1 /* Connection state tracking for netfilter. This is separated from,
2 but required by, the NAT layer; it can also be used by an iptables
5 /* (C) 1999-2001 Paul `Rusty' Russell
6 * (C) 2002-2006 Netfilter Core Team <coreteam@netfilter.org>
7 * (C) 2003,2004 USAGI/WIDE Project <http://www.linux-ipv6.org>
8 * (C) 2005-2012 Patrick McHardy <kaber@trash.net>
10 * This program is free software; you can redistribute it and/or modify
11 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
15 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
17 #include <linux/types.h>
18 #include <linux/netfilter.h>
19 #include <linux/module.h>
20 #include <linux/sched.h>
21 #include <linux/skbuff.h>
22 #include <linux/proc_fs.h>
23 #include <linux/vmalloc.h>
24 #include <linux/stddef.h>
25 #include <linux/slab.h>
26 #include <linux/random.h>
27 #include <linux/jhash.h>
28 #include <linux/siphash.h>
29 #include <linux/err.h>
30 #include <linux/percpu.h>
31 #include <linux/moduleparam.h>
32 #include <linux/notifier.h>
33 #include <linux/kernel.h>
34 #include <linux/netdevice.h>
35 #include <linux/socket.h>
37 #include <linux/nsproxy.h>
38 #include <linux/rculist_nulls.h>
40 #include <net/netfilter/nf_conntrack.h>
41 #include <net/netfilter/nf_conntrack_l3proto.h>
42 #include <net/netfilter/nf_conntrack_l4proto.h>
43 #include <net/netfilter/nf_conntrack_expect.h>
44 #include <net/netfilter/nf_conntrack_helper.h>
45 #include <net/netfilter/nf_conntrack_seqadj.h>
46 #include <net/netfilter/nf_conntrack_core.h>
47 #include <net/netfilter/nf_conntrack_extend.h>
48 #include <net/netfilter/nf_conntrack_acct.h>
49 #include <net/netfilter/nf_conntrack_ecache.h>
50 #include <net/netfilter/nf_conntrack_zones.h>
51 #include <net/netfilter/nf_conntrack_timestamp.h>
52 #include <net/netfilter/nf_conntrack_timeout.h>
53 #include <net/netfilter/nf_conntrack_labels.h>
54 #include <net/netfilter/nf_conntrack_synproxy.h>
55 #include <net/netfilter/nf_nat.h>
56 #include <net/netfilter/nf_nat_core.h>
57 #include <net/netfilter/nf_nat_helper.h>
58 #include <net/netns/hash.h>
60 #include "nf_internals.h"
62 #define NF_CONNTRACK_VERSION "0.5.0"
64 int (*nfnetlink_parse_nat_setup_hook)(struct nf_conn *ct,
65 enum nf_nat_manip_type manip,
66 const struct nlattr *attr) __read_mostly;
67 EXPORT_SYMBOL_GPL(nfnetlink_parse_nat_setup_hook);
69 __cacheline_aligned_in_smp spinlock_t nf_conntrack_locks[CONNTRACK_LOCKS];
70 EXPORT_SYMBOL_GPL(nf_conntrack_locks);
72 __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
73 EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
75 struct hlist_nulls_head *nf_conntrack_hash __read_mostly;
76 EXPORT_SYMBOL_GPL(nf_conntrack_hash);
78 struct conntrack_gc_work {
79 struct delayed_work dwork;
86 static __read_mostly struct kmem_cache *nf_conntrack_cachep;
87 static __read_mostly spinlock_t nf_conntrack_locks_all_lock;
88 static __read_mostly DEFINE_SPINLOCK(nf_conntrack_locks_all_lock);
89 static __read_mostly bool nf_conntrack_locks_all;
91 /* every gc cycle scans at most 1/GC_MAX_BUCKETS_DIV part of table */
92 #define GC_MAX_BUCKETS_DIV 128u
93 /* upper bound of full table scan */
94 #define GC_MAX_SCAN_JIFFIES (16u * HZ)
95 /* desired ratio of entries found to be expired */
96 #define GC_EVICT_RATIO 50u
98 static struct conntrack_gc_work conntrack_gc_work;
100 void nf_conntrack_lock(spinlock_t *lock) __acquires(lock)
102 /* 1) Acquire the lock */
105 /* 2) read nf_conntrack_locks_all, with ACQUIRE semantics
106 * It pairs with the smp_store_release() in nf_conntrack_all_unlock()
108 if (likely(smp_load_acquire(&nf_conntrack_locks_all) == false))
111 /* fast path failed, unlock */
114 /* Slow path 1) get global lock */
115 spin_lock(&nf_conntrack_locks_all_lock);
117 /* Slow path 2) get the lock we want */
120 /* Slow path 3) release the global lock */
121 spin_unlock(&nf_conntrack_locks_all_lock);
123 EXPORT_SYMBOL_GPL(nf_conntrack_lock);
125 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
127 h1 %= CONNTRACK_LOCKS;
128 h2 %= CONNTRACK_LOCKS;
129 spin_unlock(&nf_conntrack_locks[h1]);
131 spin_unlock(&nf_conntrack_locks[h2]);
134 /* return true if we need to recompute hashes (in case hash table was resized) */
135 static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
136 unsigned int h2, unsigned int sequence)
138 h1 %= CONNTRACK_LOCKS;
139 h2 %= CONNTRACK_LOCKS;
141 nf_conntrack_lock(&nf_conntrack_locks[h1]);
143 spin_lock_nested(&nf_conntrack_locks[h2],
144 SINGLE_DEPTH_NESTING);
146 nf_conntrack_lock(&nf_conntrack_locks[h2]);
147 spin_lock_nested(&nf_conntrack_locks[h1],
148 SINGLE_DEPTH_NESTING);
150 if (read_seqcount_retry(&nf_conntrack_generation, sequence)) {
151 nf_conntrack_double_unlock(h1, h2);
157 static void nf_conntrack_all_lock(void)
161 spin_lock(&nf_conntrack_locks_all_lock);
163 nf_conntrack_locks_all = true;
165 for (i = 0; i < CONNTRACK_LOCKS; i++) {
166 spin_lock(&nf_conntrack_locks[i]);
168 /* This spin_unlock provides the "release" to ensure that
169 * nf_conntrack_locks_all==true is visible to everyone that
170 * acquired spin_lock(&nf_conntrack_locks[]).
172 spin_unlock(&nf_conntrack_locks[i]);
176 static void nf_conntrack_all_unlock(void)
178 /* All prior stores must be complete before we clear
179 * 'nf_conntrack_locks_all'. Otherwise nf_conntrack_lock()
180 * might observe the false value but not the entire
182 * It pairs with the smp_load_acquire() in nf_conntrack_lock()
184 smp_store_release(&nf_conntrack_locks_all, false);
185 spin_unlock(&nf_conntrack_locks_all_lock);
188 unsigned int nf_conntrack_htable_size __read_mostly;
189 EXPORT_SYMBOL_GPL(nf_conntrack_htable_size);
191 unsigned int nf_conntrack_max __read_mostly;
192 seqcount_t nf_conntrack_generation __read_mostly;
193 static unsigned int nf_conntrack_hash_rnd __read_mostly;
195 static u32 hash_conntrack_raw(const struct nf_conntrack_tuple *tuple,
196 const struct net *net)
201 get_random_once(&nf_conntrack_hash_rnd, sizeof(nf_conntrack_hash_rnd));
203 /* The direction must be ignored, so we hash everything up to the
204 * destination ports (which is a multiple of 4) and treat the last
205 * three bytes manually.
207 seed = nf_conntrack_hash_rnd ^ net_hash_mix(net);
208 n = (sizeof(tuple->src) + sizeof(tuple->dst.u3)) / sizeof(u32);
209 return jhash2((u32 *)tuple, n, seed ^
210 (((__force __u16)tuple->dst.u.all << 16) |
211 tuple->dst.protonum));
214 static u32 scale_hash(u32 hash)
216 return reciprocal_scale(hash, nf_conntrack_htable_size);
219 static u32 __hash_conntrack(const struct net *net,
220 const struct nf_conntrack_tuple *tuple,
223 return reciprocal_scale(hash_conntrack_raw(tuple, net), size);
226 static u32 hash_conntrack(const struct net *net,
227 const struct nf_conntrack_tuple *tuple)
229 return scale_hash(hash_conntrack_raw(tuple, net));
233 nf_ct_get_tuple(const struct sk_buff *skb,
235 unsigned int dataoff,
239 struct nf_conntrack_tuple *tuple,
240 const struct nf_conntrack_l3proto *l3proto,
241 const struct nf_conntrack_l4proto *l4proto)
243 memset(tuple, 0, sizeof(*tuple));
245 tuple->src.l3num = l3num;
246 if (l3proto->pkt_to_tuple(skb, nhoff, tuple) == 0)
249 tuple->dst.protonum = protonum;
250 tuple->dst.dir = IP_CT_DIR_ORIGINAL;
252 return l4proto->pkt_to_tuple(skb, dataoff, net, tuple);
254 EXPORT_SYMBOL_GPL(nf_ct_get_tuple);
256 bool nf_ct_get_tuplepr(const struct sk_buff *skb, unsigned int nhoff,
258 struct net *net, struct nf_conntrack_tuple *tuple)
260 const struct nf_conntrack_l3proto *l3proto;
261 const struct nf_conntrack_l4proto *l4proto;
262 unsigned int protoff;
268 l3proto = __nf_ct_l3proto_find(l3num);
269 ret = l3proto->get_l4proto(skb, nhoff, &protoff, &protonum);
270 if (ret != NF_ACCEPT) {
275 l4proto = __nf_ct_l4proto_find(l3num, protonum);
277 ret = nf_ct_get_tuple(skb, nhoff, protoff, l3num, protonum, net, tuple,
283 EXPORT_SYMBOL_GPL(nf_ct_get_tuplepr);
286 nf_ct_invert_tuple(struct nf_conntrack_tuple *inverse,
287 const struct nf_conntrack_tuple *orig,
288 const struct nf_conntrack_l3proto *l3proto,
289 const struct nf_conntrack_l4proto *l4proto)
291 memset(inverse, 0, sizeof(*inverse));
293 inverse->src.l3num = orig->src.l3num;
294 if (l3proto->invert_tuple(inverse, orig) == 0)
297 inverse->dst.dir = !orig->dst.dir;
299 inverse->dst.protonum = orig->dst.protonum;
300 return l4proto->invert_tuple(inverse, orig);
302 EXPORT_SYMBOL_GPL(nf_ct_invert_tuple);
304 /* Generate a almost-unique pseudo-id for a given conntrack.
306 * intentionally doesn't re-use any of the seeds used for hash
307 * table location, we assume id gets exposed to userspace.
309 * Following nf_conn items do not change throughout lifetime
313 * 2. nf_conn->master address (normally NULL)
314 * 3. the associated net namespace
315 * 4. the original direction tuple
317 u32 nf_ct_get_id(const struct nf_conn *ct)
319 static __read_mostly siphash_key_t ct_id_seed;
320 unsigned long a, b, c, d;
322 net_get_random_once(&ct_id_seed, sizeof(ct_id_seed));
324 a = (unsigned long)ct;
325 b = (unsigned long)ct->master;
326 c = (unsigned long)nf_ct_net(ct);
327 d = (unsigned long)siphash(&ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
328 sizeof(ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple),
331 return siphash_4u64((u64)a, (u64)b, (u64)c, (u64)d, &ct_id_seed);
333 return siphash_4u32((u32)a, (u32)b, (u32)c, (u32)d, &ct_id_seed);
336 EXPORT_SYMBOL_GPL(nf_ct_get_id);
339 clean_from_lists(struct nf_conn *ct)
341 pr_debug("clean_from_lists(%p)\n", ct);
342 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
343 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode);
345 /* Destroy all pending expectations */
346 nf_ct_remove_expectations(ct);
349 /* must be called with local_bh_disable */
350 static void nf_ct_add_to_dying_list(struct nf_conn *ct)
352 struct ct_pcpu *pcpu;
354 /* add this conntrack to the (per cpu) dying list */
355 ct->cpu = smp_processor_id();
356 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
358 spin_lock(&pcpu->lock);
359 hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
361 spin_unlock(&pcpu->lock);
364 /* must be called with local_bh_disable */
365 static void nf_ct_add_to_unconfirmed_list(struct nf_conn *ct)
367 struct ct_pcpu *pcpu;
369 /* add this conntrack to the (per cpu) unconfirmed list */
370 ct->cpu = smp_processor_id();
371 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
373 spin_lock(&pcpu->lock);
374 hlist_nulls_add_head(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
376 spin_unlock(&pcpu->lock);
379 /* must be called with local_bh_disable */
380 static void nf_ct_del_from_dying_or_unconfirmed_list(struct nf_conn *ct)
382 struct ct_pcpu *pcpu;
384 /* We overload first tuple to link into unconfirmed or dying list.*/
385 pcpu = per_cpu_ptr(nf_ct_net(ct)->ct.pcpu_lists, ct->cpu);
387 spin_lock(&pcpu->lock);
388 BUG_ON(hlist_nulls_unhashed(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode));
389 hlist_nulls_del_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode);
390 spin_unlock(&pcpu->lock);
393 #define NFCT_ALIGN(len) (((len) + NFCT_INFOMASK) & ~NFCT_INFOMASK)
395 /* Released via destroy_conntrack() */
396 struct nf_conn *nf_ct_tmpl_alloc(struct net *net,
397 const struct nf_conntrack_zone *zone,
400 struct nf_conn *tmpl, *p;
402 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK) {
403 tmpl = kzalloc(sizeof(*tmpl) + NFCT_INFOMASK, flags);
408 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
410 tmpl = (struct nf_conn *)NFCT_ALIGN((unsigned long)p);
411 tmpl->proto.tmpl_padto = (char *)tmpl - (char *)p;
414 tmpl = kzalloc(sizeof(*tmpl), flags);
419 tmpl->status = IPS_TEMPLATE;
420 write_pnet(&tmpl->ct_net, net);
421 nf_ct_zone_add(tmpl, zone);
422 atomic_set(&tmpl->ct_general.use, 0);
426 EXPORT_SYMBOL_GPL(nf_ct_tmpl_alloc);
428 void nf_ct_tmpl_free(struct nf_conn *tmpl)
430 nf_ct_ext_destroy(tmpl);
431 nf_ct_ext_free(tmpl);
433 if (ARCH_KMALLOC_MINALIGN <= NFCT_INFOMASK)
434 kfree((char *)tmpl - tmpl->proto.tmpl_padto);
438 EXPORT_SYMBOL_GPL(nf_ct_tmpl_free);
441 destroy_conntrack(struct nf_conntrack *nfct)
443 struct nf_conn *ct = (struct nf_conn *)nfct;
444 const struct nf_conntrack_l4proto *l4proto;
446 pr_debug("destroy_conntrack(%p)\n", ct);
447 WARN_ON(atomic_read(&nfct->use) != 0);
449 if (unlikely(nf_ct_is_template(ct))) {
453 l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
454 if (l4proto->destroy)
455 l4proto->destroy(ct);
458 /* Expectations will have been removed in clean_from_lists,
459 * except TFTP can create an expectation on the first packet,
460 * before connection is in the list, so we need to clean here,
463 nf_ct_remove_expectations(ct);
465 nf_ct_del_from_dying_or_unconfirmed_list(ct);
470 nf_ct_put(ct->master);
472 pr_debug("destroy_conntrack: returning ct=%p to slab\n", ct);
473 nf_conntrack_free(ct);
476 static void nf_ct_delete_from_lists(struct nf_conn *ct)
478 struct net *net = nf_ct_net(ct);
479 unsigned int hash, reply_hash;
480 unsigned int sequence;
482 nf_ct_helper_destroy(ct);
486 sequence = read_seqcount_begin(&nf_conntrack_generation);
487 hash = hash_conntrack(net,
488 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
489 reply_hash = hash_conntrack(net,
490 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
491 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
493 clean_from_lists(ct);
494 nf_conntrack_double_unlock(hash, reply_hash);
496 nf_ct_add_to_dying_list(ct);
501 bool nf_ct_delete(struct nf_conn *ct, u32 portid, int report)
503 struct nf_conn_tstamp *tstamp;
505 if (test_and_set_bit(IPS_DYING_BIT, &ct->status))
508 tstamp = nf_conn_tstamp_find(ct);
510 s32 timeout = ct->timeout - nfct_time_stamp;
512 tstamp->stop = ktime_get_real_ns();
514 tstamp->stop -= jiffies_to_nsecs(-timeout);
517 if (nf_conntrack_event_report(IPCT_DESTROY, ct,
518 portid, report) < 0) {
519 /* destroy event was not delivered. nf_ct_put will
520 * be done by event cache worker on redelivery.
522 nf_ct_delete_from_lists(ct);
523 nf_conntrack_ecache_delayed_work(nf_ct_net(ct));
527 nf_conntrack_ecache_work(nf_ct_net(ct));
528 nf_ct_delete_from_lists(ct);
532 EXPORT_SYMBOL_GPL(nf_ct_delete);
535 nf_ct_key_equal(struct nf_conntrack_tuple_hash *h,
536 const struct nf_conntrack_tuple *tuple,
537 const struct nf_conntrack_zone *zone,
538 const struct net *net)
540 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
542 /* A conntrack can be recreated with the equal tuple,
543 * so we need to check that the conntrack is confirmed
545 return nf_ct_tuple_equal(tuple, &h->tuple) &&
546 nf_ct_zone_equal(ct, zone, NF_CT_DIRECTION(h)) &&
547 nf_ct_is_confirmed(ct) &&
548 net_eq(net, nf_ct_net(ct));
552 nf_ct_match(const struct nf_conn *ct1, const struct nf_conn *ct2)
554 return nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
555 &ct2->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
556 nf_ct_tuple_equal(&ct1->tuplehash[IP_CT_DIR_REPLY].tuple,
557 &ct2->tuplehash[IP_CT_DIR_REPLY].tuple) &&
558 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_ORIGINAL) &&
559 nf_ct_zone_equal(ct1, nf_ct_zone(ct2), IP_CT_DIR_REPLY) &&
560 net_eq(nf_ct_net(ct1), nf_ct_net(ct2));
563 /* caller must hold rcu readlock and none of the nf_conntrack_locks */
564 static void nf_ct_gc_expired(struct nf_conn *ct)
566 if (!atomic_inc_not_zero(&ct->ct_general.use))
569 if (nf_ct_should_gc(ct))
577 * - Caller must take a reference on returned object
578 * and recheck nf_ct_tuple_equal(tuple, &h->tuple)
580 static struct nf_conntrack_tuple_hash *
581 ____nf_conntrack_find(struct net *net, const struct nf_conntrack_zone *zone,
582 const struct nf_conntrack_tuple *tuple, u32 hash)
584 struct nf_conntrack_tuple_hash *h;
585 struct hlist_nulls_head *ct_hash;
586 struct hlist_nulls_node *n;
587 unsigned int bucket, hsize;
590 nf_conntrack_get_ht(&ct_hash, &hsize);
591 bucket = reciprocal_scale(hash, hsize);
593 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[bucket], hnnode) {
596 ct = nf_ct_tuplehash_to_ctrack(h);
597 if (nf_ct_is_expired(ct)) {
598 nf_ct_gc_expired(ct);
602 if (nf_ct_is_dying(ct))
605 if (nf_ct_key_equal(h, tuple, zone, net))
609 * if the nulls value we got at the end of this lookup is
610 * not the expected one, we must restart lookup.
611 * We probably met an item that was moved to another chain.
613 if (get_nulls_value(n) != bucket) {
614 NF_CT_STAT_INC_ATOMIC(net, search_restart);
621 /* Find a connection corresponding to a tuple. */
622 static struct nf_conntrack_tuple_hash *
623 __nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
624 const struct nf_conntrack_tuple *tuple, u32 hash)
626 struct nf_conntrack_tuple_hash *h;
631 h = ____nf_conntrack_find(net, zone, tuple, hash);
633 ct = nf_ct_tuplehash_to_ctrack(h);
634 if (unlikely(nf_ct_is_dying(ct) ||
635 !atomic_inc_not_zero(&ct->ct_general.use)))
638 if (unlikely(!nf_ct_key_equal(h, tuple, zone, net))) {
649 struct nf_conntrack_tuple_hash *
650 nf_conntrack_find_get(struct net *net, const struct nf_conntrack_zone *zone,
651 const struct nf_conntrack_tuple *tuple)
653 return __nf_conntrack_find_get(net, zone, tuple,
654 hash_conntrack_raw(tuple, net));
656 EXPORT_SYMBOL_GPL(nf_conntrack_find_get);
658 static void __nf_conntrack_hash_insert(struct nf_conn *ct,
660 unsigned int reply_hash)
662 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode,
663 &nf_conntrack_hash[hash]);
664 hlist_nulls_add_head_rcu(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode,
665 &nf_conntrack_hash[reply_hash]);
669 nf_conntrack_hash_check_insert(struct nf_conn *ct)
671 const struct nf_conntrack_zone *zone;
672 struct net *net = nf_ct_net(ct);
673 unsigned int hash, reply_hash;
674 struct nf_conntrack_tuple_hash *h;
675 struct hlist_nulls_node *n;
676 unsigned int sequence;
678 zone = nf_ct_zone(ct);
682 sequence = read_seqcount_begin(&nf_conntrack_generation);
683 hash = hash_conntrack(net,
684 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
685 reply_hash = hash_conntrack(net,
686 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
687 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
689 /* See if there's one in the list already, including reverse */
690 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
691 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
695 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
696 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
701 /* The caller holds a reference to this object */
702 atomic_set(&ct->ct_general.use, 2);
703 __nf_conntrack_hash_insert(ct, hash, reply_hash);
704 nf_conntrack_double_unlock(hash, reply_hash);
705 NF_CT_STAT_INC(net, insert);
710 nf_conntrack_double_unlock(hash, reply_hash);
711 NF_CT_STAT_INC(net, insert_failed);
715 EXPORT_SYMBOL_GPL(nf_conntrack_hash_check_insert);
717 static inline void nf_ct_acct_update(struct nf_conn *ct,
718 enum ip_conntrack_info ctinfo,
721 struct nf_conn_acct *acct;
723 acct = nf_conn_acct_find(ct);
725 struct nf_conn_counter *counter = acct->counter;
727 atomic64_inc(&counter[CTINFO2DIR(ctinfo)].packets);
728 atomic64_add(len, &counter[CTINFO2DIR(ctinfo)].bytes);
732 static void nf_ct_acct_merge(struct nf_conn *ct, enum ip_conntrack_info ctinfo,
733 const struct nf_conn *loser_ct)
735 struct nf_conn_acct *acct;
737 acct = nf_conn_acct_find(loser_ct);
739 struct nf_conn_counter *counter = acct->counter;
742 /* u32 should be fine since we must have seen one packet. */
743 bytes = atomic64_read(&counter[CTINFO2DIR(ctinfo)].bytes);
744 nf_ct_acct_update(ct, ctinfo, bytes);
748 /* Resolve race on insertion if this protocol allows this. */
749 static int nf_ct_resolve_clash(struct net *net, struct sk_buff *skb,
750 enum ip_conntrack_info ctinfo,
751 struct nf_conntrack_tuple_hash *h)
753 /* This is the conntrack entry already in hashes that won race. */
754 struct nf_conn *ct = nf_ct_tuplehash_to_ctrack(h);
755 const struct nf_conntrack_l4proto *l4proto;
756 enum ip_conntrack_info oldinfo;
757 struct nf_conn *loser_ct = nf_ct_get(skb, &oldinfo);
759 l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
760 if (l4proto->allow_clash &&
761 !nf_ct_is_dying(ct) &&
762 atomic_inc_not_zero(&ct->ct_general.use)) {
763 if (((ct->status & IPS_NAT_DONE_MASK) == 0) ||
764 nf_ct_match(ct, loser_ct)) {
765 nf_ct_acct_merge(ct, ctinfo, loser_ct);
766 nf_conntrack_put(&loser_ct->ct_general);
767 nf_ct_set(skb, ct, oldinfo);
772 NF_CT_STAT_INC(net, drop);
776 /* Confirm a connection given skb; places it in hash table */
778 __nf_conntrack_confirm(struct sk_buff *skb)
780 const struct nf_conntrack_zone *zone;
781 unsigned int hash, reply_hash;
782 struct nf_conntrack_tuple_hash *h;
784 struct nf_conn_help *help;
785 struct nf_conn_tstamp *tstamp;
786 struct hlist_nulls_node *n;
787 enum ip_conntrack_info ctinfo;
789 unsigned int sequence;
792 ct = nf_ct_get(skb, &ctinfo);
795 /* ipt_REJECT uses nf_conntrack_attach to attach related
796 ICMP/TCP RST packets in other direction. Actual packet
797 which created connection will be IP_CT_NEW or for an
798 expected connection, IP_CT_RELATED. */
799 if (CTINFO2DIR(ctinfo) != IP_CT_DIR_ORIGINAL)
802 zone = nf_ct_zone(ct);
806 sequence = read_seqcount_begin(&nf_conntrack_generation);
807 /* reuse the hash saved before */
808 hash = *(unsigned long *)&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev;
809 hash = scale_hash(hash);
810 reply_hash = hash_conntrack(net,
811 &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
813 } while (nf_conntrack_double_lock(net, hash, reply_hash, sequence));
815 /* We're not in hash table, and we refuse to set up related
816 * connections for unconfirmed conns. But packet copies and
817 * REJECT will give spurious warnings here.
820 /* Another skb with the same unconfirmed conntrack may
821 * win the race. This may happen for bridge(br_flood)
822 * or broadcast/multicast packets do skb_clone with
823 * unconfirmed conntrack.
825 if (unlikely(nf_ct_is_confirmed(ct))) {
827 nf_conntrack_double_unlock(hash, reply_hash);
832 pr_debug("Confirming conntrack %p\n", ct);
833 /* We have to check the DYING flag after unlink to prevent
834 * a race against nf_ct_get_next_corpse() possibly called from
835 * user context, else we insert an already 'dead' hash, blocking
836 * further use of that particular connection -JM.
838 nf_ct_del_from_dying_or_unconfirmed_list(ct);
840 if (unlikely(nf_ct_is_dying(ct))) {
841 nf_ct_add_to_dying_list(ct);
845 /* See if there's one in the list already, including reverse:
846 NAT could have grabbed it without realizing, since we're
847 not in the hash. If there is, we lost race. */
848 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[hash], hnnode)
849 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
853 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[reply_hash], hnnode)
854 if (nf_ct_key_equal(h, &ct->tuplehash[IP_CT_DIR_REPLY].tuple,
858 /* Timer relative to confirmation time, not original
859 setting time, otherwise we'd get timer wrap in
860 weird delay cases. */
861 ct->timeout += nfct_time_stamp;
862 atomic_inc(&ct->ct_general.use);
863 ct->status |= IPS_CONFIRMED;
865 /* set conntrack timestamp, if enabled. */
866 tstamp = nf_conn_tstamp_find(ct);
868 if (skb->tstamp == 0)
869 __net_timestamp(skb);
871 tstamp->start = ktime_to_ns(skb->tstamp);
873 /* Since the lookup is lockless, hash insertion must be done after
874 * starting the timer and setting the CONFIRMED bit. The RCU barriers
875 * guarantee that no other CPU can find the conntrack before the above
876 * stores are visible.
878 __nf_conntrack_hash_insert(ct, hash, reply_hash);
879 nf_conntrack_double_unlock(hash, reply_hash);
882 help = nfct_help(ct);
883 if (help && help->helper)
884 nf_conntrack_event_cache(IPCT_HELPER, ct);
886 nf_conntrack_event_cache(master_ct(ct) ?
887 IPCT_RELATED : IPCT_NEW, ct);
891 nf_ct_add_to_dying_list(ct);
892 ret = nf_ct_resolve_clash(net, skb, ctinfo, h);
894 nf_conntrack_double_unlock(hash, reply_hash);
895 NF_CT_STAT_INC(net, insert_failed);
899 EXPORT_SYMBOL_GPL(__nf_conntrack_confirm);
901 /* Returns true if a connection correspondings to the tuple (required
904 nf_conntrack_tuple_taken(const struct nf_conntrack_tuple *tuple,
905 const struct nf_conn *ignored_conntrack)
907 struct net *net = nf_ct_net(ignored_conntrack);
908 const struct nf_conntrack_zone *zone;
909 struct nf_conntrack_tuple_hash *h;
910 struct hlist_nulls_head *ct_hash;
911 unsigned int hash, hsize;
912 struct hlist_nulls_node *n;
915 zone = nf_ct_zone(ignored_conntrack);
919 nf_conntrack_get_ht(&ct_hash, &hsize);
920 hash = __hash_conntrack(net, tuple, hsize);
922 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[hash], hnnode) {
923 ct = nf_ct_tuplehash_to_ctrack(h);
925 if (ct == ignored_conntrack)
928 if (nf_ct_is_expired(ct)) {
929 nf_ct_gc_expired(ct);
933 if (nf_ct_key_equal(h, tuple, zone, net)) {
934 /* Tuple is taken already, so caller will need to find
935 * a new source port to use.
938 * If the *original tuples* are identical, then both
939 * conntracks refer to the same flow.
940 * This is a rare situation, it can occur e.g. when
941 * more than one UDP packet is sent from same socket
942 * in different threads.
944 * Let nf_ct_resolve_clash() deal with this later.
946 if (nf_ct_tuple_equal(&ignored_conntrack->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
947 &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple) &&
948 nf_ct_zone_equal(ct, zone, IP_CT_DIR_ORIGINAL))
951 NF_CT_STAT_INC_ATOMIC(net, found);
957 if (get_nulls_value(n) != hash) {
958 NF_CT_STAT_INC_ATOMIC(net, search_restart);
966 EXPORT_SYMBOL_GPL(nf_conntrack_tuple_taken);
968 #define NF_CT_EVICTION_RANGE 8
970 /* There's a small race here where we may free a just-assured
971 connection. Too bad: we're in trouble anyway. */
972 static unsigned int early_drop_list(struct net *net,
973 struct hlist_nulls_head *head)
975 struct nf_conntrack_tuple_hash *h;
976 struct hlist_nulls_node *n;
977 unsigned int drops = 0;
980 hlist_nulls_for_each_entry_rcu(h, n, head, hnnode) {
981 tmp = nf_ct_tuplehash_to_ctrack(h);
983 if (nf_ct_is_expired(tmp)) {
984 nf_ct_gc_expired(tmp);
988 if (test_bit(IPS_ASSURED_BIT, &tmp->status) ||
989 !net_eq(nf_ct_net(tmp), net) ||
993 if (!atomic_inc_not_zero(&tmp->ct_general.use))
996 /* kill only if still in same netns -- might have moved due to
997 * SLAB_TYPESAFE_BY_RCU rules.
999 * We steal the timer reference. If that fails timer has
1000 * already fired or someone else deleted it. Just drop ref
1001 * and move to next entry.
1003 if (net_eq(nf_ct_net(tmp), net) &&
1004 nf_ct_is_confirmed(tmp) &&
1005 nf_ct_delete(tmp, 0, 0))
1014 static noinline int early_drop(struct net *net, unsigned int hash)
1016 unsigned int i, bucket;
1018 for (i = 0; i < NF_CT_EVICTION_RANGE; i++) {
1019 struct hlist_nulls_head *ct_hash;
1020 unsigned int hsize, drops;
1023 nf_conntrack_get_ht(&ct_hash, &hsize);
1025 bucket = reciprocal_scale(hash, hsize);
1027 bucket = (bucket + 1) % hsize;
1029 drops = early_drop_list(net, &ct_hash[bucket]);
1033 NF_CT_STAT_ADD_ATOMIC(net, early_drop, drops);
1041 static bool gc_worker_skip_ct(const struct nf_conn *ct)
1043 return !nf_ct_is_confirmed(ct) || nf_ct_is_dying(ct);
1046 static bool gc_worker_can_early_drop(const struct nf_conn *ct)
1048 const struct nf_conntrack_l4proto *l4proto;
1050 if (!test_bit(IPS_ASSURED_BIT, &ct->status))
1053 l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
1054 if (l4proto->can_early_drop && l4proto->can_early_drop(ct))
1060 static void gc_worker(struct work_struct *work)
1062 unsigned int min_interval = max(HZ / GC_MAX_BUCKETS_DIV, 1u);
1063 unsigned int i, goal, buckets = 0, expired_count = 0;
1064 unsigned int nf_conntrack_max95 = 0;
1065 struct conntrack_gc_work *gc_work;
1066 unsigned int ratio, scanned = 0;
1067 unsigned long next_run;
1069 gc_work = container_of(work, struct conntrack_gc_work, dwork.work);
1071 goal = nf_conntrack_htable_size / GC_MAX_BUCKETS_DIV;
1072 i = gc_work->last_bucket;
1073 if (gc_work->early_drop)
1074 nf_conntrack_max95 = nf_conntrack_max / 100u * 95u;
1077 struct nf_conntrack_tuple_hash *h;
1078 struct hlist_nulls_head *ct_hash;
1079 struct hlist_nulls_node *n;
1080 unsigned int hashsz;
1081 struct nf_conn *tmp;
1086 nf_conntrack_get_ht(&ct_hash, &hashsz);
1090 hlist_nulls_for_each_entry_rcu(h, n, &ct_hash[i], hnnode) {
1093 tmp = nf_ct_tuplehash_to_ctrack(h);
1096 if (nf_ct_is_expired(tmp)) {
1097 nf_ct_gc_expired(tmp);
1102 if (nf_conntrack_max95 == 0 || gc_worker_skip_ct(tmp))
1105 net = nf_ct_net(tmp);
1106 if (atomic_read(&net->ct.count) < nf_conntrack_max95)
1109 /* need to take reference to avoid possible races */
1110 if (!atomic_inc_not_zero(&tmp->ct_general.use))
1113 if (gc_worker_skip_ct(tmp)) {
1118 if (gc_worker_can_early_drop(tmp))
1124 /* could check get_nulls_value() here and restart if ct
1125 * was moved to another chain. But given gc is best-effort
1126 * we will just continue with next hash slot.
1129 cond_resched_rcu_qs();
1130 } while (++buckets < goal);
1132 if (gc_work->exiting)
1136 * Eviction will normally happen from the packet path, and not
1137 * from this gc worker.
1139 * This worker is only here to reap expired entries when system went
1140 * idle after a busy period.
1142 * The heuristics below are supposed to balance conflicting goals:
1144 * 1. Minimize time until we notice a stale entry
1145 * 2. Maximize scan intervals to not waste cycles
1147 * Normally, expire ratio will be close to 0.
1149 * As soon as a sizeable fraction of the entries have expired
1150 * increase scan frequency.
1152 ratio = scanned ? expired_count * 100 / scanned : 0;
1153 if (ratio > GC_EVICT_RATIO) {
1154 gc_work->next_gc_run = min_interval;
1156 unsigned int max = GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV;
1158 BUILD_BUG_ON((GC_MAX_SCAN_JIFFIES / GC_MAX_BUCKETS_DIV) == 0);
1160 gc_work->next_gc_run += min_interval;
1161 if (gc_work->next_gc_run > max)
1162 gc_work->next_gc_run = max;
1165 next_run = gc_work->next_gc_run;
1166 gc_work->last_bucket = i;
1167 gc_work->early_drop = false;
1168 queue_delayed_work(system_long_wq, &gc_work->dwork, next_run);
1171 static void conntrack_gc_work_init(struct conntrack_gc_work *gc_work)
1173 INIT_DEFERRABLE_WORK(&gc_work->dwork, gc_worker);
1174 gc_work->next_gc_run = HZ;
1175 gc_work->exiting = false;
1178 static struct nf_conn *
1179 __nf_conntrack_alloc(struct net *net,
1180 const struct nf_conntrack_zone *zone,
1181 const struct nf_conntrack_tuple *orig,
1182 const struct nf_conntrack_tuple *repl,
1183 gfp_t gfp, u32 hash)
1187 /* We don't want any race condition at early drop stage */
1188 atomic_inc(&net->ct.count);
1190 if (nf_conntrack_max &&
1191 unlikely(atomic_read(&net->ct.count) > nf_conntrack_max)) {
1192 if (!early_drop(net, hash)) {
1193 if (!conntrack_gc_work.early_drop)
1194 conntrack_gc_work.early_drop = true;
1195 atomic_dec(&net->ct.count);
1196 net_warn_ratelimited("nf_conntrack: table full, dropping packet\n");
1197 return ERR_PTR(-ENOMEM);
1202 * Do not use kmem_cache_zalloc(), as this cache uses
1203 * SLAB_TYPESAFE_BY_RCU.
1205 ct = kmem_cache_alloc(nf_conntrack_cachep, gfp);
1209 spin_lock_init(&ct->lock);
1210 ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple = *orig;
1211 ct->tuplehash[IP_CT_DIR_ORIGINAL].hnnode.pprev = NULL;
1212 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *repl;
1213 /* save hash for reusing when confirming */
1214 *(unsigned long *)(&ct->tuplehash[IP_CT_DIR_REPLY].hnnode.pprev) = hash;
1216 write_pnet(&ct->ct_net, net);
1217 memset(&ct->__nfct_init_offset, 0,
1218 offsetof(struct nf_conn, proto) -
1219 offsetof(struct nf_conn, __nfct_init_offset));
1221 nf_ct_zone_add(ct, zone);
1223 /* Because we use RCU lookups, we set ct_general.use to zero before
1224 * this is inserted in any list.
1226 atomic_set(&ct->ct_general.use, 0);
1229 atomic_dec(&net->ct.count);
1230 return ERR_PTR(-ENOMEM);
1233 struct nf_conn *nf_conntrack_alloc(struct net *net,
1234 const struct nf_conntrack_zone *zone,
1235 const struct nf_conntrack_tuple *orig,
1236 const struct nf_conntrack_tuple *repl,
1239 return __nf_conntrack_alloc(net, zone, orig, repl, gfp, 0);
1241 EXPORT_SYMBOL_GPL(nf_conntrack_alloc);
1243 void nf_conntrack_free(struct nf_conn *ct)
1245 struct net *net = nf_ct_net(ct);
1247 /* A freed object has refcnt == 0, that's
1248 * the golden rule for SLAB_TYPESAFE_BY_RCU
1250 WARN_ON(atomic_read(&ct->ct_general.use) != 0);
1252 nf_ct_ext_destroy(ct);
1254 kmem_cache_free(nf_conntrack_cachep, ct);
1255 smp_mb__before_atomic();
1256 atomic_dec(&net->ct.count);
1258 EXPORT_SYMBOL_GPL(nf_conntrack_free);
1261 /* Allocate a new conntrack: we return -ENOMEM if classification
1262 failed due to stress. Otherwise it really is unclassifiable. */
1263 static noinline struct nf_conntrack_tuple_hash *
1264 init_conntrack(struct net *net, struct nf_conn *tmpl,
1265 const struct nf_conntrack_tuple *tuple,
1266 const struct nf_conntrack_l3proto *l3proto,
1267 const struct nf_conntrack_l4proto *l4proto,
1268 struct sk_buff *skb,
1269 unsigned int dataoff, u32 hash)
1272 struct nf_conn_help *help;
1273 struct nf_conntrack_tuple repl_tuple;
1274 struct nf_conntrack_ecache *ecache;
1275 struct nf_conntrack_expect *exp = NULL;
1276 const struct nf_conntrack_zone *zone;
1277 struct nf_conn_timeout *timeout_ext;
1278 struct nf_conntrack_zone tmp;
1279 unsigned int *timeouts;
1281 if (!nf_ct_invert_tuple(&repl_tuple, tuple, l3proto, l4proto)) {
1282 pr_debug("Can't invert tuple.\n");
1286 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1287 ct = __nf_conntrack_alloc(net, zone, tuple, &repl_tuple, GFP_ATOMIC,
1290 return (struct nf_conntrack_tuple_hash *)ct;
1292 if (!nf_ct_add_synproxy(ct, tmpl)) {
1293 nf_conntrack_free(ct);
1294 return ERR_PTR(-ENOMEM);
1297 timeout_ext = tmpl ? nf_ct_timeout_find(tmpl) : NULL;
1299 timeouts = nf_ct_timeout_data(timeout_ext);
1300 if (unlikely(!timeouts))
1301 timeouts = l4proto->get_timeouts(net);
1303 timeouts = l4proto->get_timeouts(net);
1306 if (!l4proto->new(ct, skb, dataoff, timeouts)) {
1307 nf_conntrack_free(ct);
1308 pr_debug("can't track with proto module\n");
1313 nf_ct_timeout_ext_add(ct, rcu_dereference(timeout_ext->timeout),
1316 nf_ct_acct_ext_add(ct, GFP_ATOMIC);
1317 nf_ct_tstamp_ext_add(ct, GFP_ATOMIC);
1318 nf_ct_labels_ext_add(ct);
1320 ecache = tmpl ? nf_ct_ecache_find(tmpl) : NULL;
1321 nf_ct_ecache_ext_add(ct, ecache ? ecache->ctmask : 0,
1322 ecache ? ecache->expmask : 0,
1326 if (net->ct.expect_count) {
1327 spin_lock(&nf_conntrack_expect_lock);
1328 exp = nf_ct_find_expectation(net, zone, tuple);
1330 pr_debug("expectation arrives ct=%p exp=%p\n",
1332 /* Welcome, Mr. Bond. We've been expecting you... */
1333 __set_bit(IPS_EXPECTED_BIT, &ct->status);
1334 /* exp->master safe, refcnt bumped in nf_ct_find_expectation */
1335 ct->master = exp->master;
1337 help = nf_ct_helper_ext_add(ct, exp->helper,
1340 rcu_assign_pointer(help->helper, exp->helper);
1343 #ifdef CONFIG_NF_CONNTRACK_MARK
1344 ct->mark = exp->master->mark;
1346 #ifdef CONFIG_NF_CONNTRACK_SECMARK
1347 ct->secmark = exp->master->secmark;
1349 NF_CT_STAT_INC(net, expect_new);
1351 spin_unlock(&nf_conntrack_expect_lock);
1354 __nf_ct_try_assign_helper(ct, tmpl, GFP_ATOMIC);
1356 /* Now it is inserted into the unconfirmed list, bump refcount */
1357 nf_conntrack_get(&ct->ct_general);
1358 nf_ct_add_to_unconfirmed_list(ct);
1364 exp->expectfn(ct, exp);
1365 nf_ct_expect_put(exp);
1368 return &ct->tuplehash[IP_CT_DIR_ORIGINAL];
1371 /* On success, returns 0, sets skb->_nfct | ctinfo */
1373 resolve_normal_ct(struct net *net, struct nf_conn *tmpl,
1374 struct sk_buff *skb,
1375 unsigned int dataoff,
1378 const struct nf_conntrack_l3proto *l3proto,
1379 const struct nf_conntrack_l4proto *l4proto)
1381 const struct nf_conntrack_zone *zone;
1382 struct nf_conntrack_tuple tuple;
1383 struct nf_conntrack_tuple_hash *h;
1384 enum ip_conntrack_info ctinfo;
1385 struct nf_conntrack_zone tmp;
1389 if (!nf_ct_get_tuple(skb, skb_network_offset(skb),
1390 dataoff, l3num, protonum, net, &tuple, l3proto,
1392 pr_debug("Can't get tuple\n");
1396 /* look for tuple match */
1397 zone = nf_ct_zone_tmpl(tmpl, skb, &tmp);
1398 hash = hash_conntrack_raw(&tuple, net);
1399 h = __nf_conntrack_find_get(net, zone, &tuple, hash);
1401 h = init_conntrack(net, tmpl, &tuple, l3proto, l4proto,
1402 skb, dataoff, hash);
1408 ct = nf_ct_tuplehash_to_ctrack(h);
1410 /* It exists; we have (non-exclusive) reference. */
1411 if (NF_CT_DIRECTION(h) == IP_CT_DIR_REPLY) {
1412 ctinfo = IP_CT_ESTABLISHED_REPLY;
1414 /* Once we've had two way comms, always ESTABLISHED. */
1415 if (test_bit(IPS_SEEN_REPLY_BIT, &ct->status)) {
1416 pr_debug("normal packet for %p\n", ct);
1417 ctinfo = IP_CT_ESTABLISHED;
1418 } else if (test_bit(IPS_EXPECTED_BIT, &ct->status)) {
1419 pr_debug("related packet for %p\n", ct);
1420 ctinfo = IP_CT_RELATED;
1422 pr_debug("new packet for %p\n", ct);
1426 nf_ct_set(skb, ct, ctinfo);
1431 nf_conntrack_in(struct net *net, u_int8_t pf, unsigned int hooknum,
1432 struct sk_buff *skb)
1434 const struct nf_conntrack_l3proto *l3proto;
1435 const struct nf_conntrack_l4proto *l4proto;
1436 struct nf_conn *ct, *tmpl;
1437 enum ip_conntrack_info ctinfo;
1438 unsigned int *timeouts;
1439 unsigned int dataoff;
1443 tmpl = nf_ct_get(skb, &ctinfo);
1444 if (tmpl || ctinfo == IP_CT_UNTRACKED) {
1445 /* Previously seen (loopback or untracked)? Ignore. */
1446 if ((tmpl && !nf_ct_is_template(tmpl)) ||
1447 ctinfo == IP_CT_UNTRACKED) {
1448 NF_CT_STAT_INC_ATOMIC(net, ignore);
1454 /* rcu_read_lock()ed by nf_hook_thresh */
1455 l3proto = __nf_ct_l3proto_find(pf);
1456 ret = l3proto->get_l4proto(skb, skb_network_offset(skb),
1457 &dataoff, &protonum);
1459 pr_debug("not prepared to track yet or error occurred\n");
1460 NF_CT_STAT_INC_ATOMIC(net, error);
1461 NF_CT_STAT_INC_ATOMIC(net, invalid);
1466 l4proto = __nf_ct_l4proto_find(pf, protonum);
1468 /* It may be an special packet, error, unclean...
1469 * inverse of the return code tells to the netfilter
1470 * core what to do with the packet. */
1471 if (l4proto->error != NULL) {
1472 ret = l4proto->error(net, tmpl, skb, dataoff, pf, hooknum);
1474 NF_CT_STAT_INC_ATOMIC(net, error);
1475 NF_CT_STAT_INC_ATOMIC(net, invalid);
1479 /* ICMP[v6] protocol trackers may assign one conntrack. */
1484 ret = resolve_normal_ct(net, tmpl, skb, dataoff, pf, protonum,
1487 /* Too stressed to deal. */
1488 NF_CT_STAT_INC_ATOMIC(net, drop);
1493 ct = nf_ct_get(skb, &ctinfo);
1495 /* Not valid part of a connection */
1496 NF_CT_STAT_INC_ATOMIC(net, invalid);
1501 /* Decide what timeout policy we want to apply to this flow. */
1502 timeouts = nf_ct_timeout_lookup(net, ct, l4proto);
1504 ret = l4proto->packet(ct, skb, dataoff, ctinfo, pf, timeouts);
1506 /* Invalid: inverse of the return code tells
1507 * the netfilter core what to do */
1508 pr_debug("nf_conntrack_in: Can't track with proto module\n");
1509 nf_conntrack_put(&ct->ct_general);
1511 NF_CT_STAT_INC_ATOMIC(net, invalid);
1512 if (ret == -NF_DROP)
1513 NF_CT_STAT_INC_ATOMIC(net, drop);
1514 /* Special case: TCP tracker reports an attempt to reopen a
1515 * closed/aborted connection. We have to go back and create a
1518 if (ret == -NF_REPEAT)
1524 if (ctinfo == IP_CT_ESTABLISHED_REPLY &&
1525 !test_and_set_bit(IPS_SEEN_REPLY_BIT, &ct->status))
1526 nf_conntrack_event_cache(IPCT_REPLY, ct);
1533 EXPORT_SYMBOL_GPL(nf_conntrack_in);
1535 bool nf_ct_invert_tuplepr(struct nf_conntrack_tuple *inverse,
1536 const struct nf_conntrack_tuple *orig)
1541 ret = nf_ct_invert_tuple(inverse, orig,
1542 __nf_ct_l3proto_find(orig->src.l3num),
1543 __nf_ct_l4proto_find(orig->src.l3num,
1544 orig->dst.protonum));
1548 EXPORT_SYMBOL_GPL(nf_ct_invert_tuplepr);
1550 /* Alter reply tuple (maybe alter helper). This is for NAT, and is
1551 implicitly racy: see __nf_conntrack_confirm */
1552 void nf_conntrack_alter_reply(struct nf_conn *ct,
1553 const struct nf_conntrack_tuple *newreply)
1555 struct nf_conn_help *help = nfct_help(ct);
1557 /* Should be unconfirmed, so not in hash table yet */
1558 WARN_ON(nf_ct_is_confirmed(ct));
1560 pr_debug("Altering reply tuple of %p to ", ct);
1561 nf_ct_dump_tuple(newreply);
1563 ct->tuplehash[IP_CT_DIR_REPLY].tuple = *newreply;
1564 if (ct->master || (help && !hlist_empty(&help->expectations)))
1568 __nf_ct_try_assign_helper(ct, NULL, GFP_ATOMIC);
1571 EXPORT_SYMBOL_GPL(nf_conntrack_alter_reply);
1573 /* Refresh conntrack for this many jiffies and do accounting if do_acct is 1 */
1574 void __nf_ct_refresh_acct(struct nf_conn *ct,
1575 enum ip_conntrack_info ctinfo,
1576 const struct sk_buff *skb,
1577 unsigned long extra_jiffies,
1582 /* Only update if this is not a fixed timeout */
1583 if (test_bit(IPS_FIXED_TIMEOUT_BIT, &ct->status))
1586 /* If not in hash table, timer will not be active yet */
1587 if (nf_ct_is_confirmed(ct))
1588 extra_jiffies += nfct_time_stamp;
1590 ct->timeout = extra_jiffies;
1593 nf_ct_acct_update(ct, ctinfo, skb->len);
1595 EXPORT_SYMBOL_GPL(__nf_ct_refresh_acct);
1597 bool nf_ct_kill_acct(struct nf_conn *ct,
1598 enum ip_conntrack_info ctinfo,
1599 const struct sk_buff *skb)
1601 nf_ct_acct_update(ct, ctinfo, skb->len);
1603 return nf_ct_delete(ct, 0, 0);
1605 EXPORT_SYMBOL_GPL(nf_ct_kill_acct);
1607 #if IS_ENABLED(CONFIG_NF_CT_NETLINK)
1609 #include <linux/netfilter/nfnetlink.h>
1610 #include <linux/netfilter/nfnetlink_conntrack.h>
1611 #include <linux/mutex.h>
1613 /* Generic function for tcp/udp/sctp/dccp and alike. This needs to be
1614 * in ip_conntrack_core, since we don't want the protocols to autoload
1615 * or depend on ctnetlink */
1616 int nf_ct_port_tuple_to_nlattr(struct sk_buff *skb,
1617 const struct nf_conntrack_tuple *tuple)
1619 if (nla_put_be16(skb, CTA_PROTO_SRC_PORT, tuple->src.u.tcp.port) ||
1620 nla_put_be16(skb, CTA_PROTO_DST_PORT, tuple->dst.u.tcp.port))
1621 goto nla_put_failure;
1627 EXPORT_SYMBOL_GPL(nf_ct_port_tuple_to_nlattr);
1629 const struct nla_policy nf_ct_port_nla_policy[CTA_PROTO_MAX+1] = {
1630 [CTA_PROTO_SRC_PORT] = { .type = NLA_U16 },
1631 [CTA_PROTO_DST_PORT] = { .type = NLA_U16 },
1633 EXPORT_SYMBOL_GPL(nf_ct_port_nla_policy);
1635 int nf_ct_port_nlattr_to_tuple(struct nlattr *tb[],
1636 struct nf_conntrack_tuple *t)
1638 if (!tb[CTA_PROTO_SRC_PORT] || !tb[CTA_PROTO_DST_PORT])
1641 t->src.u.tcp.port = nla_get_be16(tb[CTA_PROTO_SRC_PORT]);
1642 t->dst.u.tcp.port = nla_get_be16(tb[CTA_PROTO_DST_PORT]);
1646 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_to_tuple);
1648 int nf_ct_port_nlattr_tuple_size(void)
1650 return nla_policy_len(nf_ct_port_nla_policy, CTA_PROTO_MAX + 1);
1652 EXPORT_SYMBOL_GPL(nf_ct_port_nlattr_tuple_size);
1655 /* Used by ipt_REJECT and ip6t_REJECT. */
1656 static void nf_conntrack_attach(struct sk_buff *nskb, const struct sk_buff *skb)
1659 enum ip_conntrack_info ctinfo;
1661 /* This ICMP is in reverse direction to the packet which caused it */
1662 ct = nf_ct_get(skb, &ctinfo);
1663 if (CTINFO2DIR(ctinfo) == IP_CT_DIR_ORIGINAL)
1664 ctinfo = IP_CT_RELATED_REPLY;
1666 ctinfo = IP_CT_RELATED;
1668 /* Attach to new skbuff, and increment count */
1669 nf_ct_set(nskb, ct, ctinfo);
1670 nf_conntrack_get(skb_nfct(nskb));
1673 /* Bring out ya dead! */
1674 static struct nf_conn *
1675 get_next_corpse(int (*iter)(struct nf_conn *i, void *data),
1676 void *data, unsigned int *bucket)
1678 struct nf_conntrack_tuple_hash *h;
1680 struct hlist_nulls_node *n;
1683 for (; *bucket < nf_conntrack_htable_size; (*bucket)++) {
1684 lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
1686 nf_conntrack_lock(lockp);
1687 if (*bucket < nf_conntrack_htable_size) {
1688 hlist_nulls_for_each_entry(h, n, &nf_conntrack_hash[*bucket], hnnode) {
1689 if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
1691 ct = nf_ct_tuplehash_to_ctrack(h);
1703 atomic_inc(&ct->ct_general.use);
1709 static void nf_ct_iterate_cleanup(int (*iter)(struct nf_conn *i, void *data),
1710 void *data, u32 portid, int report)
1712 unsigned int bucket = 0, sequence;
1718 sequence = read_seqcount_begin(&nf_conntrack_generation);
1720 while ((ct = get_next_corpse(iter, data, &bucket)) != NULL) {
1721 /* Time to push up daises... */
1723 nf_ct_delete(ct, portid, report);
1728 if (!read_seqcount_retry(&nf_conntrack_generation, sequence))
1735 int (*iter)(struct nf_conn *i, void *data);
1740 static int iter_net_only(struct nf_conn *i, void *data)
1742 struct iter_data *d = data;
1744 if (!net_eq(d->net, nf_ct_net(i)))
1747 return d->iter(i, d->data);
1751 __nf_ct_unconfirmed_destroy(struct net *net)
1755 for_each_possible_cpu(cpu) {
1756 struct nf_conntrack_tuple_hash *h;
1757 struct hlist_nulls_node *n;
1758 struct ct_pcpu *pcpu;
1760 pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
1762 spin_lock_bh(&pcpu->lock);
1763 hlist_nulls_for_each_entry(h, n, &pcpu->unconfirmed, hnnode) {
1766 ct = nf_ct_tuplehash_to_ctrack(h);
1768 /* we cannot call iter() on unconfirmed list, the
1769 * owning cpu can reallocate ct->ext at any time.
1771 set_bit(IPS_DYING_BIT, &ct->status);
1773 spin_unlock_bh(&pcpu->lock);
1778 void nf_ct_unconfirmed_destroy(struct net *net)
1782 if (atomic_read(&net->ct.count) > 0) {
1783 __nf_ct_unconfirmed_destroy(net);
1784 nf_queue_nf_hook_drop(net);
1788 EXPORT_SYMBOL_GPL(nf_ct_unconfirmed_destroy);
1790 void nf_ct_iterate_cleanup_net(struct net *net,
1791 int (*iter)(struct nf_conn *i, void *data),
1792 void *data, u32 portid, int report)
1798 if (atomic_read(&net->ct.count) == 0)
1805 nf_ct_iterate_cleanup(iter_net_only, &d, portid, report);
1807 EXPORT_SYMBOL_GPL(nf_ct_iterate_cleanup_net);
1810 * nf_ct_iterate_destroy - destroy unconfirmed conntracks and iterate table
1811 * @iter: callback to invoke for each conntrack
1812 * @data: data to pass to @iter
1814 * Like nf_ct_iterate_cleanup, but first marks conntracks on the
1815 * unconfirmed list as dying (so they will not be inserted into
1818 * Can only be called in module exit path.
1821 nf_ct_iterate_destroy(int (*iter)(struct nf_conn *i, void *data), void *data)
1827 if (atomic_read(&net->ct.count) == 0)
1829 __nf_ct_unconfirmed_destroy(net);
1830 nf_queue_nf_hook_drop(net);
1834 /* Need to wait for netns cleanup worker to finish, if its
1835 * running -- it might have deleted a net namespace from
1836 * the global list, so our __nf_ct_unconfirmed_destroy() might
1837 * not have affected all namespaces.
1841 /* a conntrack could have been unlinked from unconfirmed list
1842 * before we grabbed pcpu lock in __nf_ct_unconfirmed_destroy().
1843 * This makes sure its inserted into conntrack table.
1847 nf_ct_iterate_cleanup(iter, data, 0, 0);
1849 EXPORT_SYMBOL_GPL(nf_ct_iterate_destroy);
1851 static int kill_all(struct nf_conn *i, void *data)
1853 return net_eq(nf_ct_net(i), data);
1856 void nf_ct_free_hashtable(void *hash, unsigned int size)
1858 if (is_vmalloc_addr(hash))
1861 free_pages((unsigned long)hash,
1862 get_order(sizeof(struct hlist_head) * size));
1864 EXPORT_SYMBOL_GPL(nf_ct_free_hashtable);
1866 void nf_conntrack_cleanup_start(void)
1868 conntrack_gc_work.exiting = true;
1869 RCU_INIT_POINTER(ip_ct_attach, NULL);
1872 void nf_conntrack_cleanup_end(void)
1874 RCU_INIT_POINTER(nf_ct_destroy, NULL);
1876 cancel_delayed_work_sync(&conntrack_gc_work.dwork);
1877 nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
1879 nf_conntrack_proto_fini();
1880 nf_conntrack_seqadj_fini();
1881 nf_conntrack_labels_fini();
1882 nf_conntrack_helper_fini();
1883 nf_conntrack_timeout_fini();
1884 nf_conntrack_ecache_fini();
1885 nf_conntrack_tstamp_fini();
1886 nf_conntrack_acct_fini();
1887 nf_conntrack_expect_fini();
1889 kmem_cache_destroy(nf_conntrack_cachep);
1893 * Mishearing the voices in his head, our hero wonders how he's
1894 * supposed to kill the mall.
1896 void nf_conntrack_cleanup_net(struct net *net)
1900 list_add(&net->exit_list, &single);
1901 nf_conntrack_cleanup_net_list(&single);
1904 void nf_conntrack_cleanup_net_list(struct list_head *net_exit_list)
1910 * This makes sure all current packets have passed through
1911 * netfilter framework. Roll on, two-stage module
1917 list_for_each_entry(net, net_exit_list, exit_list) {
1918 nf_ct_iterate_cleanup(kill_all, net, 0, 0);
1919 if (atomic_read(&net->ct.count) != 0)
1924 goto i_see_dead_people;
1927 list_for_each_entry(net, net_exit_list, exit_list) {
1928 nf_conntrack_proto_pernet_fini(net);
1929 nf_conntrack_helper_pernet_fini(net);
1930 nf_conntrack_ecache_pernet_fini(net);
1931 nf_conntrack_tstamp_pernet_fini(net);
1932 nf_conntrack_acct_pernet_fini(net);
1933 nf_conntrack_expect_pernet_fini(net);
1934 free_percpu(net->ct.stat);
1935 free_percpu(net->ct.pcpu_lists);
1939 void *nf_ct_alloc_hashtable(unsigned int *sizep, int nulls)
1941 struct hlist_nulls_head *hash;
1942 unsigned int nr_slots, i;
1945 if (*sizep > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1948 BUILD_BUG_ON(sizeof(struct hlist_nulls_head) != sizeof(struct hlist_head));
1949 nr_slots = *sizep = roundup(*sizep, PAGE_SIZE / sizeof(struct hlist_nulls_head));
1951 if (nr_slots > (UINT_MAX / sizeof(struct hlist_nulls_head)))
1954 sz = nr_slots * sizeof(struct hlist_nulls_head);
1955 hash = (void *)__get_free_pages(GFP_KERNEL | __GFP_NOWARN | __GFP_ZERO,
1961 for (i = 0; i < nr_slots; i++)
1962 INIT_HLIST_NULLS_HEAD(&hash[i], i);
1966 EXPORT_SYMBOL_GPL(nf_ct_alloc_hashtable);
1968 int nf_conntrack_hash_resize(unsigned int hashsize)
1971 unsigned int old_size;
1972 struct hlist_nulls_head *hash, *old_hash;
1973 struct nf_conntrack_tuple_hash *h;
1979 hash = nf_ct_alloc_hashtable(&hashsize, 1);
1983 old_size = nf_conntrack_htable_size;
1984 if (old_size == hashsize) {
1985 nf_ct_free_hashtable(hash, hashsize);
1990 nf_conntrack_all_lock();
1991 write_seqcount_begin(&nf_conntrack_generation);
1993 /* Lookups in the old hash might happen in parallel, which means we
1994 * might get false negatives during connection lookup. New connections
1995 * created because of a false negative won't make it into the hash
1996 * though since that required taking the locks.
1999 for (i = 0; i < nf_conntrack_htable_size; i++) {
2000 while (!hlist_nulls_empty(&nf_conntrack_hash[i])) {
2001 h = hlist_nulls_entry(nf_conntrack_hash[i].first,
2002 struct nf_conntrack_tuple_hash, hnnode);
2003 ct = nf_ct_tuplehash_to_ctrack(h);
2004 hlist_nulls_del_rcu(&h->hnnode);
2005 bucket = __hash_conntrack(nf_ct_net(ct),
2006 &h->tuple, hashsize);
2007 hlist_nulls_add_head_rcu(&h->hnnode, &hash[bucket]);
2010 old_size = nf_conntrack_htable_size;
2011 old_hash = nf_conntrack_hash;
2013 nf_conntrack_hash = hash;
2014 nf_conntrack_htable_size = hashsize;
2016 write_seqcount_end(&nf_conntrack_generation);
2017 nf_conntrack_all_unlock();
2021 nf_ct_free_hashtable(old_hash, old_size);
2025 int nf_conntrack_set_hashsize(const char *val, struct kernel_param *kp)
2027 unsigned int hashsize;
2030 if (current->nsproxy->net_ns != &init_net)
2033 /* On boot, we can set this without any fancy locking. */
2034 if (!nf_conntrack_hash)
2035 return param_set_uint(val, kp);
2037 rc = kstrtouint(val, 0, &hashsize);
2041 return nf_conntrack_hash_resize(hashsize);
2043 EXPORT_SYMBOL_GPL(nf_conntrack_set_hashsize);
2045 module_param_call(hashsize, nf_conntrack_set_hashsize, param_get_uint,
2046 &nf_conntrack_htable_size, 0600);
2048 static __always_inline unsigned int total_extension_size(void)
2050 /* remember to add new extensions below */
2051 BUILD_BUG_ON(NF_CT_EXT_NUM > 9);
2053 return sizeof(struct nf_ct_ext) +
2054 sizeof(struct nf_conn_help)
2055 #if IS_ENABLED(CONFIG_NF_NAT)
2056 + sizeof(struct nf_conn_nat)
2058 + sizeof(struct nf_conn_seqadj)
2059 + sizeof(struct nf_conn_acct)
2060 #ifdef CONFIG_NF_CONNTRACK_EVENTS
2061 + sizeof(struct nf_conntrack_ecache)
2063 #ifdef CONFIG_NF_CONNTRACK_TIMESTAMP
2064 + sizeof(struct nf_conn_tstamp)
2066 #ifdef CONFIG_NF_CONNTRACK_TIMEOUT
2067 + sizeof(struct nf_conn_timeout)
2069 #ifdef CONFIG_NF_CONNTRACK_LABELS
2070 + sizeof(struct nf_conn_labels)
2072 #if IS_ENABLED(CONFIG_NETFILTER_SYNPROXY)
2073 + sizeof(struct nf_conn_synproxy)
2078 int nf_conntrack_init_start(void)
2084 /* struct nf_ct_ext uses u8 to store offsets/size */
2085 BUILD_BUG_ON(total_extension_size() > 255u);
2087 seqcount_init(&nf_conntrack_generation);
2089 for (i = 0; i < CONNTRACK_LOCKS; i++)
2090 spin_lock_init(&nf_conntrack_locks[i]);
2092 if (!nf_conntrack_htable_size) {
2093 /* Idea from tcp.c: use 1/16384 of memory.
2094 * On i386: 32MB machine has 512 buckets.
2095 * >= 1GB machines have 16384 buckets.
2096 * >= 4GB machines have 65536 buckets.
2098 nf_conntrack_htable_size
2099 = (((totalram_pages << PAGE_SHIFT) / 16384)
2100 / sizeof(struct hlist_head));
2101 if (totalram_pages > (4 * (1024 * 1024 * 1024 / PAGE_SIZE)))
2102 nf_conntrack_htable_size = 65536;
2103 else if (totalram_pages > (1024 * 1024 * 1024 / PAGE_SIZE))
2104 nf_conntrack_htable_size = 16384;
2105 if (nf_conntrack_htable_size < 32)
2106 nf_conntrack_htable_size = 32;
2108 /* Use a max. factor of four by default to get the same max as
2109 * with the old struct list_heads. When a table size is given
2110 * we use the old value of 8 to avoid reducing the max.
2115 nf_conntrack_hash = nf_ct_alloc_hashtable(&nf_conntrack_htable_size, 1);
2116 if (!nf_conntrack_hash)
2119 nf_conntrack_max = max_factor * nf_conntrack_htable_size;
2121 nf_conntrack_cachep = kmem_cache_create("nf_conntrack",
2122 sizeof(struct nf_conn),
2124 SLAB_TYPESAFE_BY_RCU | SLAB_HWCACHE_ALIGN, NULL);
2125 if (!nf_conntrack_cachep)
2128 printk(KERN_INFO "nf_conntrack version %s (%u buckets, %d max)\n",
2129 NF_CONNTRACK_VERSION, nf_conntrack_htable_size,
2132 ret = nf_conntrack_expect_init();
2136 ret = nf_conntrack_acct_init();
2140 ret = nf_conntrack_tstamp_init();
2144 ret = nf_conntrack_ecache_init();
2148 ret = nf_conntrack_timeout_init();
2152 ret = nf_conntrack_helper_init();
2156 ret = nf_conntrack_labels_init();
2160 ret = nf_conntrack_seqadj_init();
2164 ret = nf_conntrack_proto_init();
2168 conntrack_gc_work_init(&conntrack_gc_work);
2169 queue_delayed_work(system_long_wq, &conntrack_gc_work.dwork, HZ);
2174 nf_conntrack_seqadj_fini();
2176 nf_conntrack_labels_fini();
2178 nf_conntrack_helper_fini();
2180 nf_conntrack_timeout_fini();
2182 nf_conntrack_ecache_fini();
2184 nf_conntrack_tstamp_fini();
2186 nf_conntrack_acct_fini();
2188 nf_conntrack_expect_fini();
2190 kmem_cache_destroy(nf_conntrack_cachep);
2192 nf_ct_free_hashtable(nf_conntrack_hash, nf_conntrack_htable_size);
2196 void nf_conntrack_init_end(void)
2198 /* For use by REJECT target */
2199 RCU_INIT_POINTER(ip_ct_attach, nf_conntrack_attach);
2200 RCU_INIT_POINTER(nf_ct_destroy, destroy_conntrack);
2204 * We need to use special "null" values, not used in hash table
2206 #define UNCONFIRMED_NULLS_VAL ((1<<30)+0)
2207 #define DYING_NULLS_VAL ((1<<30)+1)
2208 #define TEMPLATE_NULLS_VAL ((1<<30)+2)
2210 int nf_conntrack_init_net(struct net *net)
2215 BUILD_BUG_ON(IP_CT_UNTRACKED == IP_CT_NUMBER);
2216 atomic_set(&net->ct.count, 0);
2218 net->ct.pcpu_lists = alloc_percpu(struct ct_pcpu);
2219 if (!net->ct.pcpu_lists)
2222 for_each_possible_cpu(cpu) {
2223 struct ct_pcpu *pcpu = per_cpu_ptr(net->ct.pcpu_lists, cpu);
2225 spin_lock_init(&pcpu->lock);
2226 INIT_HLIST_NULLS_HEAD(&pcpu->unconfirmed, UNCONFIRMED_NULLS_VAL);
2227 INIT_HLIST_NULLS_HEAD(&pcpu->dying, DYING_NULLS_VAL);
2230 net->ct.stat = alloc_percpu(struct ip_conntrack_stat);
2232 goto err_pcpu_lists;
2234 ret = nf_conntrack_expect_pernet_init(net);
2237 ret = nf_conntrack_acct_pernet_init(net);
2240 ret = nf_conntrack_tstamp_pernet_init(net);
2243 ret = nf_conntrack_ecache_pernet_init(net);
2246 ret = nf_conntrack_helper_pernet_init(net);
2249 ret = nf_conntrack_proto_pernet_init(net);
2255 nf_conntrack_helper_pernet_fini(net);
2257 nf_conntrack_ecache_pernet_fini(net);
2259 nf_conntrack_tstamp_pernet_fini(net);
2261 nf_conntrack_acct_pernet_fini(net);
2263 nf_conntrack_expect_pernet_fini(net);
2265 free_percpu(net->ct.stat);
2267 free_percpu(net->ct.pcpu_lists);