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 nf_conntrack_locks_all = true;
154 for (i = 0; i < CONNTRACK_LOCKS; i++) {
155 spin_lock(&nf_conntrack_locks[i]);
157 /* This spin_unlock provides the "release" to ensure that
158 * nf_conntrack_locks_all==true is visible to everyone that
159 * acquired spin_lock(&nf_conntrack_locks[]).
161 spin_unlock(&nf_conntrack_locks[i]);
165 static void nf_conntrack_all_unlock(void)
166 __releases(&nf_conntrack_locks_all_lock)
168 /* All prior stores must be complete before we clear
169 * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
170 * might observe the false value but not the entire
172 * It pairs with the smp_load_acquire() in nf_conntrack_lock()
174 smp_store_release(&nf_conntrack_locks_all, false);
175 spin_unlock(&nf_conntrack_locks_all_lock);
178 unsigned int nf_conntrack_htable_size __read_mostly;
179 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
181 unsigned int nf_conntrack_max __read_mostly;
182 EXPORT_SYMBOL_GPL(nf_conntrack_max);
183 seqcount_spinlock_t nf_conntrack_generation __read_mostly;
184 static unsigned int nf_conntrack_hash_rnd __read_mostly;
186 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
187 const struct net *net)
192 get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
194 /* The direction must be ignored, so we hash everything up to the
195 * destination ports (which is a multiple of 4) and treat the last
196 * three bytes manually.
198 seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
199 n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
200 return jhash2((u32 *)tuple, n, seed ^
201 (((__force __u16)tuple->dst.u.all << 16) |
202 tuple->dst.protonum));
205 static u32 scale_hash(u32 hash)
207 return reciprocal_scale(hash, nf_conntrack_htable_size);
210 static u32 __hash_conntrack(const struct net *net,
211 const struct nf_conntrack_tuple *tuple,
214 return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
217 static u32 hash_conntrack(const struct net *net,
218 const struct nf_conntrack_tuple *tuple)
220 return scale_hash(hash_conntrack_raw(tuple, net));
223 static bool nf_ct_get_tuple_ports(const struct sk_buff *skb,
224 unsigned int dataoff,
225 struct nf_conntrack_tuple *tuple)
229 } _inet_hdr, *inet_hdr;
231 /* Actually only need first 4 bytes to get ports. */
232 inet_hdr = skb_header_pointer(skb, dataoff, sizeof(_inet_hdr), &_inet_hdr);
236 tuple->src.u.udp.port = inet_hdr->sport;
237 tuple->dst.u.udp.port = inet_hdr->dport;
242 nf_ct_get_tuple(const struct sk_buff *skb,
244 unsigned int dataoff,
248 struct nf_conntrack_tuple *tuple)
254 memset(tuple, 0, sizeof(*tuple));
256 tuple->src.l3num = l3num;
259 nhoff += offsetof(struct iphdr, saddr);
260 size = 2 * sizeof(__be32);
263 nhoff += offsetof(struct ipv6hdr, saddr);
264 size = sizeof(_addrs);
270 ap = skb_header_pointer(skb, nhoff, size, _addrs);
276 tuple->src.u3.ip = ap[0];
277 tuple->dst.u3.ip = ap[1];
280 memcpy(tuple->src.u3.ip6, ap, sizeof(tuple->src.u3.ip6));
281 memcpy(tuple->dst.u3.ip6, ap + 4, sizeof(tuple->dst.u3.ip6));
285 tuple->dst.protonum = protonum;
286 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
289 #if IS_ENABLED(CONFIG_IPV6)
291 return icmpv6_pkt_to_tuple(skb, dataoff, net, tuple);
294 return icmp_pkt_to_tuple(skb, dataoff, net, tuple);
295 #ifdef CONFIG_NF_CT_PROTO_GRE
297 return gre_pkt_to_tuple(skb, dataoff, net, tuple);
300 case IPPROTO_UDP: /* fallthrough */
301 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
302 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
303 case IPPROTO_UDPLITE:
304 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
306 #ifdef CONFIG_NF_CT_PROTO_SCTP
308 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
310 #ifdef CONFIG_NF_CT_PROTO_DCCP
312 return nf_ct_get_tuple_ports(skb, dataoff, tuple);
321 static int ipv4_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
325 const struct iphdr *iph;
328 iph = skb_header_pointer(skb, nhoff, sizeof(_iph), &_iph);
332 /* Conntrack defragments packets, we might still see fragments
333 * inside ICMP packets though.
335 if (iph->frag_off & htons(IP_OFFSET))
338 dataoff = nhoff + (iph->ihl << 2);
339 *protonum = iph->protocol;
341 /* Check bogus IP headers */
342 if (dataoff > skb->len) {
343 pr_debug("bogus IPv4 packet: nhoff %u, ihl %u, skblen %u\n",
344 nhoff, iph->ihl << 2, skb->len);
350 #if IS_ENABLED(CONFIG_IPV6)
351 static int ipv6_get_l4proto(const struct sk_buff *skb, unsigned int nhoff,
355 unsigned int extoff = nhoff + sizeof(struct ipv6hdr);
359 if (skb_copy_bits(skb, nhoff + offsetof(struct ipv6hdr, nexthdr),
360 &nexthdr, sizeof(nexthdr)) != 0) {
361 pr_debug("can't get nexthdr\n");
364 protoff = ipv6_skip_exthdr(skb, extoff, &nexthdr, &frag_off);
366 * (protoff == skb->len) means the packet has not data, just
367 * IPv6 and possibly extensions headers, but it is tracked anyway
369 if (protoff < 0 || (frag_off & htons(~0x7)) != 0) {
370 pr_debug("can't find proto in pkt\n");
379 static int get_l4proto(const struct sk_buff *skb,
380 unsigned int nhoff, u8 pf, u8 *l4num)
384 return ipv4_get_l4proto(skb, nhoff, l4num);
385 #if IS_ENABLED(CONFIG_IPV6)
387 return ipv6_get_l4proto(skb, nhoff, l4num);
396 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
398 struct net *net, struct nf_conntrack_tuple *tuple)
403 protoff = get_l4proto(skb, nhoff, l3num, &protonum);
407 return nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple);
409 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
412 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
413 const struct nf_conntrack_tuple *orig)
415 memset(inverse, 0, sizeof(*inverse));
417 inverse->src.l3num = orig->src.l3num;
419 switch (orig->src.l3num) {
421 inverse->src.u3.ip = orig->dst.u3.ip;
422 inverse->dst.u3.ip = orig->src.u3.ip;
425 inverse->src.u3.in6 = orig->dst.u3.in6;
426 inverse->dst.u3.in6 = orig->src.u3.in6;
432 inverse->dst.dir = !orig->dst.dir;
434 inverse->dst.protonum = orig->dst.protonum;
436 switch (orig->dst.protonum) {
438 return nf_conntrack_invert_icmp_tuple(inverse, orig);
439 #if IS_ENABLED(CONFIG_IPV6)
441 return nf_conntrack_invert_icmpv6_tuple(inverse, orig);
445 inverse->src.u.all = orig->dst.u.all;
446 inverse->dst.u.all = orig->src.u.all;
449 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
451 /* Generate a almost-unique pseudo-id for a given conntrack.
453 * intentionally doesn't re-use any of the seeds used for hash
454 * table location, we assume id gets exposed to userspace.
456 * Following nf_conn items do not change throughout lifetime
460 * 2. nf_conn->master address (normally NULL)
461 * 3. the associated net namespace
462 * 4. the original direction tuple
464 u32 nf_ct_get_id(const struct nf_conn *ct)
466 static __read_mostly siphash_key_t ct_id_seed;
467 unsigned long a, b, c, d;
469 net_get_random_once(&ct_id_seed, sizeof(ct_id_seed));
471 a = (unsigned long)ct;
472 b = (unsigned long)ct->master;
473 c = (unsigned long)nf_ct_net(ct);
474 d = (unsigned long)siphash(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
475 sizeof(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple),
478 return siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &ct_id_seed);
480 return siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &ct_id_seed);
483 EXPORT_SYMBOL_GPL(nf_ct_get_id);
486 clean_from_lists(struct nf_conn *ct)
488 pr_debug("clean_from_lists(%p)\n", ct);
489 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
490 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
492 /* Destroy all pending expectations */
493 nf_ct_remove_expectations(ct);
496 /* must be called with local_bh_disable */
497 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
499 struct ct_pcpu *pcpu;
501 /* add this conntrack to the (per cpu) dying list */
502 ct->cpu = smp_processor_id();
503 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
505 spin_lock(&pcpu->lock);
506 hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
508 spin_unlock(&pcpu->lock);
511 /* must be called with local_bh_disable */
512 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
514 struct ct_pcpu *pcpu;
516 /* add this conntrack to the (per cpu) unconfirmed list */
517 ct->cpu = smp_processor_id();
518 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
520 spin_lock(&pcpu->lock);
521 hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
523 spin_unlock(&pcpu->lock);
526 /* must be called with local_bh_disable */
527 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
529 struct ct_pcpu *pcpu;
531 /* We overload first tuple to link into unconfirmed or dying list.*/
532 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
534 spin_lock(&pcpu->lock);
535 BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
536 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
537 spin_unlock(&pcpu->lock);
540 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
542 /* Released via destroy_conntrack() */
543 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
544 const struct nf_conntrack_zone *zone,
547 struct nf_conn *tmpl, *p;
549 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
550 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
555 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
557 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
558 tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
561 tmpl = kzalloc(sizeof(*tmpl), flags);
566 tmpl->status = IPS_TEMPLATE;
567 write_pnet(&tmpl->ct_net, net);
568 nf_ct_zone_add(tmpl, zone);
569 atomic_set(&tmpl->ct_general.use, 0);
573 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
575 void nf_ct_tmpl_free(struct nf_conn *tmpl)
577 nf_ct_ext_destroy(tmpl);
579 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
580 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
584 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
586 static void destroy_gre_conntrack(struct nf_conn *ct)
588 #ifdef CONFIG_NF_CT_PROTO_GRE
589 struct nf_conn *master = ct->master;
592 nf_ct_gre_keymap_destroy(master);
597 destroy_conntrack(struct nf_conntrack *nfct)
599 struct nf_conn *ct = (struct nf_conn *)nfct;
601 pr_debug("destroy_conntrack(%p)\n", ct);
602 WARN_ON(atomic_read(&nfct->use) != 0);
604 if (unlikely(nf_ct_is_template(ct))) {
609 if (unlikely(nf_ct_protonum(ct) == IPPROTO_GRE))
610 destroy_gre_conntrack(ct);
613 /* Expectations will have been removed in clean_from_lists,
614 * except TFTP can create an expectation on the first packet,
615 * before connection is in the list, so we need to clean here,
618 nf_ct_remove_expectations(ct);
620 nf_ct_del_from_dying_or_unconfirmed_list(ct);
625 nf_ct_put(ct->master);
627 pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
628 nf_conntrack_free(ct);
631 static void nf_ct_delete_from_lists(struct nf_conn *ct)
633 struct net *net = nf_ct_net(ct);
634 unsigned int hash, reply_hash;
635 unsigned int sequence;
637 nf_ct_helper_destroy(ct);
641 sequence = read_seqcount_begin(&nf_conntrack_generation);
642 hash = hash_conntrack(net,
643 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
644 reply_hash = hash_conntrack(net,
645 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
646 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
648 clean_from_lists(ct);
649 nf_conntrack_double_unlock(hash, reply_hash);
651 nf_ct_add_to_dying_list(ct);
656 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
658 struct nf_conn_tstamp *tstamp;
661 if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
664 tstamp = nf_conn_tstamp_find(ct);
665 if (tstamp && tstamp->stop == 0)
666 tstamp->stop = ktime_get_real_ns();
668 if (nf_conntrack_event_report(IPCT_DESTROY, ct,
669 portid, report) < 0) {
670 /* destroy event was not delivered. nf_ct_put will
671 * be done by event cache worker on redelivery.
673 nf_ct_delete_from_lists(ct);
674 nf_conntrack_ecache_work(nf_ct_net(ct), NFCT_ECACHE_DESTROY_FAIL);
679 if (nf_conntrack_ecache_dwork_pending(net))
680 nf_conntrack_ecache_work(net, NFCT_ECACHE_DESTROY_SENT);
681 nf_ct_delete_from_lists(ct);
685 EXPORT_SYMBOL_GPL(nf_ct_delete);
688 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
689 const struct nf_conntrack_tuple *tuple,
690 const struct nf_conntrack_zone *zone,
691 const struct net *net)
693 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
695 /* A conntrack can be recreated with the equal tuple,
696 * so we need to check that the conntrack is confirmed
698 return nf_ct_tuple_equal(tuple, &h->tuple) &&
699 nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
700 nf_ct_is_confirmed(ct) &&
701 net_eq(net, nf_ct_net(ct));
705 nf_ct_match(const struct nf_conn *ct1, const struct nf_conn *ct2)
707 return nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
708 &ct2->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
709 nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_REPLY].tuple,
710 &ct2->tuplehash[IP_CT_DIR_REPLY].tuple) &&
711 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_ORIGINAL) &&
712 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_REPLY) &&
713 net_eq(nf_ct_net(ct1), nf_ct_net(ct2));
716 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
717 static void nf_ct_gc_expired(struct nf_conn *ct)
719 if (!atomic_inc_not_zero(&ct->ct_general.use))
722 if (nf_ct_should_gc(ct))
730 * - Caller must take a reference on returned object
731 * and recheck nf_ct_tuple_equal(tuple, &h->tuple)
733 static struct nf_conntrack_tuple_hash *
734 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
735 const struct nf_conntrack_tuple *tuple, u32 hash)
737 struct nf_conntrack_tuple_hash *h;
738 struct hlist_nulls_head *ct_hash;
739 struct hlist_nulls_node *n;
740 unsigned int bucket, hsize;
743 nf_conntrack_get_ht(&ct_hash, &hsize);
744 bucket = reciprocal_scale(hash, hsize);
746 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
749 ct = nf_ct_tuplehash_to_ctrack(h);
750 if (nf_ct_is_expired(ct)) {
751 nf_ct_gc_expired(ct);
755 if (nf_ct_key_equal(h, tuple, zone, net))
759 * if the nulls value we got at the end of this lookup is
760 * not the expected one, we must restart lookup.
761 * We probably met an item that was moved to another chain.
763 if (get_nulls_value(n) != bucket) {
764 NF_CT_STAT_INC_ATOMIC(net, search_restart);
771 /* Find a connection corresponding to a tuple. */
772 static struct nf_conntrack_tuple_hash *
773 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
774 const struct nf_conntrack_tuple *tuple, u32 hash)
776 struct nf_conntrack_tuple_hash *h;
781 h = ____nf_conntrack_find(net, zone, tuple, hash);
783 /* We have a candidate that matches the tuple we're interested
784 * in, try to obtain a reference and re-check tuple
786 ct = nf_ct_tuplehash_to_ctrack(h);
787 if (likely(atomic_inc_not_zero(&ct->ct_general.use))) {
788 if (likely(nf_ct_key_equal(h, tuple, zone, net)))
791 /* TYPESAFE_BY_RCU recycled the candidate */
803 struct nf_conntrack_tuple_hash *
804 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
805 const struct nf_conntrack_tuple *tuple)
807 return __nf_conntrack_find_get(net, zone, tuple,
808 hash_conntrack_raw(tuple, net));
810 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
812 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
814 unsigned int reply_hash)
816 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
817 &nf_conntrack_hash[hash]);
818 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
819 &nf_conntrack_hash[reply_hash]);
823 nf_conntrack_hash_check_insert(struct nf_conn *ct)
825 const struct nf_conntrack_zone *zone;
826 struct net *net = nf_ct_net(ct);
827 unsigned int hash, reply_hash;
828 struct nf_conntrack_tuple_hash *h;
829 struct hlist_nulls_node *n;
830 unsigned int sequence;
832 zone = nf_ct_zone(ct);
836 sequence = read_seqcount_begin(&nf_conntrack_generation);
837 hash = hash_conntrack(net,
838 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
839 reply_hash = hash_conntrack(net,
840 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
841 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
843 /* See if there's one in the list already, including reverse */
844 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
845 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
849 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
850 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
855 /* The caller holds a reference to this object */
856 atomic_set(&ct->ct_general.use, 2);
857 __nf_conntrack_hash_insert(ct, hash, reply_hash);
858 nf_conntrack_double_unlock(hash, reply_hash);
859 NF_CT_STAT_INC(net, insert);
864 nf_conntrack_double_unlock(hash, reply_hash);
868 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
870 void nf_ct_acct_add(struct nf_conn *ct, u32 dir, unsigned int packets,
873 struct nf_conn_acct *acct;
875 acct = nf_conn_acct_find(ct);
877 struct nf_conn_counter *counter = acct->counter;
879 atomic64_add(packets, &counter[dir].packets);
880 atomic64_add(bytes, &counter[dir].bytes);
883 EXPORT_SYMBOL_GPL(nf_ct_acct_add);
885 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
886 const struct nf_conn *loser_ct)
888 struct nf_conn_acct *acct;
890 acct = nf_conn_acct_find(loser_ct);
892 struct nf_conn_counter *counter = acct->counter;
895 /* u32 should be fine since we must have seen one packet. */
896 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
897 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), bytes);
901 static void __nf_conntrack_insert_prepare(struct nf_conn *ct)
903 struct nf_conn_tstamp *tstamp;
905 atomic_inc(&ct->ct_general.use);
906 ct->status |= IPS_CONFIRMED;
908 /* set conntrack timestamp, if enabled. */
909 tstamp = nf_conn_tstamp_find(ct);
911 tstamp->start = ktime_get_real_ns();
914 /* caller must hold locks to prevent concurrent changes */
915 static int __nf_ct_resolve_clash(struct sk_buff *skb,
916 struct nf_conntrack_tuple_hash *h)
918 /* This is the conntrack entry already in hashes that won race. */
919 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
920 enum ip_conntrack_info ctinfo;
921 struct nf_conn *loser_ct;
923 loser_ct = nf_ct_get(skb, &ctinfo);
925 if (nf_ct_is_dying(ct))
928 if (((ct->status & IPS_NAT_DONE_MASK) == 0) ||
929 nf_ct_match(ct, loser_ct)) {
930 struct net *net = nf_ct_net(ct);
932 nf_conntrack_get(&ct->ct_general);
934 nf_ct_acct_merge(ct, ctinfo, loser_ct);
935 nf_ct_add_to_dying_list(loser_ct);
936 nf_conntrack_put(&loser_ct->ct_general);
937 nf_ct_set(skb, ct, ctinfo);
939 NF_CT_STAT_INC(net, clash_resolve);
947 * nf_ct_resolve_clash_harder - attempt to insert clashing conntrack entry
949 * @skb: skb that causes the collision
950 * @repl_idx: hash slot for reply direction
952 * Called when origin or reply direction had a clash.
953 * The skb can be handled without packet drop provided the reply direction
954 * is unique or there the existing entry has the identical tuple in both
957 * Caller must hold conntrack table locks to prevent concurrent updates.
959 * Returns NF_DROP if the clash could not be handled.
961 static int nf_ct_resolve_clash_harder(struct sk_buff *skb, u32 repl_idx)
963 struct nf_conn *loser_ct = (struct nf_conn *)skb_nfct(skb);
964 const struct nf_conntrack_zone *zone;
965 struct nf_conntrack_tuple_hash *h;
966 struct hlist_nulls_node *n;
969 zone = nf_ct_zone(loser_ct);
970 net = nf_ct_net(loser_ct);
972 /* Reply direction must never result in a clash, unless both origin
973 * and reply tuples are identical.
975 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[repl_idx], hnnode) {
976 if (nf_ct_key_equal(h,
977 &loser_ct->tuplehash[IP_CT_DIR_REPLY].tuple,
979 return __nf_ct_resolve_clash(skb, h);
982 /* We want the clashing entry to go away real soon: 1 second timeout. */
983 loser_ct->timeout = nfct_time_stamp + HZ;
985 /* IPS_NAT_CLASH removes the entry automatically on the first
986 * reply. Also prevents UDP tracker from moving the entry to
987 * ASSURED state, i.e. the entry can always be evicted under
990 loser_ct->status |= IPS_FIXED_TIMEOUT | IPS_NAT_CLASH;
992 __nf_conntrack_insert_prepare(loser_ct);
994 /* fake add for ORIGINAL dir: we want lookups to only find the entry
995 * already in the table. This also hides the clashing entry from
996 * ctnetlink iteration, i.e. conntrack -L won't show them.
998 hlist_nulls_add_fake(&loser_ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
1000 hlist_nulls_add_head_rcu(&loser_ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
1001 &nf_conntrack_hash[repl_idx]);
1003 NF_CT_STAT_INC(net, clash_resolve);
1008 * nf_ct_resolve_clash - attempt to handle clash without packet drop
1010 * @skb: skb that causes the clash
1011 * @h: tuplehash of the clashing entry already in table
1012 * @reply_hash: hash slot for reply direction
1014 * A conntrack entry can be inserted to the connection tracking table
1015 * if there is no existing entry with an identical tuple.
1017 * If there is one, @skb (and the assocated, unconfirmed conntrack) has
1018 * to be dropped. In case @skb is retransmitted, next conntrack lookup
1019 * will find the already-existing entry.
1021 * The major problem with such packet drop is the extra delay added by
1022 * the packet loss -- it will take some time for a retransmit to occur
1023 * (or the sender to time out when waiting for a reply).
1025 * This function attempts to handle the situation without packet drop.
1027 * If @skb has no NAT transformation or if the colliding entries are
1028 * exactly the same, only the to-be-confirmed conntrack entry is discarded
1029 * and @skb is associated with the conntrack entry already in the table.
1031 * Failing that, the new, unconfirmed conntrack is still added to the table
1032 * provided that the collision only occurs in the ORIGINAL direction.
1033 * The new entry will be added only in the non-clashing REPLY direction,
1034 * so packets in the ORIGINAL direction will continue to match the existing
1035 * entry. The new entry will also have a fixed timeout so it expires --
1036 * due to the collision, it will only see reply traffic.
1038 * Returns NF_DROP if the clash could not be resolved.
1040 static __cold noinline int
1041 nf_ct_resolve_clash(struct sk_buff *skb, struct nf_conntrack_tuple_hash *h,
1044 /* This is the conntrack entry already in hashes that won race. */
1045 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
1046 const struct nf_conntrack_l4proto *l4proto;
1047 enum ip_conntrack_info ctinfo;
1048 struct nf_conn *loser_ct;
1052 loser_ct = nf_ct_get(skb, &ctinfo);
1053 net = nf_ct_net(loser_ct);
1055 l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1056 if (!l4proto->allow_clash)
1059 ret = __nf_ct_resolve_clash(skb, h);
1060 if (ret == NF_ACCEPT)
1063 ret = nf_ct_resolve_clash_harder(skb, reply_hash);
1064 if (ret == NF_ACCEPT)
1068 nf_ct_add_to_dying_list(loser_ct);
1069 NF_CT_STAT_INC(net, drop);
1070 NF_CT_STAT_INC(net, insert_failed);
1074 /* Confirm a connection given skb; places it in hash table */
1076 __nf_conntrack_confirm(struct sk_buff *skb)
1078 const struct nf_conntrack_zone *zone;
1079 unsigned int hash, reply_hash;
1080 struct nf_conntrack_tuple_hash *h;
1082 struct nf_conn_help *help;
1083 struct hlist_nulls_node *n;
1084 enum ip_conntrack_info ctinfo;
1086 unsigned int sequence;
1089 ct = nf_ct_get(skb, &ctinfo);
1090 net = nf_ct_net(ct);
1092 /* ipt_REJECT uses nf_conntrack_attach to attach related
1093 ICMP/TCP RST packets in other direction. Actual packet
1094 which created connection will be IP_CT_NEW or for an
1095 expected connection, IP_CT_RELATED. */
1096 if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
1099 zone = nf_ct_zone(ct);
1103 sequence = read_seqcount_begin(&nf_conntrack_generation);
1104 /* reuse the hash saved before */
1105 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
1106 hash = scale_hash(hash);
1107 reply_hash = hash_conntrack(net,
1108 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
1110 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
1112 /* We're not in hash table, and we refuse to set up related
1113 * connections for unconfirmed conns. But packet copies and
1114 * REJECT will give spurious warnings here.
1117 /* Another skb with the same unconfirmed conntrack may
1118 * win the race. This may happen for bridge(br_flood)
1119 * or broadcast/multicast packets do skb_clone with
1120 * unconfirmed conntrack.
1122 if (unlikely(nf_ct_is_confirmed(ct))) {
1124 nf_conntrack_double_unlock(hash, reply_hash);
1129 pr_debug("Confirming conntrack %p\n", ct);
1130 /* We have to check the DYING flag after unlink to prevent
1131 * a race against nf_ct_get_next_corpse() possibly called from
1132 * user context, else we insert an already 'dead' hash, blocking
1133 * further use of that particular connection -JM.
1135 nf_ct_del_from_dying_or_unconfirmed_list(ct);
1137 if (unlikely(nf_ct_is_dying(ct))) {
1138 nf_ct_add_to_dying_list(ct);
1139 NF_CT_STAT_INC(net, insert_failed);
1143 /* See if there's one in the list already, including reverse:
1144 NAT could have grabbed it without realizing, since we're
1145 not in the hash. If there is, we lost race. */
1146 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
1147 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1151 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
1152 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
1156 /* Timer relative to confirmation time, not original
1157 setting time, otherwise we'd get timer wrap in
1158 weird delay cases. */
1159 ct->timeout += nfct_time_stamp;
1161 __nf_conntrack_insert_prepare(ct);
1163 /* Since the lookup is lockless, hash insertion must be done after
1164 * starting the timer and setting the CONFIRMED bit. The RCU barriers
1165 * guarantee that no other CPU can find the conntrack before the above
1166 * stores are visible.
1168 __nf_conntrack_hash_insert(ct, hash, reply_hash);
1169 nf_conntrack_double_unlock(hash, reply_hash);
1172 help = nfct_help(ct);
1173 if (help && help->helper)
1174 nf_conntrack_event_cache(IPCT_HELPER, ct);
1176 nf_conntrack_event_cache(master_ct(ct) ?
1177 IPCT_RELATED : IPCT_NEW, ct);
1181 ret = nf_ct_resolve_clash(skb, h, reply_hash);
1183 nf_conntrack_double_unlock(hash, reply_hash);
1187 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
1189 /* Returns true if a connection correspondings to the tuple (required
1192 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
1193 const struct nf_conn *ignored_conntrack)
1195 struct net *net = nf_ct_net(ignored_conntrack);
1196 const struct nf_conntrack_zone *zone;
1197 struct nf_conntrack_tuple_hash *h;
1198 struct hlist_nulls_head *ct_hash;
1199 unsigned int hash, hsize;
1200 struct hlist_nulls_node *n;
1203 zone = nf_ct_zone(ignored_conntrack);
1207 nf_conntrack_get_ht(&ct_hash, &hsize);
1208 hash = __hash_conntrack(net, tuple, hsize);
1210 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
1211 ct = nf_ct_tuplehash_to_ctrack(h);
1213 if (ct == ignored_conntrack)
1216 if (nf_ct_is_expired(ct)) {
1217 nf_ct_gc_expired(ct);
1221 if (nf_ct_key_equal(h, tuple, zone, net)) {
1222 /* Tuple is taken already, so caller will need to find
1223 * a new source port to use.
1226 * If the *original tuples* are identical, then both
1227 * conntracks refer to the same flow.
1228 * This is a rare situation, it can occur e.g. when
1229 * more than one UDP packet is sent from same socket
1230 * in different threads.
1232 * Let nf_ct_resolve_clash() deal with this later.
1234 if (nf_ct_tuple_equal(&ignored_conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
1235 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
1236 nf_ct_zone_equal(ct, zone, IP_CT_DIR_ORIGINAL))
1239 NF_CT_STAT_INC_ATOMIC(net, found);
1245 if (get_nulls_value(n) != hash) {
1246 NF_CT_STAT_INC_ATOMIC(net, search_restart);
1254 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
1256 #define NF_CT_EVICTION_RANGE 8
1258 /* There's a small race here where we may free a just-assured
1259 connection. Too bad: we're in trouble anyway. */
1260 static unsigned int early_drop_list(struct net *net,
1261 struct hlist_nulls_head *head)
1263 struct nf_conntrack_tuple_hash *h;
1264 struct hlist_nulls_node *n;
1265 unsigned int drops = 0;
1266 struct nf_conn *tmp;
1268 hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
1269 tmp = nf_ct_tuplehash_to_ctrack(h);
1271 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status))
1274 if (nf_ct_is_expired(tmp)) {
1275 nf_ct_gc_expired(tmp);
1279 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
1280 !net_eq(nf_ct_net(tmp), net) ||
1281 nf_ct_is_dying(tmp))
1284 if (!atomic_inc_not_zero(&tmp->ct_general.use))
1287 /* kill only if still in same netns -- might have moved due to
1288 * SLAB_TYPESAFE_BY_RCU rules.
1290 * We steal the timer reference. If that fails timer has
1291 * already fired or someone else deleted it. Just drop ref
1292 * and move to next entry.
1294 if (net_eq(nf_ct_net(tmp), net) &&
1295 nf_ct_is_confirmed(tmp) &&
1296 nf_ct_delete(tmp, 0, 0))
1305 static noinline int early_drop(struct net *net, unsigned int hash)
1307 unsigned int i, bucket;
1309 for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
1310 struct hlist_nulls_head *ct_hash;
1311 unsigned int hsize, drops;
1314 nf_conntrack_get_ht(&ct_hash, &hsize);
1316 bucket = reciprocal_scale(hash, hsize);
1318 bucket = (bucket + 1) % hsize;
1320 drops = early_drop_list(net, &ct_hash[bucket]);
1324 NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
1332 static bool gc_worker_skip_ct(const struct nf_conn *ct)
1334 return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
1337 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
1339 const struct nf_conntrack_l4proto *l4proto;
1341 if (!test_bit(IPS_ASSURED_BIT, &ct->status))
1344 l4proto = nf_ct_l4proto_find(nf_ct_protonum(ct));
1345 if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
1351 static void gc_worker(struct work_struct *work)
1353 unsigned int min_interval = max(HZ / GC_MAX_BUCKETS_DIV, 1u);
1354 unsigned int i, goal, buckets = 0, expired_count = 0;
1355 unsigned int nf_conntrack_max95 = 0;
1356 struct conntrack_gc_work *gc_work;
1357 unsigned int ratio, scanned = 0;
1358 unsigned long next_run;
1360 gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
1362 goal = nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV;
1363 i = gc_work->last_bucket;
1364 if (gc_work->early_drop)
1365 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
1368 struct nf_conntrack_tuple_hash *h;
1369 struct hlist_nulls_head *ct_hash;
1370 struct hlist_nulls_node *n;
1371 unsigned int hashsz;
1372 struct nf_conn *tmp;
1377 nf_conntrack_get_ht(&ct_hash, &hashsz);
1381 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1382 struct nf_conntrack_net *cnet;
1385 tmp = nf_ct_tuplehash_to_ctrack(h);
1388 if (test_bit(IPS_OFFLOAD_BIT, &tmp->status)) {
1389 nf_ct_offload_timeout(tmp);
1393 if (nf_ct_is_expired(tmp)) {
1394 nf_ct_gc_expired(tmp);
1399 if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1402 net = nf_ct_net(tmp);
1403 cnet = nf_ct_pernet(net);
1404 if (atomic_read(&cnet->count) < nf_conntrack_max95)
1407 /* need to take reference to avoid possible races */
1408 if (!atomic_inc_not_zero(&tmp->ct_general.use))
1411 if (gc_worker_skip_ct(tmp)) {
1416 if (gc_worker_can_early_drop(tmp))
1422 /* could check get_nulls_value() here and restart if ct
1423 * was moved to another chain. But given gc is best-effort
1424 * we will just continue with next hash slot.
1428 } while (++buckets < goal);
1430 if (gc_work->exiting)
1434 * Eviction will normally happen from the packet path, and not
1435 * from this gc worker.
1437 * This worker is only here to reap expired entries when system went
1438 * idle after a busy period.
1440 * The heuristics below are supposed to balance conflicting goals:
1442 * 1. Minimize time until we notice a stale entry
1443 * 2. Maximize scan intervals to not waste cycles
1445 * Normally, expire ratio will be close to 0.
1447 * As soon as a sizeable fraction of the entries have expired
1448 * increase scan frequency.
1450 ratio = scanned ? expired_count * 100 / scanned : 0;
1451 if (ratio > GC_EVICT_RATIO) {
1452 gc_work->next_gc_run = min_interval;
1454 unsigned int max = GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV;
1456 BUILD_BUG_ON((GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV) == 0);
1458 gc_work->next_gc_run += min_interval;
1459 if (gc_work->next_gc_run > max)
1460 gc_work->next_gc_run = max;
1463 next_run = gc_work->next_gc_run;
1464 gc_work->last_bucket = i;
1465 gc_work->early_drop = false;
1466 queue_delayed_work(system_power_efficient_wq, &gc_work->dwork, next_run);
1469 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1471 INIT_DEFERRABLE_WORK(&gc_work->dwork, gc_worker);
1472 gc_work->next_gc_run = HZ;
1473 gc_work->exiting = false;
1476 static struct nf_conn *
1477 __nf_conntrack_alloc(struct net *net,
1478 const struct nf_conntrack_zone *zone,
1479 const struct nf_conntrack_tuple *orig,
1480 const struct nf_conntrack_tuple *repl,
1481 gfp_t gfp, u32 hash)
1483 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
1484 unsigned int ct_count;
1487 /* We don't want any race condition at early drop stage */
1488 ct_count = atomic_inc_return(&cnet->count);
1490 if (nf_conntrack_max && unlikely(ct_count > nf_conntrack_max)) {
1491 if (!early_drop(net, hash)) {
1492 if (!conntrack_gc_work.early_drop)
1493 conntrack_gc_work.early_drop = true;
1494 atomic_dec(&cnet->count);
1495 net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1496 return ERR_PTR(-ENOMEM);
1501 * Do not use kmem_cache_zalloc(), as this cache uses
1502 * SLAB_TYPESAFE_BY_RCU.
1504 ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1508 spin_lock_init(&ct->lock);
1509 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1510 ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1511 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1512 /* save hash for reusing when confirming */
1513 *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1516 write_pnet(&ct->ct_net, net);
1517 memset(&ct->__nfct_init_offset, 0,
1518 offsetof(struct nf_conn, proto) -
1519 offsetof(struct nf_conn, __nfct_init_offset));
1521 nf_ct_zone_add(ct, zone);
1523 /* Because we use RCU lookups, we set ct_general.use to zero before
1524 * this is inserted in any list.
1526 atomic_set(&ct->ct_general.use, 0);
1529 atomic_dec(&cnet->count);
1530 return ERR_PTR(-ENOMEM);
1533 struct nf_conn *nf_conntrack_alloc(struct net *net,
1534 const struct nf_conntrack_zone *zone,
1535 const struct nf_conntrack_tuple *orig,
1536 const struct nf_conntrack_tuple *repl,
1539 return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1541 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1543 void nf_conntrack_free(struct nf_conn *ct)
1545 struct net *net = nf_ct_net(ct);
1546 struct nf_conntrack_net *cnet;
1548 /* A freed object has refcnt == 0, that's
1549 * the golden rule for SLAB_TYPESAFE_BY_RCU
1551 WARN_ON(atomic_read(&ct->ct_general.use) != 0);
1553 nf_ct_ext_destroy(ct);
1554 kmem_cache_free(nf_conntrack_cachep, ct);
1555 cnet = nf_ct_pernet(net);
1557 smp_mb__before_atomic();
1558 atomic_dec(&cnet->count);
1560 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1563 /* Allocate a new conntrack: we return -ENOMEM if classification
1564 failed due to stress. Otherwise it really is unclassifiable. */
1565 static noinline struct nf_conntrack_tuple_hash *
1566 init_conntrack(struct net *net, struct nf_conn *tmpl,
1567 const struct nf_conntrack_tuple *tuple,
1568 struct sk_buff *skb,
1569 unsigned int dataoff, u32 hash)
1572 struct nf_conn_help *help;
1573 struct nf_conntrack_tuple repl_tuple;
1574 struct nf_conntrack_ecache *ecache;
1575 struct nf_conntrack_expect *exp = NULL;
1576 const struct nf_conntrack_zone *zone;
1577 struct nf_conn_timeout *timeout_ext;
1578 struct nf_conntrack_zone tmp;
1579 struct nf_conntrack_net *cnet;
1581 if (!nf_ct_invert_tuple(&repl_tuple, tuple)) {
1582 pr_debug("Can't invert tuple.\n");
1586 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1587 ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1590 return (struct nf_conntrack_tuple_hash *)ct;
1592 if (!nf_ct_add_synproxy(ct, tmpl)) {
1593 nf_conntrack_free(ct);
1594 return ERR_PTR(-ENOMEM);
1597 timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1600 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1603 nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1604 nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1605 nf_ct_labels_ext_add(ct);
1607 ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1608 nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1609 ecache ? ecache->expmask : 0,
1613 cnet = nf_ct_pernet(net);
1614 if (cnet->expect_count) {
1615 spin_lock(&nf_conntrack_expect_lock);
1616 exp = nf_ct_find_expectation(net, zone, tuple);
1618 pr_debug("expectation arrives ct=%p exp=%p\n",
1620 /* Welcome, Mr. Bond. We've been expecting you... */
1621 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1622 /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1623 ct->master = exp->master;
1625 help = nf_ct_helper_ext_add(ct, GFP_ATOMIC);
1627 rcu_assign_pointer(help->helper, exp->helper);
1630 #ifdef CONFIG_NF_CONNTRACK_MARK
1631 ct->mark = exp->master->mark;
1633 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1634 ct->secmark = exp->master->secmark;
1636 NF_CT_STAT_INC(net, expect_new);
1638 spin_unlock(&nf_conntrack_expect_lock);
1641 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1643 /* Now it is inserted into the unconfirmed list, bump refcount */
1644 nf_conntrack_get(&ct->ct_general);
1645 nf_ct_add_to_unconfirmed_list(ct);
1651 exp->expectfn(ct, exp);
1652 nf_ct_expect_put(exp);
1655 return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1658 /* On success, returns 0, sets skb->_nfct | ctinfo */
1660 resolve_normal_ct(struct nf_conn *tmpl,
1661 struct sk_buff *skb,
1662 unsigned int dataoff,
1664 const struct nf_hook_state *state)
1666 const struct nf_conntrack_zone *zone;
1667 struct nf_conntrack_tuple tuple;
1668 struct nf_conntrack_tuple_hash *h;
1669 enum ip_conntrack_info ctinfo;
1670 struct nf_conntrack_zone tmp;
1674 if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1675 dataoff, state->pf, protonum, state->net,
1677 pr_debug("Can't get tuple\n");
1681 /* look for tuple match */
1682 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1683 hash = hash_conntrack_raw(&tuple, state->net);
1684 h = __nf_conntrack_find_get(state->net, zone, &tuple, hash);
1686 h = init_conntrack(state->net, tmpl, &tuple,
1687 skb, dataoff, hash);
1693 ct = nf_ct_tuplehash_to_ctrack(h);
1695 /* It exists; we have (non-exclusive) reference. */
1696 if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1697 ctinfo = IP_CT_ESTABLISHED_REPLY;
1699 /* Once we've had two way comms, always ESTABLISHED. */
1700 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1701 pr_debug("normal packet for %p\n", ct);
1702 ctinfo = IP_CT_ESTABLISHED;
1703 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1704 pr_debug("related packet for %p\n", ct);
1705 ctinfo = IP_CT_RELATED;
1707 pr_debug("new packet for %p\n", ct);
1711 nf_ct_set(skb, ct, ctinfo);
1716 * icmp packets need special treatment to handle error messages that are
1717 * related to a connection.
1719 * Callers need to check if skb has a conntrack assigned when this
1720 * helper returns; in such case skb belongs to an already known connection.
1722 static unsigned int __cold
1723 nf_conntrack_handle_icmp(struct nf_conn *tmpl,
1724 struct sk_buff *skb,
1725 unsigned int dataoff,
1727 const struct nf_hook_state *state)
1731 if (state->pf == NFPROTO_IPV4 && protonum == IPPROTO_ICMP)
1732 ret = nf_conntrack_icmpv4_error(tmpl, skb, dataoff, state);
1733 #if IS_ENABLED(CONFIG_IPV6)
1734 else if (state->pf == NFPROTO_IPV6 && protonum == IPPROTO_ICMPV6)
1735 ret = nf_conntrack_icmpv6_error(tmpl, skb, dataoff, state);
1741 NF_CT_STAT_INC_ATOMIC(state->net, error);
1746 static int generic_packet(struct nf_conn *ct, struct sk_buff *skb,
1747 enum ip_conntrack_info ctinfo)
1749 const unsigned int *timeout = nf_ct_timeout_lookup(ct);
1752 timeout = &nf_generic_pernet(nf_ct_net(ct))->timeout;
1754 nf_ct_refresh_acct(ct, ctinfo, skb, *timeout);
1758 /* Returns verdict for packet, or -1 for invalid. */
1759 static int nf_conntrack_handle_packet(struct nf_conn *ct,
1760 struct sk_buff *skb,
1761 unsigned int dataoff,
1762 enum ip_conntrack_info ctinfo,
1763 const struct nf_hook_state *state)
1765 switch (nf_ct_protonum(ct)) {
1767 return nf_conntrack_tcp_packet(ct, skb, dataoff,
1770 return nf_conntrack_udp_packet(ct, skb, dataoff,
1773 return nf_conntrack_icmp_packet(ct, skb, ctinfo, state);
1774 #if IS_ENABLED(CONFIG_IPV6)
1775 case IPPROTO_ICMPV6:
1776 return nf_conntrack_icmpv6_packet(ct, skb, ctinfo, state);
1778 #ifdef CONFIG_NF_CT_PROTO_UDPLITE
1779 case IPPROTO_UDPLITE:
1780 return nf_conntrack_udplite_packet(ct, skb, dataoff,
1783 #ifdef CONFIG_NF_CT_PROTO_SCTP
1785 return nf_conntrack_sctp_packet(ct, skb, dataoff,
1788 #ifdef CONFIG_NF_CT_PROTO_DCCP
1790 return nf_conntrack_dccp_packet(ct, skb, dataoff,
1793 #ifdef CONFIG_NF_CT_PROTO_GRE
1795 return nf_conntrack_gre_packet(ct, skb, dataoff,
1800 return generic_packet(ct, skb, ctinfo);
1804 nf_conntrack_in(struct sk_buff *skb, const struct nf_hook_state *state)
1806 enum ip_conntrack_info ctinfo;
1807 struct nf_conn *ct, *tmpl;
1811 tmpl = nf_ct_get(skb, &ctinfo);
1812 if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1813 /* Previously seen (loopback or untracked)? Ignore. */
1814 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1815 ctinfo == IP_CT_UNTRACKED)
1820 /* rcu_read_lock()ed by nf_hook_thresh */
1821 dataoff = get_l4proto(skb, skb_network_offset(skb), state->pf, &protonum);
1823 pr_debug("not prepared to track yet or error occurred\n");
1824 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1829 if (protonum == IPPROTO_ICMP || protonum == IPPROTO_ICMPV6) {
1830 ret = nf_conntrack_handle_icmp(tmpl, skb, dataoff,
1836 /* ICMP[v6] protocol trackers may assign one conntrack. */
1841 ret = resolve_normal_ct(tmpl, skb, dataoff,
1844 /* Too stressed to deal. */
1845 NF_CT_STAT_INC_ATOMIC(state->net, drop);
1850 ct = nf_ct_get(skb, &ctinfo);
1852 /* Not valid part of a connection */
1853 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1858 ret = nf_conntrack_handle_packet(ct, skb, dataoff, ctinfo, state);
1860 /* Invalid: inverse of the return code tells
1861 * the netfilter core what to do */
1862 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1863 nf_conntrack_put(&ct->ct_general);
1865 NF_CT_STAT_INC_ATOMIC(state->net, invalid);
1866 if (ret == -NF_DROP)
1867 NF_CT_STAT_INC_ATOMIC(state->net, drop);
1868 /* Special case: TCP tracker reports an attempt to reopen a
1869 * closed/aborted connection. We have to go back and create a
1872 if (ret == -NF_REPEAT)
1878 if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
1879 !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1880 nf_conntrack_event_cache(IPCT_REPLY, ct);
1887 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1889 /* Alter reply tuple (maybe alter helper). This is for NAT, and is
1890 implicitly racy: see __nf_conntrack_confirm */
1891 void nf_conntrack_alter_reply(struct nf_conn *ct,
1892 const struct nf_conntrack_tuple *newreply)
1894 struct nf_conn_help *help = nfct_help(ct);
1896 /* Should be unconfirmed, so not in hash table yet */
1897 WARN_ON(nf_ct_is_confirmed(ct));
1899 pr_debug("Altering reply tuple of %p to ", ct);
1900 nf_ct_dump_tuple(newreply);
1902 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1903 if (ct->master || (help && !hlist_empty(&help->expectations)))
1907 __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1910 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1912 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1913 void __nf_ct_refresh_acct(struct nf_conn *ct,
1914 enum ip_conntrack_info ctinfo,
1915 const struct sk_buff *skb,
1919 /* Only update if this is not a fixed timeout */
1920 if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1923 /* If not in hash table, timer will not be active yet */
1924 if (nf_ct_is_confirmed(ct))
1925 extra_jiffies += nfct_time_stamp;
1927 if (READ_ONCE(ct->timeout) != extra_jiffies)
1928 WRITE_ONCE(ct->timeout, extra_jiffies);
1931 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
1933 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1935 bool nf_ct_kill_acct(struct nf_conn *ct,
1936 enum ip_conntrack_info ctinfo,
1937 const struct sk_buff *skb)
1939 nf_ct_acct_update(ct, CTINFO2DIR(ctinfo), skb->len);
1941 return nf_ct_delete(ct, 0, 0);
1943 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
1945 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1947 #include <linux/netfilter/nfnetlink.h>
1948 #include <linux/netfilter/nfnetlink_conntrack.h>
1949 #include <linux/mutex.h>
1951 /* Generic function for tcp/udp/sctp/dccp and alike. */
1952 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1953 const struct nf_conntrack_tuple *tuple)
1955 if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1956 nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1957 goto nla_put_failure;
1963 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1965 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1966 [CTA_PROTO_SRC_PORT] = { .type = NLA_U16 },
1967 [CTA_PROTO_DST_PORT] = { .type = NLA_U16 },
1969 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1971 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1972 struct nf_conntrack_tuple *t,
1975 if (flags & CTA_FILTER_FLAG(CTA_PROTO_SRC_PORT)) {
1976 if (!tb[CTA_PROTO_SRC_PORT])
1979 t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1982 if (flags & CTA_FILTER_FLAG(CTA_PROTO_DST_PORT)) {
1983 if (!tb[CTA_PROTO_DST_PORT])
1986 t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1991 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1993 unsigned int nf_ct_port_nlattr_tuple_size(void)
1995 static unsigned int size __read_mostly;
1998 size = nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
2002 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
2005 /* Used by ipt_REJECT and ip6t_REJECT. */
2006 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
2009 enum ip_conntrack_info ctinfo;
2011 /* This ICMP is in reverse direction to the packet which caused it */
2012 ct = nf_ct_get(skb, &ctinfo);
2013 if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
2014 ctinfo = IP_CT_RELATED_REPLY;
2016 ctinfo = IP_CT_RELATED;
2018 /* Attach to new skbuff, and increment count */
2019 nf_ct_set(nskb, ct, ctinfo);
2020 nf_conntrack_get(skb_nfct(nskb));
2023 static int __nf_conntrack_update(struct net *net, struct sk_buff *skb,
2025 enum ip_conntrack_info ctinfo)
2027 struct nf_conntrack_tuple_hash *h;
2028 struct nf_conntrack_tuple tuple;
2029 struct nf_nat_hook *nat_hook;
2030 unsigned int status;
2035 l3num = nf_ct_l3num(ct);
2037 dataoff = get_l4proto(skb, skb_network_offset(skb), l3num, &l4num);
2041 if (!nf_ct_get_tuple(skb, skb_network_offset(skb), dataoff, l3num,
2042 l4num, net, &tuple))
2045 if (ct->status & IPS_SRC_NAT) {
2046 memcpy(tuple.src.u3.all,
2047 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u3.all,
2048 sizeof(tuple.src.u3.all));
2050 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.src.u.all;
2053 if (ct->status & IPS_DST_NAT) {
2054 memcpy(tuple.dst.u3.all,
2055 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u3.all,
2056 sizeof(tuple.dst.u3.all));
2058 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple.dst.u.all;
2061 h = nf_conntrack_find_get(net, nf_ct_zone(ct), &tuple);
2065 /* Store status bits of the conntrack that is clashing to re-do NAT
2066 * mangling according to what it has been done already to this packet.
2068 status = ct->status;
2071 ct = nf_ct_tuplehash_to_ctrack(h);
2072 nf_ct_set(skb, ct, ctinfo);
2074 nat_hook = rcu_dereference(nf_nat_hook);
2078 if (status & IPS_SRC_NAT &&
2079 nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_SRC,
2080 IP_CT_DIR_ORIGINAL) == NF_DROP)
2083 if (status & IPS_DST_NAT &&
2084 nat_hook->manip_pkt(skb, ct, NF_NAT_MANIP_DST,
2085 IP_CT_DIR_ORIGINAL) == NF_DROP)
2091 /* This packet is coming from userspace via nf_queue, complete the packet
2092 * processing after the helper invocation in nf_confirm().
2094 static int nf_confirm_cthelper(struct sk_buff *skb, struct nf_conn *ct,
2095 enum ip_conntrack_info ctinfo)
2097 const struct nf_conntrack_helper *helper;
2098 const struct nf_conn_help *help;
2101 help = nfct_help(ct);
2105 helper = rcu_dereference(help->helper);
2106 if (!(helper->flags & NF_CT_HELPER_F_USERSPACE))
2109 switch (nf_ct_l3num(ct)) {
2111 protoff = skb_network_offset(skb) + ip_hdrlen(skb);
2113 #if IS_ENABLED(CONFIG_IPV6)
2114 case NFPROTO_IPV6: {
2118 pnum = ipv6_hdr(skb)->nexthdr;
2119 protoff = ipv6_skip_exthdr(skb, sizeof(struct ipv6hdr), &pnum,
2121 if (protoff < 0 || (frag_off & htons(~0x7)) != 0)
2130 if (test_bit(IPS_SEQ_ADJUST_BIT, &ct->status) &&
2131 !nf_is_loopback_packet(skb)) {
2132 if (!nf_ct_seq_adjust(skb, ct, ctinfo, protoff)) {
2133 NF_CT_STAT_INC_ATOMIC(nf_ct_net(ct), drop);
2138 /* We've seen it coming out the other side: confirm it */
2139 return nf_conntrack_confirm(skb) == NF_DROP ? - 1 : 0;
2142 static int nf_conntrack_update(struct net *net, struct sk_buff *skb)
2144 enum ip_conntrack_info ctinfo;
2148 ct = nf_ct_get(skb, &ctinfo);
2152 if (!nf_ct_is_confirmed(ct)) {
2153 err = __nf_conntrack_update(net, skb, ct, ctinfo);
2157 ct = nf_ct_get(skb, &ctinfo);
2160 return nf_confirm_cthelper(skb, ct, ctinfo);
2163 static bool nf_conntrack_get_tuple_skb(struct nf_conntrack_tuple *dst_tuple,
2164 const struct sk_buff *skb)
2166 const struct nf_conntrack_tuple *src_tuple;
2167 const struct nf_conntrack_tuple_hash *hash;
2168 struct nf_conntrack_tuple srctuple;
2169 enum ip_conntrack_info ctinfo;
2172 ct = nf_ct_get(skb, &ctinfo);
2174 src_tuple = nf_ct_tuple(ct, CTINFO2DIR(ctinfo));
2175 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2179 if (!nf_ct_get_tuplepr(skb, skb_network_offset(skb),
2180 NFPROTO_IPV4, dev_net(skb->dev),
2184 hash = nf_conntrack_find_get(dev_net(skb->dev),
2190 ct = nf_ct_tuplehash_to_ctrack(hash);
2191 src_tuple = nf_ct_tuple(ct, !hash->tuple.dst.dir);
2192 memcpy(dst_tuple, src_tuple, sizeof(*dst_tuple));
2198 /* Bring out ya dead! */
2199 static struct nf_conn *
2200 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
2201 void *data, unsigned int *bucket)
2203 struct nf_conntrack_tuple_hash *h;
2205 struct hlist_nulls_node *n;
2208 for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
2209 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
2211 nf_conntrack_lock(lockp);
2212 if (*bucket < nf_conntrack_htable_size) {
2213 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
2214 if (NF_CT_DIRECTION(h) != IP_CT_DIR_REPLY)
2216 /* All nf_conn objects are added to hash table twice, one
2217 * for original direction tuple, once for the reply tuple.
2219 * Exception: In the IPS_NAT_CLASH case, only the reply
2220 * tuple is added (the original tuple already existed for
2221 * a different object).
2223 * We only need to call the iterator once for each
2224 * conntrack, so we just use the 'reply' direction
2225 * tuple while iterating.
2227 ct = nf_ct_tuplehash_to_ctrack(h);
2239 atomic_inc(&ct->ct_general.use);
2245 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
2246 void *data, u32 portid, int report)
2248 unsigned int bucket = 0, sequence;
2254 sequence = read_seqcount_begin(&nf_conntrack_generation);
2256 while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
2257 /* Time to push up daises... */
2259 nf_ct_delete(ct, portid, report);
2264 if (!read_seqcount_retry(&nf_conntrack_generation, sequence))
2271 int (*iter)(struct nf_conn *i, void *data);
2276 static int iter_net_only(struct nf_conn *i, void *data)
2278 struct iter_data *d = data;
2280 if (!net_eq(d->net, nf_ct_net(i)))
2283 return d->iter(i, d->data);
2287 __nf_ct_unconfirmed_destroy(struct net *net)
2291 for_each_possible_cpu(cpu) {
2292 struct nf_conntrack_tuple_hash *h;
2293 struct hlist_nulls_node *n;
2294 struct ct_pcpu *pcpu;
2296 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2298 spin_lock_bh(&pcpu->lock);
2299 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
2302 ct = nf_ct_tuplehash_to_ctrack(h);
2304 /* we cannot call iter() on unconfirmed list, the
2305 * owning cpu can reallocate ct->ext at any time.
2307 set_bit(IPS_DYING_BIT, &ct->status);
2309 spin_unlock_bh(&pcpu->lock);
2314 void nf_ct_unconfirmed_destroy(struct net *net)
2316 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2320 if (atomic_read(&cnet->count) > 0) {
2321 __nf_ct_unconfirmed_destroy(net);
2322 nf_queue_nf_hook_drop(net);
2326 EXPORT_SYMBOL_GPL(nf_ct_unconfirmed_destroy);
2328 void nf_ct_iterate_cleanup_net(struct net *net,
2329 int (*iter)(struct nf_conn *i, void *data),
2330 void *data, u32 portid, int report)
2332 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2337 if (atomic_read(&cnet->count) == 0)
2344 nf_ct_iterate_cleanup(iter_net_only, &d, portid, report);
2346 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
2349 * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
2350 * @iter: callback to invoke for each conntrack
2351 * @data: data to pass to @iter
2353 * Like nf_ct_iterate_cleanup, but first marks conntracks on the
2354 * unconfirmed list as dying (so they will not be inserted into
2357 * Can only be called in module exit path.
2360 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
2364 down_read(&net_rwsem);
2366 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2368 if (atomic_read(&cnet->count) == 0)
2370 __nf_ct_unconfirmed_destroy(net);
2371 nf_queue_nf_hook_drop(net);
2373 up_read(&net_rwsem);
2375 /* Need to wait for netns cleanup worker to finish, if its
2376 * running -- it might have deleted a net namespace from
2377 * the global list, so our __nf_ct_unconfirmed_destroy() might
2378 * not have affected all namespaces.
2382 /* a conntrack could have been unlinked from unconfirmed list
2383 * before we grabbed pcpu lock in __nf_ct_unconfirmed_destroy().
2384 * This makes sure its inserted into conntrack table.
2388 nf_ct_iterate_cleanup(iter, data, 0, 0);
2390 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
2392 static int kill_all(struct nf_conn *i, void *data)
2394 return net_eq(nf_ct_net(i), data);
2397 void nf_conntrack_cleanup_start(void)
2399 conntrack_gc_work.exiting = true;
2400 RCU_INIT_POINTER(ip_ct_attach, NULL);
2403 void nf_conntrack_cleanup_end(void)
2405 RCU_INIT_POINTER(nf_ct_hook, NULL);
2406 cancel_delayed_work_sync(&conntrack_gc_work.dwork);
2407 kvfree(nf_conntrack_hash);
2409 nf_conntrack_proto_fini();
2410 nf_conntrack_seqadj_fini();
2411 nf_conntrack_labels_fini();
2412 nf_conntrack_helper_fini();
2413 nf_conntrack_timeout_fini();
2414 nf_conntrack_ecache_fini();
2415 nf_conntrack_tstamp_fini();
2416 nf_conntrack_acct_fini();
2417 nf_conntrack_expect_fini();
2419 kmem_cache_destroy(nf_conntrack_cachep);
2423 * Mishearing the voices in his head, our hero wonders how he's
2424 * supposed to kill the mall.
2426 void nf_conntrack_cleanup_net(struct net *net)
2430 list_add(&net->exit_list, &single);
2431 nf_conntrack_cleanup_net_list(&single);
2434 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
2440 * This makes sure all current packets have passed through
2441 * netfilter framework. Roll on, two-stage module
2447 list_for_each_entry(net, net_exit_list, exit_list) {
2448 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2450 nf_ct_iterate_cleanup(kill_all, net, 0, 0);
2451 if (atomic_read(&cnet->count) != 0)
2456 goto i_see_dead_people;
2459 list_for_each_entry(net, net_exit_list, exit_list) {
2460 nf_conntrack_proto_pernet_fini(net);
2461 nf_conntrack_ecache_pernet_fini(net);
2462 nf_conntrack_expect_pernet_fini(net);
2463 free_percpu(net->ct.stat);
2464 free_percpu(net->ct.pcpu_lists);
2468 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
2470 struct hlist_nulls_head *hash;
2471 unsigned int nr_slots, i;
2473 if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
2476 BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
2477 nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
2479 hash = kvcalloc(nr_slots, sizeof(struct hlist_nulls_head), GFP_KERNEL);
2482 for (i = 0; i < nr_slots; i++)
2483 INIT_HLIST_NULLS_HEAD(&hash[i], i);
2487 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
2489 int nf_conntrack_hash_resize(unsigned int hashsize)
2492 unsigned int old_size;
2493 struct hlist_nulls_head *hash, *old_hash;
2494 struct nf_conntrack_tuple_hash *h;
2500 hash = nf_ct_alloc_hashtable(&hashsize, 1);
2504 old_size = nf_conntrack_htable_size;
2505 if (old_size == hashsize) {
2511 nf_conntrack_all_lock();
2512 write_seqcount_begin(&nf_conntrack_generation);
2514 /* Lookups in the old hash might happen in parallel, which means we
2515 * might get false negatives during connection lookup. New connections
2516 * created because of a false negative won't make it into the hash
2517 * though since that required taking the locks.
2520 for (i = 0; i < nf_conntrack_htable_size; i++) {
2521 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
2522 h = hlist_nulls_entry(nf_conntrack_hash[i].first,
2523 struct nf_conntrack_tuple_hash, hnnode);
2524 ct = nf_ct_tuplehash_to_ctrack(h);
2525 hlist_nulls_del_rcu(&h->hnnode);
2526 bucket = __hash_conntrack(nf_ct_net(ct),
2527 &h->tuple, hashsize);
2528 hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
2531 old_size = nf_conntrack_htable_size;
2532 old_hash = nf_conntrack_hash;
2534 nf_conntrack_hash = hash;
2535 nf_conntrack_htable_size = hashsize;
2537 write_seqcount_end(&nf_conntrack_generation);
2538 nf_conntrack_all_unlock();
2546 int nf_conntrack_set_hashsize(const char *val, const struct kernel_param *kp)
2548 unsigned int hashsize;
2551 if (current->nsproxy->net_ns != &init_net)
2554 /* On boot, we can set this without any fancy locking. */
2555 if (!nf_conntrack_hash)
2556 return param_set_uint(val, kp);
2558 rc = kstrtouint(val, 0, &hashsize);
2562 return nf_conntrack_hash_resize(hashsize);
2565 static __always_inline unsigned int total_extension_size(void)
2567 /* remember to add new extensions below */
2568 BUILD_BUG_ON(NF_CT_EXT_NUM > 9);
2570 return sizeof(struct nf_ct_ext) +
2571 sizeof(struct nf_conn_help)
2572 #if IS_ENABLED(CONFIG_NF_NAT)
2573 + sizeof(struct nf_conn_nat)
2575 + sizeof(struct nf_conn_seqadj)
2576 + sizeof(struct nf_conn_acct)
2577 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2578 + sizeof(struct nf_conntrack_ecache)
2580 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
2581 + sizeof(struct nf_conn_tstamp)
2583 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
2584 + sizeof(struct nf_conn_timeout)
2586 #ifdef CONFIG_NF_CONNTRACK_LABELS
2587 + sizeof(struct nf_conn_labels)
2589 #if IS_ENABLED(CONFIG_NETFILTER_SYNPROXY)
2590 + sizeof(struct nf_conn_synproxy)
2595 int nf_conntrack_init_start(void)
2597 unsigned long nr_pages = totalram_pages();
2602 /* struct nf_ct_ext uses u8 to store offsets/size */
2603 BUILD_BUG_ON(total_extension_size() > 255u);
2605 seqcount_spinlock_init(&nf_conntrack_generation,
2606 &nf_conntrack_locks_all_lock);
2608 for (i = 0; i < CONNTRACK_LOCKS; i++)
2609 spin_lock_init(&nf_conntrack_locks[i]);
2611 if (!nf_conntrack_htable_size) {
2612 /* Idea from tcp.c: use 1/16384 of memory.
2613 * On i386: 32MB machine has 512 buckets.
2614 * >= 1GB machines have 16384 buckets.
2615 * >= 4GB machines have 65536 buckets.
2617 nf_conntrack_htable_size
2618 = (((nr_pages << PAGE_SHIFT) / 16384)
2619 / sizeof(struct hlist_head));
2620 if (nr_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2621 nf_conntrack_htable_size = 65536;
2622 else if (nr_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2623 nf_conntrack_htable_size = 16384;
2624 if (nf_conntrack_htable_size < 32)
2625 nf_conntrack_htable_size = 32;
2627 /* Use a max. factor of four by default to get the same max as
2628 * with the old struct list_heads. When a table size is given
2629 * we use the old value of 8 to avoid reducing the max.
2634 nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2635 if (!nf_conntrack_hash)
2638 nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2640 nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2641 sizeof(struct nf_conn),
2643 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2644 if (!nf_conntrack_cachep)
2647 ret = nf_conntrack_expect_init();
2651 ret = nf_conntrack_acct_init();
2655 ret = nf_conntrack_tstamp_init();
2659 ret = nf_conntrack_ecache_init();
2663 ret = nf_conntrack_timeout_init();
2667 ret = nf_conntrack_helper_init();
2671 ret = nf_conntrack_labels_init();
2675 ret = nf_conntrack_seqadj_init();
2679 ret = nf_conntrack_proto_init();
2683 conntrack_gc_work_init(&conntrack_gc_work);
2684 queue_delayed_work(system_power_efficient_wq, &conntrack_gc_work.dwork, HZ);
2689 nf_conntrack_seqadj_fini();
2691 nf_conntrack_labels_fini();
2693 nf_conntrack_helper_fini();
2695 nf_conntrack_timeout_fini();
2697 nf_conntrack_ecache_fini();
2699 nf_conntrack_tstamp_fini();
2701 nf_conntrack_acct_fini();
2703 nf_conntrack_expect_fini();
2705 kmem_cache_destroy(nf_conntrack_cachep);
2707 kvfree(nf_conntrack_hash);
2711 static struct nf_ct_hook nf_conntrack_hook = {
2712 .update = nf_conntrack_update,
2713 .destroy = destroy_conntrack,
2714 .get_tuple_skb = nf_conntrack_get_tuple_skb,
2717 void nf_conntrack_init_end(void)
2719 /* For use by REJECT target */
2720 RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
2721 RCU_INIT_POINTER(nf_ct_hook, &nf_conntrack_hook);
2725 * We need to use special "null" values, not used in hash table
2727 #define UNCONFIRMED_NULLS_VAL ((1<<30)+0)
2728 #define DYING_NULLS_VAL ((1<<30)+1)
2730 int nf_conntrack_init_net(struct net *net)
2732 struct nf_conntrack_net *cnet = nf_ct_pernet(net);
2736 BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2737 BUILD_BUG_ON_NOT_POWER_OF_2(CONNTRACK_LOCKS);
2738 atomic_set(&cnet->count, 0);
2740 net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
2741 if (!net->ct.pcpu_lists)
2744 for_each_possible_cpu(cpu) {
2745 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2747 spin_lock_init(&pcpu->lock);
2748 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
2749 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
2752 net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2754 goto err_pcpu_lists;
2756 ret = nf_conntrack_expect_pernet_init(net);
2760 nf_conntrack_acct_pernet_init(net);
2761 nf_conntrack_tstamp_pernet_init(net);
2762 nf_conntrack_ecache_pernet_init(net);
2763 nf_conntrack_helper_pernet_init(net);
2764 nf_conntrack_proto_pernet_init(net);
2769 free_percpu(net->ct.stat);
2771 free_percpu(net->ct.pcpu_lists);