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

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 <redacted>
---
 include/net/netfilter/nf_conntrack_core.h |    8 ++---
 net/netfilter/nf_conntrack_core.c         |   55 +++++++++++++++++++++--------
 net/netfilter/nf_conntrack_helper.c       |    2 +-
 net/netfilter/nf_conntrack_netlink.c      |    2 +-
 net/netfilter/nfnetlink_cttimeout.c       |    4 +--
 5 files changed, 48 insertions(+), 23 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..3c008ce 100644
--- a/net/netfilter/nf_conntrack_core.c
+++ b/net/netfilter/nf_conntrack_core.c

@@ -66,6 +66,32 @@ 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_lock_nested(spinlock_t *lock)
+{
+	spin_lock_nested(lock, SINGLE_DEPTH_NESTING);
+	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_nested(lock, SINGLE_DEPTH_NESTING);
+	}
+}
+
 static void nf_conntrack_double_unlock(unsigned int h1, unsigned int h2)
 {
 	h1 %= CONNTRACK_LOCKS;

@@ -82,14 +108,12 @@ 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);
+			nf_conntrack_lock_nested(&nf_conntrack_locks[h2]);
 	} else {
-		spin_lock(&nf_conntrack_locks[h2]);
-		spin_lock_nested(&nf_conntrack_locks[h1],
-				 SINGLE_DEPTH_NESTING);
+		nf_conntrack_lock(&nf_conntrack_locks[h2]);
+		nf_conntrack_lock_nested(&nf_conntrack_locks[h1]);
 	}
 	if (read_seqcount_retry(&net->ct.generation, sequence)) {
 		nf_conntrack_double_unlock(h1, h2);

@@ -102,16 +126,19 @@ 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 +784,7 @@ 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 +1409,7 @@ 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 9f52729..81e3f70 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 3921d54..8a30ca6 100644
--- a/net/netfilter/nfnetlink_cttimeout.c
+++ b/net/netfilter/nfnetlink_cttimeout.c

@@ -309,12 +309,12 @@ 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

From: Sasha Levin

Sent: 05 January 2016 02:26
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).

...

+static void nf_conntrack_lock_nested(spinlock_t *lock)
+{
+	spin_lock_nested(lock, SINGLE_DEPTH_NESTING);
+	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_nested(lock, SINGLE_DEPTH_NESTING);
+	}
+}

...

If spin_lock_nested() does anything like what I think its
name suggests then I suspect that deadlocks.

	David

Sasha Levin [off-list ref] 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.

You're right.

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).

Seems both of these can be static and __read_mostly too --
as you already note resizing virtually never happens.

Otherwise:
Reviewed-by: Florian Westphal <fw@strlen.de>

On Sun, Jan 10, 2016 at 02:06:37AM +0100, Florian Westphal wrote:

Sasha Levin [off-list ref] wrote:

quoted

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).

quoted

diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
index 3cb3cb8..3c008ce 100644
--- a/net/netfilter/nf_conntrack_core.c
+++ b/net/netfilter/nf_conntrack_core.c

@@ -66,6 +66,32 @@ 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;

Seems both of these can be static and __read_mostly too --
as you already note resizing virtually never happens.

Otherwise:
Reviewed-by: Florian Westphal <fw@strlen.de>

Sasha, would you resubmit addressing Florian's feedback?

Thanks.

On 01/13/2016 11:54 AM, Pablo Neira Ayuso wrote:

On Sun, Jan 10, 2016 at 02:06:37AM +0100, Florian Westphal wrote:

quoted

quoted

Sasha Levin [off-list ref] wrote:

quoted

quoted

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).

quoted

quoted

diff --git a/net/netfilter/nf_conntrack_core.c b/net/netfilter/nf_conntrack_core.c
index 3cb3cb8..3c008ce 100644
--- a/net/netfilter/nf_conntrack_core.c
+++ b/net/netfilter/nf_conntrack_core.c

@@ -66,6 +66,32 @@ 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;

Seems both of these can be static and __read_mostly too --
as you already note resizing virtually never happens.

Otherwise:
Reviewed-by: Florian Westphal <fw@strlen.de>

Sasha, would you resubmit addressing Florian's feedback?

Yup, sorry - still catching up with the holidays vacation :(


Thanks,
Sasha

On Mon,  4 Jan 2016 21:25:46 -0500
Sasha Levin [off-list ref] 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 <redacted>
---

Looks good to me, and I like the idea.

Acked-by: Jesper Dangaard Brouer <redacted>

-- 
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