_ct_get(skb, &ctinfo);
1061 net = nf_ct_net(loser_ct);
1063 l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1064 if (!l4proto->allow_clash)
1067 ret = __nf_ct_resolve_clash(skb, h);
1068 if (ret == NF_ACCEPT)
1071 ret = nf_ct_resolve_clash_harder(skb, reply_hash);
1072 if (ret == NF_ACCEPT)
1076 nf_ct_add_to_dying_list(loser_ct);
1077 NF_CT_STAT_INC(net, drop);
1078 NF_CT_STAT_INC(net, insert_failed);
1082 /* Confirm a connection given skb; places it in hash table */
1084 __nf_conntrack_confirm(struct sk_buff *skb)
1086 const struct nf_conntrack_zone *zone;
1087 unsigned int hash, reply_hash;
1088 struct nf_conntrack_tuple_hash *h;
1090 struct nf_conn_help *help;
1091 struct hlist_nulls_node *n;
1092 enum ip_conntrack_info ctinfo;
1094 unsigned int sequence;
1097 ct = nf_ct_get(skb, &ctinfo);
1098 net = nf_ct_net(ct);
1100 /* ipt_REJECT uses nf_conntrack_attach to attach related
1101 ICMP/TCP RST packets in other direction. Actual packet
1102 which created connection will be IP_CT_NEW or for an
1103 expected connection, IP_CT_RELATED. */
1104 if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
1107 zone = nf_ct_zone(ct);
1111 sequence = read_seqcount_begin(&nf_conntrack_generation);
1112 /* reuse the hash saved before */
1113 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
1114 hash = scale_hash(hash);
1115 reply_hash = hash_conntrack(net,
1116 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
1118 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
1120 /* We're not in hash table, and we refuse to set up related
1121 * connections for unconfirmed conns. But packet copies and
1122 * REJECT will give spurious warnings here.
1125 /* Another skb with the same unconfirmed conntrack may
1126 * win the race. This may happen for bridge(br_flood)
1127 * or broadcast/multicast packets do skb_clone with
1128 * unconfirmed conntrack.
1130 if (unlikely(nf_ct_is_confirmed(ct))) {
1132 nf_conntrack_double_unlock(hash, reply_hash);
1137 pr_debug("Confirming conntrack %p\n", ct);
1138 /* We have to check the DYING flag after unlink to prevent
1139 * a race against nf_ct_get_next_corpse() possibly called from
1140 * user context, else we insert an already 'dead' hash, blocking
1141 * further use of that particular connection -JM.
1143 nf_ct_del_from_dying_or_unconfirmed_list(ct);
1145 if (unlikely(nf_ct_is_dying(ct))) {
1146 nf_ct_add_to_dying_list(ct);
1147 NF_CT_STAT_INC(net, insert_failed);
1151 /* See if there's one in the list already, including reverse:
1152 NAT could have grabbed it without realizing, since we're
1153 not in the hash. If there is, we lost race. */
1154 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
1155 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1159 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
1160 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1164 /* Timer relative to confirmation time, not original
1165 setting time, otherwise we'd get timer wrap in
1166 weird delay cases. */
1167 ct->timeout += nfct_time_stamp;
1169 __nf_conntrack_insert_prepare(ct);
1171 /* Since the lookup is lockless, hash insertion must be done after
1172 * starting the timer and setting the CONFIRMED bit. The RCU barriers
1173 * guarantee that no other CPU can find the conntrack before the above
1174 * stores are visible.
1176 __nf_conntrack_hash_insert(ct, hash, reply_hash);
1177 nf_conntrack_double_unlock(hash, reply_hash);
1180 help = nfct_help(ct);
1181 if (help && help->helper)
1182 nf_conntrack_event_cache(IPCT_HELPER, ct);
1184 nf_conntrack_event_cache(master_ct(ct) ?
1185 IPCT_RELATED : IPCT_NEW, ct);
1189 ret = nf_ct_resolve_clash(skb, h, reply_hash);
1191 nf_conntrack_double_unlock(hash, reply_hash);
1195 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
1197 /* Returns true if a connection correspondings to the tuple (required
1200 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
1201 const struct nf_conn *ignored_conntrack)
1203 struct net *net = nf_ct_net(ignored_conntrack);
1204 const struct nf_conntrack_zone *zone;
1205 struct nf_conntrack_tuple_hash *h;
1206 struct hlist_nulls_head *ct_hash;
1207 unsigned int hash, hsize;
1208 struct hlist_nulls_node *n;
1211 zone = nf_ct_zone(ignored_conntrack);
1215 nf_conntrack_get_ht(&ct_hash, &hsize);
1216 hash = __hash_conntrack(net, tuple, hsize);
1218 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
1219 ct = nf_ct_tuplehash_to_ctrack(h);
1221 if (ct == ignored_conntrack)
1224 if (nf_ct_is_expired(ct)) {
1225 nf_ct_gc_expired(ct);
1229 if (nf_ct_key_equal(h, tuple, zone, net)) {
1230 /* Tuple is taken already, so caller will need to find
1231 * a new source port to use.
1234 * If the *original tuples* are identical, then both
1235 * conntracks refer to the same flow.
1236 * This is a rare situation, it can occur e.g. when
1237 * more than one UDP packet is sent from same socket
1238 * in different threads.
1240 * Let nf_ct_resolve_clash() deal with this later.
1242 if (nf_ct_tuple_equal(&ignored_conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1243 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
1244 nf_ct_zone_equal(ct, zone, IP_CT_DIR_ORIGINAL))
1247 NF_CT_STAT_INC_ATOMIC(net, found);
1253 if (get_nulls_value(n) != hash) {
1254 NF_CT_STAT_INC_ATOMIC(net, search_restart);
1262 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
1264 #define NF_CT_EVICTION_RANGE 8
1266 /* There's a small race here where we may free a just-assured
1267 connection. Too bad: we're in trouble anyway. */
1268 static unsigned int early_drop_list(struct net *net,
1269 struct hlist_nulls_head *head)
1271 struct nf_conntrack_tuple_hash *h;
1272 struct hlist_nulls_node *n;
1273 unsigned int drops = 0;
1274 struct nf_conn *tmp;
1276 hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
1277 tmp = nf_ct_tuplehash_to_ctrack(h);
1279 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status))
1282 if (nf_ct_is_expired(tmp)) {
1283 nf_ct_gc_expired(tmp);
1287 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
1288 !net_eq(nf_ct_net(tmp), net) ||
1289 nf_ct_is_dying(tmp))
1292 if (!atomic_inc_not_zero(&tmp->ct_general.use))
1295 /* kill only if still in same netns -- might have moved due to
1296 * SLAB_TYPESAFE_BY_RCU rules.
1298 * We steal the timer reference. If that fails timer has
1299 * already fired or someone else deleted it. Just drop ref
1300 * and move to next entry.
1302 if (net_eq(nf_ct_net(tmp), net) &&
1303 nf_ct_is_confirmed(tmp) &&
1304 nf_ct_delete(tmp, 0, 0))
1313 static noinline int early_drop(struct net *net, unsigned int hash)
1315 unsigned int i, bucket;
1317 for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
1318 struct hlist_nulls_head *ct_hash;
1319 unsigned int hsize, drops;
1322 nf_conntrack_get_ht(&ct_hash, &hsize);
1324 bucket = reciprocal_scale(hash, hsize);
1326 bucket = (bucket + 1) % hsize;
1328 drops = early_drop_list(net, &ct_hash[bucket]);
1332 NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
1340 static bool gc_worker_skip_ct(const struct nf_conn *ct)
1342 return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
1345 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
1347 const struct nf_conntrack_l4proto *l4proto;
1349 if (!test_bit(IPS_ASSURED_BIT, &ct->status))
1352 l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1353 if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
1359 static void gc_worker(struct work_struct *work)
1361 unsigned int min_interval = max(HZ / GC_MAX_BUCKETS_DIV, 1u);
1362 unsigned int i, goal, buckets = 0, expired_count = 0;
1363 unsigned int nf_conntrack_max95 = 0;
1364 struct conntrack_gc_work *gc_work;
1365 unsigned int ratio, scanned = 0;
1366 unsigned long next_run;
1368 gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
1370 goal = nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV;
1371 i = gc_work->last_bucket;
1372 if (gc_work->early_drop)
1373 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
1376 struct nf_conntrack_tuple_hash *h;
1377 struct hlist_nulls_head *ct_hash;
1378 struct hlist_nulls_node *n;
1379 unsigned int hashsz;
1380 struct nf_conn *tmp;
1385 nf_conntrack_get_ht(&ct_hash, &hashsz);
1389 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1390 struct nf_conntrack_net *cnet;
1393 tmp = nf_ct_tuplehash_to_ctrack(h);
1396 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status)) {
1397 nf_ct_offload_timeout(tmp);
1401 if (nf_ct_is_expired(tmp)) {
1402 nf_ct_gc_expired(tmp);
1407 if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1410 net = nf_ct_net(tmp);
1411 cnet = nf_ct_pernet(net);
1412 if (atomic_read(&cnet->count) < nf_conntrack_max95)
1415 /* need to take reference to avoid possible races */
1416 if (!atomic_inc_not_zero(&tmp->ct_general.use))
1419 if (gc_worker_skip_ct(tmp)) {
1424 if (gc_worker_can_early_drop(tmp))
1430 /* could check get_nulls_value() here and restart if ct
1431 * was moved to another chain. But given gc is best-effort
1432 * we will just continue with next hash slot.
1436 } while (++buckets < goal);
1438 if (gc_work->exiting)
1442 * Eviction will normally happen from the packet path, and not
1443 * from this gc worker.
1445 * This worker is only here to reap expired entries when system went
1446 * idle after a busy period.
1448 * The heuristics below are supposed to balance conflicting goals:
1450 * 1. Minimize time until we notice a stale entry
1451 * 2. Maximize scan intervals to not waste cycles
1453 * Normally, expire ratio will be close to 0.
1455 * As soon as a sizeable fraction of the entries have expired
1456 * increase scan frequency.
1458 ratio = scanned ? expired_count * 100 / scanned : 0;
1459 if (ratio > GC_EVICT_RATIO) {
1460 gc_work->next_gc_run = min_interval;
1462 unsigned int max = GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV;
1464 BUILD_BUG_ON((GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV) == 0);
1466 gc_work->next_gc_run += min_interval;
1467 if (gc_work->next_gc_run > max)
1468 gc_work->next_gc_run = max;
1471 next_run = gc_work->next_gc_run;
1472 gc_work->last_bucket = i;
1473 gc_work->early_drop = false;
1474 queue_delayed_work(system_power_efficient_wq, &gc_work->dwork, next_run);
1477 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1479 INIT_DEFERRABLE_WORK(&gc_work->dwork, gc_worker);
1480 gc_work->next_gc_run = HZ;
1481 gc_work->exiting = false;
1484 static struct nf_conn *
1485 __nf_conntrack_alloc(struct net *net,
1486 const struct nf_conntrack_zone *zone,
1487 const struct nf_conntrack_tuple *orig,
1488 const struct nf_conntrack_tuple *repl,
1489 gfp_t gfp, u32 hash)
1491 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
1492 unsigned int ct_count;
1495 /* We don't want any race condition at early drop stage */
1496 ct_count = atomic_inc_return(&cnet->count);
1498 if (nf_conntrack_max && unlikely(ct_count > nf_conntrack_max)) {
1499 if (!early_drop(net, hash)) {
1500 if (!conntrack_gc_work.early_drop)
1501 conntrack_gc_work.early_drop = true;
1502 atomic_dec(&cnet->count);
1503 net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1504 return ERR_PTR(-ENOMEM);
1509 * Do not use kmem_cache_zalloc(), as this cache uses
1510 * SLAB_TYPESAFE_BY_RCU.
1512 ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1516 spin_lock_init(&ct->lock);
1517 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1518 ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1519 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1520 /* save hash for reusing when confirming */
1521 *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1524 write_pnet(&ct->ct_net, net);
1525 memset(&ct->__nfct_init_offset, 0,
1526 offsetof(struct nf_conn, proto) -
1527 offsetof(struct nf_conn, __nfct_init_offset));
1529 nf_ct_zone_add(ct, zone);
1531 /* Because we use RCU lookups, we set ct_general.use to zero before
1532 * this is inserted in any list.
1534 atomic_set(&ct->ct_general.use, 0);
1537 atomic_dec(&cnet->count);
1538 return ERR_PTR(-ENOMEM);
1541 struct nf_conn *nf_conntrack_alloc(struct net *net,
1542 const struct nf_conntrack_zone *zone,
1543 const struct nf_conntrack_tuple *orig,
1544 const struct nf_conntrack_tuple *repl,
1547 return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1549 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1551 void nf_conntrack_free(struct nf_conn *ct)
1553 struct net *net = nf_ct_net(ct);
1554 struct nf_conntrack_net *cnet;
1556 /* A freed object has refcnt == 0, that's
1557 * the golden rule for SLAB_TYPESAFE_BY_RCU
1559 WARN_ON(atomic_read(&ct->ct_general.use) != 0);
1561 nf_ct_ext_destroy(ct);
1562 kmem_cache_free(nf_conntrack_cachep, ct);
1563 cnet = nf_ct_pernet(net);
1565 smp_mb__before_atomic();
1566 atomic_dec(&cnet->count);
1568 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1571 /* Allocate a new conntrack: we return -ENOMEM if classification
1572 failed due to stress. Otherwise it really is unclassifiable. */
1573 static noinline struct nf_conntrack_tuple_hash *
1574 init_conntrack(struct net *net, struct nf_conn *tmpl,
1575 const struct nf_conntrack_tuple *tuple,
1576 struct sk_buff *skb,
1577 unsigned int dataoff, u32 hash)
1580 struct nf_conn_help *help;
1581 struct nf_conntrack_tuple repl_tuple;
1582 struct nf_conntrack_ecache *ecache;
1583 struct nf_conntrack_expect *exp = NULL;
1584 const struct nf_conntrack_zone *zone;
1585 struct nf_conn_timeout *timeout_ext;
1586 struct nf_conntrack_zone tmp;
1587 struct nf_conntrack_net *cnet;
1589 if (!nf_ct_invert_tuple(&repl_tuple, tuple)) {
1590 pr_debug("Can't invert tuple.\n");
1594 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1595 ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1598 return (struct nf_conntrack_tuple_hash *)ct;
1600 if (!nf_ct_add_synproxy(ct, tmpl)) {
1601 nf_conntrack_free(ct);
1602 return ERR_PTR(-ENOMEM);
1605 timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1608 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1611 nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1612 nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1613 nf_ct_labels_ext_add(ct);
1615 ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1616 nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1617 ecache ? ecache->expmask : 0,
1621 cnet = nf_ct_pernet(net);
1622 if (cnet->expect_count) {
1623 spin_lock(&nf_conntrack_expect_lock);
1624 exp = nf_ct_find_expectation(net, zone, tuple);
1626 pr_debug("expectation arrives ct=%p exp=%p\n",
1628 /* Welcome, Mr. Bond. We've been expecting you... */
1629 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1630 /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1631 ct->master = exp->master;
1633 help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1635 rcu_assign_pointer(help->helper, exp->helper);
1638 #ifdef CONFIG_NF_CONNTRACK_MARK
1639 ct->mark = exp->master->mark;
1641 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1642 ct->secmark = exp->master->secmark;
1644 NF_CT_STAT_INC(net, expect_new);
1646 spin_unlock(&nf_conntrack_expect_lock);
1649 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1651 /* Now it is inserted into the unconfirmed list, bump refcount */
1652 nf_conntrack_get(&ct->ct_general);
1653 nf_ct_add_to_unconfirmed_list(ct);
1659 exp->expectfn(ct, exp);
1660 nf_ct_expect_put(exp);
1663 return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1666 /* On success, returns 0, sets skb->_nfct | ctinfo */
1668 resolve_normal_ct(struct nf_conn *tmpl,
1669 struct sk_buff *skb,
1670 unsigned int dataoff,
1672 const struct nf_hook_state *state)
1674 const struct nf_conntrack_zone *zone;
1675 struct nf_conntrack_tuple tuple;
1676 struct nf_conntrack_tuple_hash *h;
1677 enum ip_conntrack_info ctinfo;
1678 struct nf_conntrack_zone tmp;
1682 if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1683 dataoff, state->pf, protonum, state->net,
1685 pr_debug("Can't get tuple\n");
1689 /* look for tuple match */
1690 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1691 hash = hash_conntrack_raw(&tuple, state->net);
1692 h = __nf_conntrack_find_get(state->net, zone, &tuple, hash);
1694 h = init_conntrack(state->net, tmpl, &tuple,
1695 skb, dataoff, hash);
1701 ct = nf_ct_tuplehash_to_ctrack(h);
1703 /* It exists; we have (non-exclusive) reference. */
1704 if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1705 ctinfo = IP_CT_ESTABLISHED_REPLY;
1707 /* Once we've had two way comms, always ESTABLISHED. */
1708 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1709 pr_debug("normal packet for %p\n", ct);
1710 ctinfo = IP_CT_ESTABLISHED;
1711 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1712 pr_debug("related packet for %p\n", ct);
1713 ctinfo = IP_CT_RELATED;
1715 pr_debug("new packet for %p\n", ct);
1719 nf_ct_set(skb, ct, ctinfo);
1724 * icmp packets need special treatment to handle error messages that are
1725 * related to a connection.
1727 * Callers need to check if skb has a conntrack assigned when this
1728 * helper returns; in such case skb belongs to an already known connection.
1730 static unsigned int __cold
1731 nf_conntrack_handle_icmp(struct nf_conn *tmpl,
1732 struct sk_buff *skb,
1733 unsigned int dataoff,
1735 const struct nf_hook_state *state)
1739 if (state->pf == NFPROTO_IPV4 && protonum == IPPROTO_ICMP)
1740 ret = nf_conntrack_icmpv4_error(tmpl, skb, dataoff, state);
1741 #if IS_ENABLED(CONFIG_IPV6)
1742 else if (state->pf == NFPROTO_IPV6 && protonum == IPPROTO_ICMPV6)
1743 ret = nf_conntrack_icmpv6_error(tmpl, skb, dataoff, state);
1749 NF_CT_STAT_INC_ATOMIC(state->net, error);
1754 static int generic_packet(struct nf_conn *ct, struct sk_buff *skb,
1755 enum ip_conntrack_info ctinfo)
1757 const unsigned int *timeout = nf_ct_timeout_lookup(ct);
1760 timeout = &nf_generic_pernet(nf_ct_net(ct))->timeout;
1762 nf_ct_refresh_acct(ct, ctinfo, skb, *timeout);
1766 /* Returns verdict for packet, or -1 for invalid. */
1767 static int nf_conntrack_handle_packet(struct nf_conn *ct,
1768 struct sk_buff *skb,
1769 unsigned int dataoff,
1770 enum ip_conntrack_info ctinfo,
1771 const struct nf_hook_state *state)
1773 switch (nf_ct_protonum(ct)) {
1775 return nf_conntrack_tcp_packet(ct, skb, dataoff,
1778 return nf_conntrack_udp_packet(ct, skb, dataoff,
1781 return nf_conntrack_icmp_packet(ct, skb, ctinfo, state);
1782 #if IS_ENABLED(CONFIG_IPV6)
1783 case IPPROTO_ICMPV6:
1784 return nf_conntrack_icmpv6_packet(ct, skb, ctinfo, state);
1786 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
1787 case IPPROTO_UDPLITE:
1788 return nf_conntrack_udplite_packet(ct, skb, dataoff,
1791 #ifdef CONFIG_NF_CT_PROTO_SCTP
1793 return nf_conntrack_sctp_packet(ct, skb, dataoff,
1796 #ifdef CONFIG_NF_CT_PROTO_DCCP
1798 return nf_conntrack_dccp_packet(ct, skb, dataoff,
1801 #ifdef CONFIG_NF_CT_PROTO_GRE
1803 return nf_conntrack_gre_packet(ct, skb, dataoff,
1808 return generic_packet(ct, skb, ctinfo);
1812 nf_conntrack_in(struct sk_buff *skb, const struct nf_hook_state *state)
1814 enum ip_conntrack_info ctinfo;
1815 struct nf_conn *ct, *tmpl;
1819 tmpl = nf_ct_get(skb, &ctinfo);
1820 if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1821 /* Previously seen (loopback or untracked)? Ignore. */
1822 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1823 ctinfo == IP_CT_UNTRACKED)
1828 /* rcu_read_lock()ed by nf_hook_thresh */
1829 dataoff = get_l4proto(skb, skb_network_offset(skb), state->pf, &protonum);
1831 pr_debug("not prepared to track yet or error occurred\n");
1832 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1837 if (protonum == IPPROTO_ICMP || protonum == IPPROTO_ICMPV6) {
1838 ret = nf_conntrack_handle_icmp(tmpl, skb, dataoff,
1844 /* ICMP[v6] protocol trackers may assign one conntrack. */
1849 ret = resolve_normal_ct(tmpl, skb, dataoff,
1852 /* Too stressed to deal. */
1853 NF_CT_STAT_INC_ATOMIC(state->net, drop);
1858 ct = nf_ct_get(skb, &ctinfo);
1860 /* Not valid part of a connection */
1861 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1866 ret = nf_conntrack_handle_packet(ct, skb, dataoff, ctinfo, state);
1868 /* Invalid: inverse of the return code tells
1869 * the netfilter core what to do */
1870 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1871 nf_conntrack_put(&ct->ct_general);
1873 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1874 if (ret == -NF_DROP)
1875 NF_CT_STAT_INC_ATOMIC(state->net, drop);
1876 /* Special case: TCP tracker reports an attempt to reopen a
1877 * closed/aborted connection. We have to go back and create a
1880 if (ret == -NF_REPEAT)
1886 if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
1887 !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1888 nf_conntrack_event_cache(IPCT_REPLY, ct);
1895 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1897 /* Alter reply tuple (maybe alter helper). This is for NAT, and is
1898 implicitly racy: see __nf_conntrack_confirm */
1899 void nf_conntrack_alter_reply(struct nf_conn *ct,
1900 const struct nf_conntrack_tuple *newreply)
1902 struct nf_conn_help *help = nfct_help(ct);
1904 /* Should be unconfirmed, so not in hash table yet */
1905 WARN_ON(nf_ct_is_confirmed(ct));
1907 pr_debug("Altering reply tuple of %p to ", ct);
1908 nf_ct_dump_tuple(newreply);
1910 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1911 if (ct->master || (help && !hlist_empty(&help->expectations)))
1915 __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1918 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1920 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1921 void __nf_ct_refresh_acct(struct nf_conn *ct,
1922 enum ip_conntrack_info ctinfo,
1923 const struct sk_buff *skb,
1927 /* Only update if this is not a fixed timeout */
1928 if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1931 /* If not in hash table, timer will not be active yet */
1932 if (nf_ct_is_confirmed(ct))
1933 extra_jiffies += nfct_time_stamp;
1935 if (READ_ONCE(ct->timeout) != extra_jiffies)
1936 WRITE_ONCE(ct->timeout, extra_jiffies);
1939 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
1941 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1943 bool nf_ct_kill_acct(struct nf_conn *ct,
1944 enum ip_conntrack_info ctinfo,
1945 const struct sk_buff *skb)
1947 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
1949 return nf_ct_delete(ct, 0, 0);
1951 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
1953 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1955 #include <linux/netfilter/nfnetlink.h>
1956 #include <linux/netfilter/nfnetlink_conntrack.h>
1957 #include <linux/mutex.h>
1959 /* Generic function for tcp/udp/sctp/dccp and alike. */
1960 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1961 const struct nf_conntrack_tuple *tuple)
1963 if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1964 nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1965 goto nla_put_failure;
1971 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1973 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1974 [CTA_PROTO_SRC_PORT] = { .type = NLA_U16 },
1975 [CTA_PROTO_DST_PORT] = { .type = NLA_U16 },
1977 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1979 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1980 struct nf_conntrack_tuple *t,
1983 if (flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) {
1984 if (!tb[CTA_PROTO_SRC_PORT])
1987 t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1990 if (flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) {
1991 if (!tb[CTA_PROTO_DST_PORT])
1994 t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1999 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
2001 unsigned int nf_ct_port_nlattr_tuple_size(void)
2003 static unsigned int size __read_mostly;
2006 size = nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
2010 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
2013 /* Used by ipt_REJECT and ip6t_REJECT. */
2014 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
2017 enum ip_conntrack_info ctinfo;
2019 /* This ICMP is in reverse direction to the packet which caused it */
2020 ct = nf_ct_get(skb, &ctinfo);
2021 if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
2022 ctinfo = IP_CT_RELATED_REPLY;
2024 ctinfo = IP_CT_RELATED;
2026 /* Attach to new skbuff, and increment count */
2027 nf_ct_set(nskb, ct, ctinfo);
2028 nf_conntrack_get(skb_nfct(nskb));
2031 static int __nf_conntrack_update(struct net *net, struct sk_buff *skb,
2033 enum ip_conntrack_info ctinfo)
2035 struct nf_conntrack_tuple_hash *h;
2036 struct nf_conntrack_tuple tuple;
2037 struct nf_nat_hook *nat_hook;
2038 unsigned int status;
2043 l3num = nf_ct_l3num(ct);
2045 dataoff = get_l4proto(skb, skb_network_offset(skb), l3num, &l4num);
2049 if (!nf_ct_get_tuple(skb, skb_network_offset(skb), dataoff, l3num,
2050 l4num, net, &tuple))
2053 if (ct->status & IPS_SRC_NAT) {
2054 memcpy(tuple.src.u3.all,
2055 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.all,
2056 sizeof(tuple.src.u3.all));
2058 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u.all;
2061 if (ct->status & IPS_DST_NAT) {
2062 memcpy(tuple.dst.u3.all,
2063 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.all,
2064 sizeof(tuple.dst.u3.all));
2066 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u.all;
2069 h = nf_conntrack_find_get(net, nf_ct_zone(ct), &tuple);
2073 /* Store status bits of the conntrack that is clashing to re-do NAT
2074 * mangling according to what it has been done already to this packet.
2076 status = ct->status;
2079 ct = nf_ct_tuplehash_to_ctrack(h);
2080 nf_ct_set(skb, ct, ctinfo);
2082 nat_hook = rcu_dereference(nf_nat_hook);
2086 if (status & IPS_SRC_NAT &&
2087 nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_SRC,
2088 IP_CT_DIR_ORIGINAL) == NF_DROP)
2091 if (status & IPS_DST_NAT &&
2092 nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_DST,
2093 IP_CT_DIR_ORIGINAL) == NF_DROP)
2099 /* This packet is coming from userspace via nf_queue, complete the packet
2100 * processing after the helper invocation in nf_confirm().
2102 static int nf_confirm_cthelper(struct sk_buff *skb, struct nf_conn *ct,
2103 enum ip_conntrack_info ctinfo)
2105 const struct nf_conntrack_helper *helper;
2106 const struct nf_conn_help *help;
2109 help = nfct_help(ct);
2113 helper = rcu_dereference(help->helper);
2114 if (!(helper->flags & NF_CT_HELPER_F_USERSPACE))
2117 switch (nf_ct_l3num(ct)) {
2119 protoff = skb_network_offset(skb) + ip_hdrlen(skb);
2121 #if IS_ENABLED(CONFIG_IPV6)
2122 case NFPROTO_IPV6: {
2126 pnum = ipv6_hdr(skb)->nexthdr;
2127 protoff = ipv6_skip_exthdr(skb, sizeof(struct ipv6hdr), &pnum,
2129 if (protoff < 0 || (frag_off & htons(~0x7)) != 0)
2138 if (test_bit(IPS_SEQ_ADJUST_BIT, &ct->status) &&
2139 !nf_is_loopback_packet(skb)) {
2140 if (!nf_ct_seq_adjust(skb, ct, ctinfo, protoff)) {
2141 NF_CT_STAT_INC_ATOMIC(nf_ct_net(ct), drop);
2146 /* We've seen it coming out the other side: confirm it */
2147 return nf_conntrack_confirm(skb) == NF_DROP ? - 1 : 0;
2150 static int nf_conntrack_update(struct net *net, struct sk_buff *skb)
2152 enum ip_conntrack_info ctinfo;
2156 ct = nf_ct_get(skb, &ctinfo);
2160 if (!nf_ct_is_confirmed(ct)) {
2161 err = __nf_conntrack_update(net, skb, ct, ctinfo);
2165 ct = nf_ct_get(skb, &ctinfo);
2168 return nf_confirm_cthelper(skb, ct, ctinfo);
2171 static bool nf_conntrack_get_tuple_skb(struct nf_conntrack_tuple *dst_tuple,
2172 const struct sk_buff *skb)
2174 const struct nf_conntrack_tuple *src_tuple;
2175 const struct nf_conntrack_tuple_hash *hash;
2176 struct nf_conntrack_tuple srctuple;
2177 enum ip_conntrack_info ctinfo;
2180 ct = nf_ct_get(skb, &ctinfo);
2182 src_tuple = nf_ct_tuple(ct, CTINFO2DIR(ctinfo));
2183 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2187 if (!nf_ct_get_tuplepr(skb, skb_network_offset(skb),
2188 NFPROTO_IPV4, dev_net(skb->dev),
2192 hash = nf_conntrack_find_get(dev_net(skb->dev),
2198 ct = nf_ct_tuplehash_to_ctrack(hash);
2199 src_tuple = nf_ct_tuple(ct, !hash->tuple.dst.dir);
2200 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2206 /* Bring out ya dead! */
2207 static struct nf_conn *
2208 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
2209 void *data, unsigned int *bucket)
2211 struct nf_conntrack_tuple_hash *h;
2213 struct hlist_nulls_node *n;
2216 for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
2217 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
2219 nf_conntrack_lock(lockp);
2220 if (*bucket < nf_conntrack_htable_size) {
2221 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
2222 if (NF_CT_DIRECTION(h) != IP_CT_DIR_REPLY)
2224 /* All nf_conn objects are added to hash table twice, one
2225 * for original direction tuple, once for the reply tuple.
2227 * Exception: In the IPS_NAT_CLASH case, only the reply
2228 * tuple is added (the original tuple already existed for
2229 * a different object).
2231 * We only need to call the iterator once for each
2232 * conntrack, so we just use the 'reply' direction
2233 * tuple while iterating.
2235 ct = nf_ct_tuplehash_to_ctrack(h);
2247 atomic_inc(&ct->ct_general.use);
2253 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
2254 void *data, u32 portid, int report)
2256 unsigned int bucket = 0, sequence;
2262 sequence = read_seqcount_begin(&nf_conntrack_generation);
2264 while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
2265 /* Time to push up daises... */
2267 nf_ct_delete(ct, portid, report);
2272 if (!read_seqcount_retry(&nf_conntrack_generation, sequence))
2279 int (*iter)(struct nf_conn *i, void *data);
2284 static int iter_net_only(struct nf_conn *i, void *data)
2286 struct iter_data *d = data;
2288 if (!net_eq(d->net, nf_ct_net(i)))
2291 return d->iter(i, d->data);
2295 __nf_ct_unconfirmed_destroy(struct net *net)
2299 for_each_possible_cpu(cpu) {
2300 struct nf_conntrack_tuple_hash *h;
2301 struct hlist_nulls_node *n;
2302 struct ct_pcpu *pcpu;
2304 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2306 spin_lock_bh(&pcpu->lock);
2307 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
2310 ct = nf_ct_tuplehash_to_ctrack(h);
2312 /* we cannot call iter() on unconfirmed list, the
2313 * owning cpu can reallocate ct->ext at any time.
2315 set_bit(IPS_DYING_BIT, &ct->status);
2317 spin_unlock_bh(&pcpu->lock);
2322 void nf_ct_unconfirmed_destroy(struct net *net)
2324 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2328 if (atomic_read(&cnet->count) > 0) {
2329 __nf_ct_unconfirmed_destroy(net);
2330 nf_queue_nf_hook_drop(net);
2334 EXPORT_SYMBOL_GPL(nf_ct_unconfirmed_destroy);
2336 void nf_ct_iterate_cleanup_net(struct net *net,
2337 int (*iter)(struct nf_conn *i, void *data),
2338 void *data, u32 portid, int report)
2340 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2345 if (atomic_read(&cnet->count) == 0)
2352 nf_ct_iterate_cleanup(iter_net_only, &d, portid, report);
2354 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
2357 * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
2358 * @iter: callback to invoke for each conntrack
2359 * @data: data to pass to @iter
2361 * Like nf_ct_iterate_cleanup, but first marks conntracks on the
2362 * unconfirmed list as dying (so they will not be inserted into
2365 * Can only be called in module exit path.
2368 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
2372 down_read(&net_rwsem);
2374 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2376 if (atomic_read(&cnet->count) == 0)
2378 __nf_ct_unconfirmed_destroy(net);
2379 nf_queue_nf_hook_drop(net);
2381 up_read(&net_rwsem);
2383 /* Need to wait for netns cleanup worker to finish, if its
2384 * running -- it might have deleted a net namespace from
2385 * the global list, so our __nf_ct_unconfirmed_destroy() might
2386 * not have affected all namespaces.
2390 /* a conntrack could have been unlinked from unconfirmed list
2391 * before we grabbed pcpu lock in __nf_ct_unconfirmed_destroy().
2392 * This makes sure its inserted into conntrack table.
2396 nf_ct_iterate_cleanup(iter, data, 0, 0);
2398 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
2400 static int kill_all(struct nf_conn *i, void *data)
2402 return net_eq(nf_ct_net(i), data);
2405 void nf_conntrack_cleanup_start(void)
2407 conntrack_gc_work.exiting = true;
2408 RCU_INIT_POINTER(ip_ct_attach, NULL);
2411 void nf_conntrack_cleanup_end(void)
2413 RCU_INIT_POINTER(nf_ct_hook, NULL);
2414 cancel_delayed_work_sync(&conntrack_gc_work.dwork);
2415 kvfree(nf_conntrack_hash);
2417 nf_conntrack_proto_fini();
2418 nf_conntrack_seqadj_fini();
2419 nf_conntrack_labels_fini();
2420 nf_conntrack_helper_fini();
2421 nf_conntrack_timeout_fini();
2422 nf_conntrack_ecache_fini();
2423 nf_conntrack_tstamp_fini();
2424 nf_conntrack_acct_fini();
2425 nf_conntrack_expect_fini();
2427 kmem_cache_destroy(nf_conntrack_cachep);
2431 * Mishearing the voices in his head, our hero wonders how he's
2432 * supposed to kill the mall.
2434 void nf_conntrack_cleanup_net(struct net *net)
2438 list_add(&net->exit_list, &single);
2439 nf_conntrack_cleanup_net_list(&single);
2442 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
2448 * This makes sure all current packets have passed through
2449 * netfilter framework. Roll on, two-stage module
2455 list_for_each_entry(net, net_exit_list, exit_list) {
2456 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2458 nf_ct_iterate_cleanup(kill_all, net, 0, 0);
2459 if (atomic_read(&cnet->count) != 0)
2464 goto i_see_dead_people;
2467 list_for_each_entry(net, net_exit_list, exit_list) {
2468 nf_conntrack_ecache_pernet_fini(net);
2469 nf_conntrack_expect_pernet_fini(net);
2470 free_percpu(net->ct.stat);
2471 free_percpu(net->ct.pcpu_lists);
2475 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
2477 struct hlist_nulls_head *hash;
2478 unsigned int nr_slots, i;
2480 if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
2483 BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
2484 nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
2486 hash = kvcalloc(nr_slots, sizeof(struct hlist_nulls_head), GFP_KERNEL);
2489 for (i = 0; i < nr_slots; i++)
2490 INIT_HLIST_NULLS_HEAD(&hash[i], i);
2494 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
2496 int nf_conntrack_hash_resize(unsigned int hashsize)
2499 unsigned int old_size;
2500 struct hlist_nulls_head *hash, *old_hash;
2501 struct nf_conntrack_tuple_hash *h;
2507 hash = nf_ct_alloc_hashtable(&hashsize, 1);
2511 old_size = nf_conntrack_htable_size;
2512 if (old_size == hashsize) {
2518 nf_conntrack_all_lock();
2519 write_seqcount_begin(&nf_conntrack_generation);
2521 /* Lookups in the old hash might happen in parallel, which means we
2522 * might get false negatives during connection lookup. New connections
2523 * created because of a false negative won't make it into the hash
2524 * though since that required taking the locks.
2527 for (i = 0; i < nf_conntrack_htable_size; i++) {
2528 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
2529 h = hlist_nulls_entry(nf_conntrack_hash[i].first,
2530 struct nf_conntrack_tuple_hash, hnnode);
2531 ct = nf_ct_tuplehash_to_ctrack(h);
2532 hlist_nulls_del_rcu(&h->hnnode);
2533 bucket = __hash_conntrack(nf_ct_net(ct),
2534 &h->tuple, hashsize);
2535 hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
2538 old_size = nf_conntrack_htable_size;
2539 old_hash = nf_conntrack_hash;
2541 nf_conntrack_hash = hash;
2542 nf_conntrack_htable_size = hashsize;
2544 write_seqcount_end(&nf_conntrack_generation);
2545 nf_conntrack_all_unlock();
2553 int nf_conntrack_set_hashsize(const char *val, const struct kernel_param *kp)
2555 unsigned int hashsize;
2558 if (current->nsproxy->net_ns != &init_net)
2561 /* On boot, we can set this without any fancy locking. */
2562 if (!nf_conntrack_hash)
2563 return param_set_uint(val, kp);
2565 rc = kstrtouint(val, 0, &hashsize);
2569 return nf_conntrack_hash_resize(hashsize);
2572 static __always_inline unsigned int total_extension_size(void)
2574 /* remember to add new extensions below */
2575 BUILD_BUG_ON(NF_CT_EXT_NUM > 9);
2577 return sizeof(struct nf_ct_ext) +
2578 sizeof(struct nf_conn_help)
2579 #if IS_ENABLED(CONFIG_NF_NAT)
2580 + sizeof(struct nf_conn_nat)
2582 + sizeof(struct nf_conn_seqadj)
2583 + sizeof(struct nf_conn_acct)
2584 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2585 + sizeof(struct nf_conntrack_ecache)
2587 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
2588 + sizeof(struct nf_conn_tstamp)
2590 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
2591 + sizeof(struct nf_conn_timeout)
2593 #ifdef CONFIG_NF_CONNTRACK_LABELS
2594 + sizeof(struct nf_conn_labels)
2596 #if IS_ENABLED(CONFIG_NETFILTER_SYNPROXY)
2597 + sizeof(struct nf_conn_synproxy)
2602 int nf_conntrack_init_start(void)
2604 unsigned long nr_pages = totalram_pages();
2609 /* struct nf_ct_ext uses u8 to store offsets/size */
2610 BUILD_BUG_ON(total_extension_size() > 255u);
2612 seqcount_spinlock_init(&nf_conntrack_generation,
2613 &nf_conntrack_locks_all_lock);
2615 for (i = 0; i < CONNTRACK_LOCKS; i++)
2616 spin_lock_init(&nf_conntrack_locks[i]);
2618 if (!nf_conntrack_htable_size) {
2619 /* Idea from tcp.c: use 1/16384 of memory.
2620 * On i386: 32MB machine has 512 buckets.
2621 * >= 1GB machines have 16384 buckets.
2622 * >= 4GB machines have 65536 buckets.
2624 nf_conntrack_htable_size
2625 = (((nr_pages << PAGE_SHIFT) / 16384)
2626 / sizeof(struct hlist_head));
2627 if (nr_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2628 nf_conntrack_htable_size = 65536;
2629 else if (nr_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2630 nf_conntrack_htable_size = 16384;
2631 if (nf_conntrack_htable_size < 32)
2632 nf_conntrack_htable_size = 32;
2634 /* Use a max. factor of four by default to get the same max as
2635 * with the old struct list_heads. When a table size is given
2636 * we use the old value of 8 to avoid reducing the max.
2641 nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2642 if (!nf_conntrack_hash)
2645 nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2647 nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2648 sizeof(struct nf_conn),
2650 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2651 if (!nf_conntrack_cachep)
2654 ret = nf_conntrack_expect_init();
2658 ret = nf_conntrack_acct_init();
2662 ret = nf_conntrack_tstamp_init();
2666 ret = nf_conntrack_ecache_init();
2670 ret = nf_conntrack_timeout_init();
2674 ret = nf_conntrack_helper_init();
2678 ret = nf_conntrack_labels_init();
2682 ret = nf_conntrack_seqadj_init();
2686 ret = nf_conntrack_proto_init();
2690 conntrack_gc_work_init(&conntrack_gc_work);
2691 queue_delayed_work(system_power_efficient_wq, &conntrack_gc_work.dwork, HZ);
2696 nf_conntrack_seqadj_fini();
2698 nf_conntrack_labels_fini();
2700 nf_conntrack_helper_fini();
2702 nf_conntrack_timeout_fini();
2704 nf_conntrack_ecache_fini();
2706 nf_conntrack_tstamp_fini();
2708 nf_conntrack_acct_fini();
2710 nf_conntrack_expect_fini();
2712 kmem_cache_destroy(nf_conntrack_cachep);
2714 kvfree(nf_conntrack_hash);
2718 static struct nf_ct_hook nf_conntrack_hook = {
2719 .update = nf_conntrack_update,
2720 .destroy = destroy_conntrack,
2721 .get_tuple_skb = nf_conntrack_get_tuple_skb,
2724 void nf_conntrack_init_end(void)
2726 /* For use by REJECT target */
2727 RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
2728 RCU_INIT_POINTER(nf_ct_hook, &nf_conntrack_hook);
2732 * We need to use special "null" values, not used in hash table
2734 #define UNCONFIRMED_NULLS_VAL ((1<<30)+0)
2735 #define DYING_NULLS_VAL ((1<<30)+1)
2737 int nf_conntrack_init_net(struct net *net)
2739 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2743 BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2744 BUILD_BUG_ON_NOT_POWER_OF_2(CONNTRACK_LOCKS);
2745 atomic_set(&cnet->count, 0);
2747 net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
2748 if (!net->ct.pcpu_lists)
2751 for_each_possible_cpu(cpu) {
2752 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2754 spin_lock_init(&pcpu->lock);
2755 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
2756 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
2759 net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2761 goto err_pcpu_lists;
2763 ret = nf_conntrack_expect_pernet_init(net);
2767 nf_conntrack_acct_pernet_init(net);
2768 nf_conntrack_tstamp_pernet_init(net);
2769 nf_conntrack_ecache_pernet_init(net);
2770 nf_conntrack_helper_pernet_init(net);
2771 nf_conntrack_proto_pernet_init(net);
2776 free_percpu(net->ct.stat);
2778 free_percpu(net->ct.pcpu_lists);