[PATCH 4.9.y 0/2] netfilter: nat: revert rhashtable conversion

[PATCH 4.9.y 0/2] netfilter: nat: revert rhashtable conversion

[Florian Westphal]

These two patches get rid of rhashtable the conversion.

First patch is an implicit dependency, rhlist remove interface
is a no-op in case element is not yet in the table, but hlist_del_rcu()
must only be called for conntracks that already have been inserted.

Second patch is the actual revert.

[PATCH 4.9.y 1/2] netfilter: nat: avoid use of nf_conn_nat extension

[Florian Westphal]

commit 6e699867f84c0f358fed233fe6162173aca28e04 upstream.

successful insert into the bysource hash sets IPS_SRC_NAT_DONE status bit
so we can check that instead of presence of nat extension which requires
extra deref.

Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
---
 net/netfilter/nf_conntrack_core.c |  2 +-
 net/netfilter/nf_nat_core.c       | 18 ++++--------------
 2 files changed, 5 insertions(+), 15 deletions(-)

diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
index ed9ce7c63252..750b8bf13e60 100644
--- a/net/netfilter/nf_conntrack_core.c
+++ b/net/netfilter/nf_conntrack_core.c
@@ -689,7 +689,7 @@ static int nf_ct_resolve_clash(struct net *net, struct sk_buff *skb,
 
 	l4proto = __nf_ct_l4proto_find(nf_ct_l3num(ct), nf_ct_protonum(ct));
 	if (l4proto->allow_clash &&
-	    !nfct_nat(ct) &&
+	    ((ct->status & IPS_NAT_DONE_MASK) == 0) &&
 	    !nf_ct_is_dying(ct) &&
 	    atomic_inc_not_zero(&ct->ct_general.use)) {
 		nf_ct_acct_merge(ct, ctinfo, (struct nf_conn *)skb->nfct);
diff --git a/net/netfilter/nf_nat_core.c b/net/netfilter/nf_nat_core.c
index 2916f4815c9c..e9fcb58d4eba 100644
--- a/net/netfilter/nf_nat_core.c
+++ b/net/netfilter/nf_nat_core.c
@@ -550,10 +550,6 @@ struct nf_nat_proto_clean {
 static int nf_nat_proto_remove(struct nf_conn *i, void *data)
 {
 	const struct nf_nat_proto_clean *clean = data;
-	struct nf_conn_nat *nat = nfct_nat(i);
-
-	if (!nat)
-		return 0;
 
 	if ((clean->l3proto && nf_ct_l3num(i) != clean->l3proto) ||
 	    (clean->l4proto && nf_ct_protonum(i) != clean->l4proto))
@@ -564,12 +560,10 @@ static int nf_nat_proto_remove(struct nf_conn *i, void *data)
 
 static int nf_nat_proto_clean(struct nf_conn *ct, void *data)
 {
-	struct nf_conn_nat *nat = nfct_nat(ct);
-
 	if (nf_nat_proto_remove(ct, data))
 		return 1;
 
-	if (!nat)
+	if ((ct->status & IPS_SRC_NAT_DONE) == 0)
 		return 0;
 
 	/* This netns is being destroyed, and conntrack has nat null binding.
@@ -705,13 +699,9 @@ EXPORT_SYMBOL_GPL(nf_nat_l3proto_unregister);
 /* No one using conntrack by the time this called. */
 static void nf_nat_cleanup_conntrack(struct nf_conn *ct)
 {
-	struct nf_conn_nat *nat = nf_ct_ext_find(ct, NF_CT_EXT_NAT);
-
-	if (!nat)
-		return;
-
-	rhltable_remove(&nf_nat_bysource_table, &ct->nat_bysource,
-			nf_nat_bysource_params);
+	if (ct->status & IPS_SRC_NAT_DONE)
+		rhltable_remove(&nf_nat_bysource_table, &ct->nat_bysource,
+				nf_nat_bysource_params);
 }
 
 static struct nf_ct_ext_type nat_extend __read_mostly = {
-- 
2.13.6

[PATCH 4.9.y 2/2] netfilter: nat: Revert "netfilter: nat: convert nat bysrc hash to rhashtable"

[Florian Westphal]

commit e1bf1687740ce1a3598a1c5e452b852ff2190682 upstream.

This reverts commit 870190a9ec9075205c0fa795a09fa931694a3ff1.

It was not a good idea. The custom hash table was a much better
fit for this purpose.

A fast lookup is not essential, in fact for most cases there is no lookup
at all because original tuple is not taken and can be used as-is.
What needs to be fast is insertion and deletion.

rhlist removal however requires a rhlist walk.
We can have thousands of entries in such a list if source port/addresses
are reused for multiple flows, if this happens removal requests are so
expensive that deletions of a few thousand flows can take several
seconds(!).

The advantages that we got from rhashtable are:
1) table auto-sizing
2) multiple locks

1) would be nice to have, but it is not essential as we have at
most one lookup per new flow, so even a million flows in the bysource
table are not a problem compared to current deletion cost.
2) is easy to add to custom hash table.

I tried to add hlist_node to rhlist to speed up rhltable_remove but this
isn't doable without changing semantics.  rhltable_remove_fast will
check that the to-be-deleted object is part of the table and that
requires a list walk that we want to avoid.

Furthermore, using hlist_node increases size of struct rhlist_head, which
in turn increases nf_conn size.

Link: https://bugzilla.kernel.org/show_bug.cgi?id=196821
Reported-by: Ivan Babrou <ibobrik@gmail.com>
Signed-off-by: Florian Westphal <fw@strlen.de>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>
---
 include/net/netfilter/nf_conntrack.h |   3 +-
 include/net/netfilter/nf_nat.h       |   1 -
 net/netfilter/nf_nat_core.c          | 134 +++++++++++++++--------------------
 3 files changed, 57 insertions(+), 81 deletions(-)

diff --git a/include/net/netfilter/nf_conntrack.h b/include/net/netfilter/nf_conntrack.h
index d9d52c020a70..9ae819e27940 100644
--- a/include/net/netfilter/nf_conntrack.h
+++ b/include/net/netfilter/nf_conntrack.h
@@ -17,7 +17,6 @@
 #include <linux/bitops.h>
 #include <linux/compiler.h>
 #include <linux/atomic.h>
-#include <linux/rhashtable.h>
 
 #include <linux/netfilter/nf_conntrack_tcp.h>
 #include <linux/netfilter/nf_conntrack_dccp.h>
@@ -101,7 +100,7 @@ struct nf_conn {
 	possible_net_t ct_net;
 
 #if IS_ENABLED(CONFIG_NF_NAT)
-	struct rhlist_head nat_bysource;
+	struct hlist_node	nat_bysource;
 #endif
 	/* all members below initialized via memset */
 	u8 __nfct_init_offset[0];
diff --git a/include/net/netfilter/nf_nat.h b/include/net/netfilter/nf_nat.h
index c327a431a6f3..02515f7ed4cc 100644
--- a/include/net/netfilter/nf_nat.h
+++ b/include/net/netfilter/nf_nat.h
@@ -1,6 +1,5 @@
 #ifndef _NF_NAT_H
 #define _NF_NAT_H
-#include <linux/rhashtable.h>
 #include <linux/netfilter_ipv4.h>
 #include <linux/netfilter/nf_nat.h>
 #include <net/netfilter/nf_conntrack_tuple.h>
diff --git a/net/netfilter/nf_nat_core.c b/net/netfilter/nf_nat_core.c
index e9fcb58d4eba..624d6e4dcd5c 100644
--- a/net/netfilter/nf_nat_core.c
+++ b/net/netfilter/nf_nat_core.c
@@ -30,19 +30,17 @@
 #include <net/netfilter/nf_conntrack_zones.h>
 #include <linux/netfilter/nf_nat.h>
 
+static DEFINE_SPINLOCK(nf_nat_lock);
+
 static DEFINE_MUTEX(nf_nat_proto_mutex);
 static const struct nf_nat_l3proto __rcu *nf_nat_l3protos[NFPROTO_NUMPROTO]
 						__read_mostly;
 static const struct nf_nat_l4proto __rcu **nf_nat_l4protos[NFPROTO_NUMPROTO]
 						__read_mostly;
 
-struct nf_nat_conn_key {
-	const struct net *net;
-	const struct nf_conntrack_tuple *tuple;
-	const struct nf_conntrack_zone *zone;
-};
-
-static struct rhltable nf_nat_bysource_table;
+static struct hlist_head *nf_nat_bysource __read_mostly;
+static unsigned int nf_nat_htable_size __read_mostly;
+static unsigned int nf_nat_hash_rnd __read_mostly;
 
 inline const struct nf_nat_l3proto *
 __nf_nat_l3proto_find(u8 family)
@@ -121,17 +119,19 @@ int nf_xfrm_me_harder(struct net *net, struct sk_buff *skb, unsigned int family)
 EXPORT_SYMBOL(nf_xfrm_me_harder);
 #endif /* CONFIG_XFRM */
 
-static u32 nf_nat_bysource_hash(const void *data, u32 len, u32 seed)
+/* We keep an extra hash for each conntrack, for fast searching. */
+static inline unsigned int
+hash_by_src(const struct net *n, const struct nf_conntrack_tuple *tuple)
 {
-	const struct nf_conntrack_tuple *t;
-	const struct nf_conn *ct = data;
+	unsigned int hash;
+
+	get_random_once(&nf_nat_hash_rnd, sizeof(nf_nat_hash_rnd));
 
-	t = &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple;
 	/* Original src, to ensure we map it consistently if poss. */
+	hash = jhash2((u32 *)&tuple->src, sizeof(tuple->src) / sizeof(u32),
+		      tuple->dst.protonum ^ nf_nat_hash_rnd ^ net_hash_mix(n));
 
-	seed ^= net_hash_mix(nf_ct_net(ct));
-	return jhash2((const u32 *)&t->src, sizeof(t->src) / sizeof(u32),
-		      t->dst.protonum ^ seed);
+	return reciprocal_scale(hash, nf_nat_htable_size);
 }
 
 /* Is this tuple already taken? (not by us) */
@@ -187,28 +187,6 @@ same_src(const struct nf_conn *ct,
 		t->src.u.all == tuple->src.u.all);
 }
 
-static int nf_nat_bysource_cmp(struct rhashtable_compare_arg *arg,
-			       const void *obj)
-{
-	const struct nf_nat_conn_key *key = arg->key;
-	const struct nf_conn *ct = obj;
-
-	if (!same_src(ct, key->tuple) ||
-	    !net_eq(nf_ct_net(ct), key->net) ||
-	    !nf_ct_zone_equal(ct, key->zone, IP_CT_DIR_ORIGINAL))
-		return 1;
-
-	return 0;
-}
-
-static struct rhashtable_params nf_nat_bysource_params = {
-	.head_offset = offsetof(struct nf_conn, nat_bysource),
-	.obj_hashfn = nf_nat_bysource_hash,
-	.obj_cmpfn = nf_nat_bysource_cmp,
-	.nelem_hint = 256,
-	.min_size = 1024,
-};
-
 /* Only called for SRC manip */
 static int
 find_appropriate_src(struct net *net,
@@ -219,26 +197,22 @@ find_appropriate_src(struct net *net,
 		     struct nf_conntrack_tuple *result,
 		     const struct nf_nat_range *range)
 {
+	unsigned int h = hash_by_src(net, tuple);
 	const struct nf_conn *ct;
-	struct nf_nat_conn_key key = {
-		.net = net,
-		.tuple = tuple,
-		.zone = zone
-	};
-	struct rhlist_head *hl, *h;
-
-	hl = rhltable_lookup(&nf_nat_bysource_table, &key,
-			     nf_nat_bysource_params);
 
-	rhl_for_each_entry_rcu(ct, h, hl, nat_bysource) {
-		nf_ct_invert_tuplepr(result,
-				     &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
-		result->dst = tuple->dst;
-
-		if (in_range(l3proto, l4proto, result, range))
-			return 1;
+	hlist_for_each_entry_rcu(ct, &nf_nat_bysource[h], nat_bysource) {
+		if (same_src(ct, tuple) &&
+		    net_eq(net, nf_ct_net(ct)) &&
+		    nf_ct_zone_equal(ct, zone, IP_CT_DIR_ORIGINAL)) {
+			/* Copy source part from reply tuple. */
+			nf_ct_invert_tuplepr(result,
+				       &ct->tuplehash[IP_CT_DIR_REPLY].tuple);
+			result->dst = tuple->dst;
+
+			if (in_range(l3proto, l4proto, result, range))
+				return 1;
+		}
 	}
-
 	return 0;
 }
 
@@ -411,6 +385,7 @@ nf_nat_setup_info(struct nf_conn *ct,
 		  const struct nf_nat_range *range,
 		  enum nf_nat_manip_type maniptype)
 {
+	struct net *net = nf_ct_net(ct);
 	struct nf_conntrack_tuple curr_tuple, new_tuple;
 	struct nf_conn_nat *nat;
 
@@ -452,19 +427,16 @@ nf_nat_setup_info(struct nf_conn *ct,
 	}
 
 	if (maniptype == NF_NAT_MANIP_SRC) {
-		struct nf_nat_conn_key key = {
-			.net = nf_ct_net(ct),
-			.tuple = &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple,
-			.zone = nf_ct_zone(ct),
-		};
-		int err;
-
-		err = rhltable_insert_key(&nf_nat_bysource_table,
-					  &key,
-					  &ct->nat_bysource,
-					  nf_nat_bysource_params);
-		if (err)
-			return NF_DROP;
+		unsigned int srchash;
+
+		srchash = hash_by_src(net,
+				      &ct->tuplehash[IP_CT_DIR_ORIGINAL].tuple);
+		spin_lock_bh(&nf_nat_lock);
+		/* nf_conntrack_alter_reply might re-allocate extension aera */
+		nat = nfct_nat(ct);
+		hlist_add_head_rcu(&ct->nat_bysource,
+				   &nf_nat_bysource[srchash]);
+		spin_unlock_bh(&nf_nat_lock);
 	}
 
 	/* It's done. */
@@ -572,9 +544,10 @@ static int nf_nat_proto_clean(struct nf_conn *ct, void *data)
 	 * Else, when the conntrack is destoyed, nf_nat_cleanup_conntrack()
 	 * will delete entry from already-freed table.
 	 */
+	spin_lock_bh(&nf_nat_lock);
+	hlist_del_rcu(&ct->nat_bysource);
 	ct->status &= ~IPS_NAT_DONE_MASK;
-	rhltable_remove(&nf_nat_bysource_table, &ct->nat_bysource,
-			nf_nat_bysource_params);
+	spin_unlock_bh(&nf_nat_lock);
 
 	/* don't delete conntrack.  Although that would make things a lot
 	 * simpler, we'd end up flushing all conntracks on nat rmmod.
@@ -699,9 +672,11 @@ EXPORT_SYMBOL_GPL(nf_nat_l3proto_unregister);
 /* No one using conntrack by the time this called. */
 static void nf_nat_cleanup_conntrack(struct nf_conn *ct)
 {
-	if (ct->status & IPS_SRC_NAT_DONE)
-		rhltable_remove(&nf_nat_bysource_table, &ct->nat_bysource,
-				nf_nat_bysource_params);
+	if (ct->status & IPS_SRC_NAT_DONE) {
+		spin_lock_bh(&nf_nat_lock);
+		hlist_del_rcu(&ct->nat_bysource);
+		spin_unlock_bh(&nf_nat_lock);
+	}
 }
 
 static struct nf_ct_ext_type nat_extend __read_mostly = {
@@ -836,13 +811,16 @@ static int __init nf_nat_init(void)
 {
 	int ret;
 
-	ret = rhltable_init(&nf_nat_bysource_table, &nf_nat_bysource_params);
-	if (ret)
-		return ret;
+	/* Leave them the same for the moment. */
+	nf_nat_htable_size = nf_conntrack_htable_size;
+
+	nf_nat_bysource = nf_ct_alloc_hashtable(&nf_nat_htable_size, 0);
+	if (!nf_nat_bysource)
+		return -ENOMEM;
 
 	ret = nf_ct_extend_register(&nat_extend);
 	if (ret < 0) {
-		rhltable_destroy(&nf_nat_bysource_table);
+		nf_ct_free_hashtable(nf_nat_bysource, nf_nat_htable_size);
 		printk(KERN_ERR "nf_nat_core: Unable to register extension\n");
 		return ret;
 	}
@@ -866,7 +844,7 @@ static int __init nf_nat_init(void)
 	return 0;
 
  cleanup_extend:
-	rhltable_destroy(&nf_nat_bysource_table);
+	nf_ct_free_hashtable(nf_nat_bysource, nf_nat_htable_size);
 	nf_ct_extend_unregister(&nat_extend);
 	return ret;
 }
@@ -886,8 +864,8 @@ static void __exit nf_nat_cleanup(void)
 
 	for (i = 0; i < NFPROTO_NUMPROTO; i++)
 		kfree(nf_nat_l4protos[i]);
-
-	rhltable_destroy(&nf_nat_bysource_table);
+	synchronize_net();
+	nf_ct_free_hashtable(nf_nat_bysource, nf_nat_htable_size);
 }
 
 MODULE_LICENSE("GPL");
-- 
2.13.6

Re: [PATCH 4.9.y 0/2] netfilter: nat: revert rhashtable conversion

[Guillaume Nault]

On Thu, Nov 16, 2017 at 11:32:54AM +0100, Florian Westphal wrote:
> These two patches get rid of rhashtable the conversion.
> 
> First patch is an implicit dependency, rhlist remove interface
> is a no-op in case element is not yet in the table, but hlist_del_rcu()
> must only be called for conntracks that already have been inserted.
> 
> Second patch is the actual revert.
> 
This version is very close to what we're running here.
Thanks for your work Florian.

Re: [PATCH 4.9.y 0/2] netfilter: nat: revert rhashtable conversion

[Greg KH]

On Thu, Nov 16, 2017 at 11:32:54AM +0100, Florian Westphal wrote:
> These two patches get rid of rhashtable the conversion.
> 
> First patch is an implicit dependency, rhlist remove interface
> is a no-op in case element is not yet in the table, but hlist_del_rcu()
> must only be called for conntracks that already have been inserted.
> 
> Second patch is the actual revert.

Thanks for the patches, now queued up.

greg k-h