[PATCH v2] netfilter: nf_conntrack: use safer way to lock all buckets

[PATCH v2] netfilter: nf_conntrack: use safer way to lock all buckets

[Sasha Levin]

When we need to lock all buckets in the connection hashtable we'd attempt to
lock 1024 spinlocks, which is way more preemption levels than supported by
the kernel. Furthermore, this behavior was hidden by checking if lockdep is
enabled, and if it was - use only 8 buckets(!).

Fix this by using a global lock and synchronize all buckets on it when we
need to lock them all. This is pretty heavyweight, but is only done when we
need to resize the hashtable, and that doesn't happen often enough (or at all).

Signed-off-by: Sasha Levin <sasha.levin@oracle.com>
---
 include/net/netfilter/nf_conntrack_core.h |    8 +++---
 net/netfilter/nf_conntrack_core.c         |   38 +++++++++++++++++++++--------
 net/netfilter/nf_conntrack_helper.c       |    2 +-
 net/netfilter/nf_conntrack_netlink.c      |    2 +-
 net/netfilter/nfnetlink_cttimeout.c       |    4 +--
 5 files changed, 35 insertions(+), 19 deletions(-)

diff --git a/include/net/netfilter/nf_conntrack_core.h b/include/net/netfilter/nf_conntrack_core.h
index 788ef58..62e17d1 100644
--- a/include/net/netfilter/nf_conntrack_core.h
+++ b/include/net/netfilter/nf_conntrack_core.h
@@ -79,12 +79,10 @@ print_tuple(struct seq_file *s, const struct nf_conntrack_tuple *tuple,
             const struct nf_conntrack_l3proto *l3proto,
             const struct nf_conntrack_l4proto *proto);
 
-#ifdef CONFIG_LOCKDEP
-# define CONNTRACK_LOCKS 8
-#else
-# define CONNTRACK_LOCKS 1024
-#endif
+#define CONNTRACK_LOCKS 1024
+
 extern spinlock_t nf_conntrack_locks[CONNTRACK_LOCKS];
+void nf_conntrack_lock(spinlock_t *lock);
 
 extern spinlock_t nf_conntrack_expect_lock;
 
diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
index 3cb3cb8..4ccf5ad 100644
--- a/net/netfilter/nf_conntrack_core.c
+++ b/net/netfilter/nf_conntrack_core.c
@@ -66,6 +66,21 @@ EXPORT_SYMBOL_GPL(nf_conntrack_locks);
 __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
 EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
 
+spinlock_t nf_conntrack_locks_all_lock;
+bool nf_conntrack_locks_all;
+
+void nf_conntrack_lock(spinlock_t *lock)
+{
+	spin_lock(lock);
+	while (unlikely(nf_conntrack_locks_all)) {
+		spin_unlock(lock);
+		spin_lock(&nf_conntrack_locks_all_lock);
+		spin_unlock(&nf_conntrack_locks_all_lock);
+		spin_lock(lock);
+	}
+}
+EXPORT_SYMBOL_GPL(nf_conntrack_lock);
+
 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
 {
 	h1 %= CONNTRACK_LOCKS;
@@ -82,12 +97,12 @@ static bool nf_conntrack_double_lock(struct net *net, unsigned int h1,
 	h1 %= CONNTRACK_LOCKS;
 	h2 %= CONNTRACK_LOCKS;
 	if (h1 <= h2) {
-		spin_lock(&nf_conntrack_locks[h1]);
+		nf_conntrack_lock(&nf_conntrack_locks[h1]);
 		if (h1 != h2)
 			spin_lock_nested(&nf_conntrack_locks[h2],
 					 SINGLE_DEPTH_NESTING);
 	} else {
-		spin_lock(&nf_conntrack_locks[h2]);
+		nf_conntrack_lock(&nf_conntrack_locks[h2]);
 		spin_lock_nested(&nf_conntrack_locks[h1],
 				 SINGLE_DEPTH_NESTING);
 	}
@@ -102,16 +117,19 @@ static void nf_conntrack_all_lock(void)
 {
 	int i;
 
-	for (i = 0; i < CONNTRACK_LOCKS; i++)
-		spin_lock_nested(&nf_conntrack_locks[i], i);
+	spin_lock(&nf_conntrack_locks_all_lock);
+	nf_conntrack_locks_all = true;
+
+	for (i = 0; i < CONNTRACK_LOCKS; i++) {
+		spin_lock(&nf_conntrack_locks[i]);
+		spin_unlock(&nf_conntrack_locks[i]);
+	}
 }
 
 static void nf_conntrack_all_unlock(void)
 {
-	int i;
-
-	for (i = 0; i < CONNTRACK_LOCKS; i++)
-		spin_unlock(&nf_conntrack_locks[i]);
+	nf_conntrack_locks_all = false;
+	spin_unlock(&nf_conntrack_locks_all_lock);
 }
 
 unsigned int nf_conntrack_htable_size __read_mostly;
@@ -757,7 +775,7 @@ restart:
 	hash = hash_bucket(_hash, net);
 	for (; i < net->ct.htable_size; i++) {
 		lockp = &nf_conntrack_locks[hash % CONNTRACK_LOCKS];
-		spin_lock(lockp);
+		nf_conntrack_lock(lockp);
 		if (read_seqcount_retry(&net->ct.generation, sequence)) {
 			spin_unlock(lockp);
 			goto restart;
@@ -1382,7 +1400,7 @@ get_next_corpse(struct net *net, int (*iter)(struct nf_conn *i, void *data),
 	for (; *bucket < net->ct.htable_size; (*bucket)++) {
 		lockp = &nf_conntrack_locks[*bucket % CONNTRACK_LOCKS];
 		local_bh_disable();
-		spin_lock(lockp);
+		nf_conntrack_lock(lockp);
 		if (*bucket < net->ct.htable_size) {
 			hlist_nulls_for_each_entry(h, n, &net->ct.hash[*bucket], hnnode) {
 				if (NF_CT_DIRECTION(h) != IP_CT_DIR_ORIGINAL)
diff --git a/net/netfilter/nf_conntrack_helper.c b/net/netfilter/nf_conntrack_helper.c
index bd9d315..3b40ec5 100644
--- a/net/netfilter/nf_conntrack_helper.c
+++ b/net/netfilter/nf_conntrack_helper.c
@@ -425,7 +425,7 @@ static void __nf_conntrack_helper_unregister(struct nf_conntrack_helper *me,
 	}
 	local_bh_disable();
 	for (i = 0; i < net->ct.htable_size; i++) {
-		spin_lock(&nf_conntrack_locks[i % CONNTRACK_LOCKS]);
+		nf_conntrack_lock(&nf_conntrack_locks[i % CONNTRACK_LOCKS]);
 		if (i < net->ct.htable_size) {
 			hlist_nulls_for_each_entry(h, nn, &net->ct.hash[i], hnnode)
 				unhelp(h, me);
diff --git a/net/netfilter/nf_conntrack_netlink.c b/net/netfilter/nf_conntrack_netlink.c
index dbb1bb3..355e855 100644
--- a/net/netfilter/nf_conntrack_netlink.c
+++ b/net/netfilter/nf_conntrack_netlink.c
@@ -840,7 +840,7 @@ ctnetlink_dump_table(struct sk_buff *skb, struct netlink_callback *cb)
 	for (; cb->args[0] < net->ct.htable_size; cb->args[0]++) {
 restart:
 		lockp = &nf_conntrack_locks[cb->args[0] % CONNTRACK_LOCKS];
-		spin_lock(lockp);
+		nf_conntrack_lock(lockp);
 		if (cb->args[0] >= net->ct.htable_size) {
 			spin_unlock(lockp);
 			goto out;
diff --git a/net/netfilter/nfnetlink_cttimeout.c b/net/netfilter/nfnetlink_cttimeout.c
index 5d010f2..94837d2 100644
--- a/net/netfilter/nfnetlink_cttimeout.c
+++ b/net/netfilter/nfnetlink_cttimeout.c
@@ -307,12 +307,12 @@ static void ctnl_untimeout(struct net *net, struct ctnl_timeout *timeout)
 
 	local_bh_disable();
 	for (i = 0; i < net->ct.htable_size; i++) {
-		spin_lock(&nf_conntrack_locks[i % CONNTRACK_LOCKS]);
+		nf_conntrack_lock(&nf_conntrack_locks[i % CONNTRACK_LOCKS]);
 		if (i < net->ct.htable_size) {
 			hlist_nulls_for_each_entry(h, nn, &net->ct.hash[i], hnnode)
 				untimeout(h, timeout);
 		}
-		spin_unlock(&nf_conntrack_locks[i % CONNTRACK_LOCKS]);
+		nf_conntrack_lock(&nf_conntrack_locks[i % CONNTRACK_LOCKS]);
 	}
 	local_bh_enable();
 }
-- 
1.7.10.4

Re: [PATCH v2] netfilter: nf_conntrack: use safer way to lock all buckets

[Florian Westphal]

Sasha Levin <sasha.levin@oracle.com> wrote:
> When we need to lock all buckets in the connection hashtable we'd attempt to
> lock 1024 spinlocks, which is way more preemption levels than supported by
> the kernel. Furthermore, this behavior was hidden by checking if lockdep is
> enabled, and if it was - use only 8 buckets(!).
> 
> Fix this by using a global lock and synchronize all buckets on it when we
> need to lock them all. This is pretty heavyweight, but is only done when we
> need to resize the hashtable, and that doesn't happen often enough (or at all).
> 
> diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
> index 3cb3cb8..4ccf5ad 100644
> --- a/net/netfilter/nf_conntrack_core.c
> +++ b/net/netfilter/nf_conntrack_core.c
> @@ -66,6 +66,21 @@ EXPORT_SYMBOL_GPL(nf_conntrack_locks);
>  __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
>  EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
>  
> +spinlock_t nf_conntrack_locks_all_lock;
> +bool nf_conntrack_locks_all;

warning: symbol 'nf_conntrack_locks_all_lock' was not declared. Should it be static?
warning: symbol 'nf_conntrack_locks_all' was not declared.  Should it be static?

> +void nf_conntrack_lock(spinlock_t *lock)

__acquires(lock)

It avoids 'context imbalance in 'nf_conntrack_lock' - wrong count at exit'
sparse error.

Other than that this looks good, thanks.

Re: [PATCH v2] netfilter: nf_conntrack: use safer way to lock all buckets

[Jesper Dangaard Brouer]

On Thu, 14 Jan 2016 12:03:56 -0500
Sasha Levin <sasha.levin@oracle.com> wrote:

> When we need to lock all buckets in the connection hashtable we'd attempt to
> lock 1024 spinlocks, which is way more preemption levels than supported by
> the kernel. Furthermore, this behavior was hidden by checking if lockdep is
> enabled, and if it was - use only 8 buckets(!).
> 
> Fix this by using a global lock and synchronize all buckets on it when we
> need to lock them all. This is pretty heavyweight, but is only done when we
> need to resize the hashtable, and that doesn't happen often enough (or at all).
> 
> Signed-off-by: Sasha Levin <sasha.levin@oracle.com>

[...]
> diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
> index 3cb3cb8..4ccf5ad 100644
> --- a/net/netfilter/nf_conntrack_core.c
> +++ b/net/netfilter/nf_conntrack_core.c
> @@ -66,6 +66,21 @@ EXPORT_SYMBOL_GPL(nf_conntrack_locks);
>  __cacheline_aligned_in_smp DEFINE_SPINLOCK(nf_conntrack_expect_lock);
>  EXPORT_SYMBOL_GPL(nf_conntrack_expect_lock);
>  
> +spinlock_t nf_conntrack_locks_all_lock;
> +bool nf_conntrack_locks_all;

You forgot to mark these static and __read_mostly as Florian pointed out...

After that you can add my:
 Acked-by: Jesper Dangaard Brouer <brouer@redhat.com>

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  Author of http://www.iptv-analyzer.org
  LinkedIn: http://www.linkedin.com/in/brouer