Revert "netfilter: x_tables: Switch synchronization to RCU"

This reverts commit cc00bcaa589914096edef7fb87ca5cee4a166b5c.

This (and the preceding) patch basically re-implemented the RCU
mechanisms of patch 784544739a25. That patch was replaced because of the
performance problems that it created when replacing tables. Now, we have
the same issue: the call to synchronize_rcu() makes replacing tables
slower by as much as an order of magnitude.

Prior to using RCU a script calling "iptables" approx. 200 times was
taking 1.16s. With RCU this increased to 11.59s.

Revert these patches and fix the issue in a different way.

Signed-off-by: Mark Tomlinson <mark.tomlinson@alliedtelesis.co.nz>
Signed-off-by: Pablo Neira Ayuso <pablo@netfilter.org>

1 files changed, 34 insertions, 15 deletions

diff --git a/net/netfilter/x_tables.c b/net/netfilter/x_tables.c
index bce6ca203d46..7df3aef39c5c 100644
--- a/net/netfilter/x_tables.c
+++ b/net/netfilter/x_tables.c
@@ -1351,14 +1351,6 @@ struct xt_counters *xt_counters_alloc(unsigned int counters)
 }
 EXPORT_SYMBOL(xt_counters_alloc);
 
-struct xt_table_info
-*xt_table_get_private_protected(const struct xt_table *table)
-{
-	return rcu_dereference_protected(table->private,
-					 mutex_is_locked(&xt[table->af].mutex));
-}
-EXPORT_SYMBOL(xt_table_get_private_protected);
-
 struct xt_table_info *
 xt_replace_table(struct xt_table *table,
 	      unsigned int num_counters,
@@ -1366,6 +1358,7 @@ xt_replace_table(struct xt_table *table,
 	      int *error)
 {
 	struct xt_table_info *private;
+	unsigned int cpu;
 	int ret;
 
 	ret = xt_jumpstack_alloc(newinfo);
@@ -1375,20 +1368,47 @@ xt_replace_table(struct xt_table *table,
 	}
 
 	/* Do the substitution. */
-	private = xt_table_get_private_protected(table);
+	local_bh_disable();
+	private = table->private;
 
 	/* Check inside lock: is the old number correct? */
 	if (num_counters != private->number) {
 		pr_debug("num_counters != table->private->number (%u/%u)\n",
 			 num_counters, private->number);
+		local_bh_enable();
 		*error = -EAGAIN;
 		return NULL;
 	}
 
 	newinfo->initial_entries = private->initial_entries;
+	/*
+	 * Ensure contents of newinfo are visible before assigning to
+	 * private.
+	 */
+	smp_wmb();
+	table->private = newinfo;
+
+	/* make sure all cpus see new ->private value */
+	smp_wmb();
 
-	rcu_assign_pointer(table->private, newinfo);
-	synchronize_rcu();
+	/*
+	 * Even though table entries have now been swapped, other CPU's
+	 * may still be using the old entries...
+	 */
+	local_bh_enable();
+
+	/* ... so wait for even xt_recseq on all cpus */
+	for_each_possible_cpu(cpu) {
+		seqcount_t *s = &per_cpu(xt_recseq, cpu);
+		u32 seq = raw_read_seqcount(s);
+
+		if (seq & 1) {
+			do {
+				cond_resched();
+				cpu_relax();
+			} while (seq == raw_read_seqcount(s));
+		}
+	}
 
 	audit_log_nfcfg(table->name, table->af, private->number,
 			!private->number ? AUDIT_XT_OP_REGISTER :
@@ -1424,12 +1444,12 @@ struct xt_table *xt_register_table(struct net *net,
 	}
 
 	/* Simplifies replace_table code. */
-	rcu_assign_pointer(table->private, bootstrap);
+	table->private = bootstrap;
 
 	if (!xt_replace_table(table, 0, newinfo, &ret))
 		goto unlock;
 
-	private = xt_table_get_private_protected(table);
+	private = table->private;
 	pr_debug("table->private->number = %u\n", private->number);
 
 	/* save number of initial entries */
@@ -1452,8 +1472,7 @@ void *xt_unregister_table(struct xt_table *table)
 	struct xt_table_info *private;
 
 	mutex_lock(&xt[table->af].mutex);
-	private = xt_table_get_private_protected(table);
-	RCU_INIT_POINTER(table->private, NULL);
+	private = table->private;
 	list_del(&table->list);
 	mutex_unlock(&xt[table->af].mutex);
 	audit_log_nfcfg(table->name, table->af, private->number,