xfrm: Add an inbound percpu state cache.

Now that we can have percpu xfrm states, the number of active
states might increase. To get a better lookup performance,
we add a percpu cache to cache the used inbound xfrm states.

Signed-off-by: Steffen Klassert <[e-mail]>
Tested-by: Antony Antony <[e-mail]>
Tested-by: Tobias Brunner <[e-mail]>

include/net/netns/xfrm.h

···
 	struct hlist_head	__rcu *state_bysrc;
 	struct hlist_head	__rcu *state_byspi;
 	struct hlist_head	__rcu *state_byseq;
+	struct hlist_head	 __percpu *state_cache_input;
 	unsigned int		state_hmask;
 	unsigned int		state_num;
 	struct work_struct	state_hash_work;

include/net/xfrm.h

···
 	struct hlist_node	byspi;
 	struct hlist_node	byseq;
 	struct hlist_node	state_cache;
+	struct hlist_node	state_cache_input;
 
 	refcount_t		refcnt;
 	spinlock_t		lock;
···
 struct xfrm_state *xfrm_state_lookup(struct net *net, u32 mark,
 				     const xfrm_address_t *daddr, __be32 spi,
 				     u8 proto, unsigned short family);
+struct xfrm_state *xfrm_input_state_lookup(struct net *net, u32 mark,
+					   const xfrm_address_t *daddr,
+					   __be32 spi, u8 proto,
+					   unsigned short family);
 struct xfrm_state *xfrm_state_lookup_byaddr(struct net *net, u32 mark,
 					    const xfrm_address_t *daddr,
 					    const xfrm_address_t *saddr,

net/ipv4/esp4_offload.c

···
 		if (sp->len == XFRM_MAX_DEPTH)
 			goto out_reset;
 
-		x = xfrm_state_lookup(dev_net(skb->dev), skb->mark,
-				      (xfrm_address_t *)&ip_hdr(skb)->daddr,
-				      spi, IPPROTO_ESP, AF_INET);
+		x = xfrm_input_state_lookup(dev_net(skb->dev), skb->mark,
+					    (xfrm_address_t *)&ip_hdr(skb)->daddr,
+					    spi, IPPROTO_ESP, AF_INET);
 
 		if (unlikely(x && x->dir && x->dir != XFRM_SA_DIR_IN)) {
 			/* non-offload path will record the error and audit log */

net/ipv6/esp6_offload.c

···
 		if (sp->len == XFRM_MAX_DEPTH)
 			goto out_reset;
 
-		x = xfrm_state_lookup(dev_net(skb->dev), skb->mark,
-				      (xfrm_address_t *)&ipv6_hdr(skb)->daddr,
-				      spi, IPPROTO_ESP, AF_INET6);
+		x = xfrm_input_state_lookup(dev_net(skb->dev), skb->mark,
+					    (xfrm_address_t *)&ipv6_hdr(skb)->daddr,
+					    spi, IPPROTO_ESP, AF_INET6);
 
 		if (unlikely(x && x->dir && x->dir != XFRM_SA_DIR_IN)) {
 			/* non-offload path will record the error and audit log */

net/xfrm/xfrm_input.c

···
 			goto drop;
 		}
 
-		x = xfrm_state_lookup(net, mark, daddr, spi, nexthdr, family);
+		x = xfrm_input_state_lookup(net, mark, daddr, spi, nexthdr, family);
 		if (x == NULL) {
 			secpath_reset(skb);
 			XFRM_INC_STATS(net, LINUX_MIB_XFRMINNOSTATES);

net/xfrm/xfrm_state.c

···
 			hlist_del_rcu(&x->byseq);
 		if (!hlist_unhashed(&x->state_cache))
 			hlist_del_rcu(&x->state_cache);
+		if (!hlist_unhashed(&x->state_cache_input))
+			hlist_del_rcu(&x->state_cache_input);
+
 		if (x->id.spi)
 			hlist_del_rcu(&x->byspi);
 		net->xfrm.state_num--;
···
 
 	return NULL;
 }
+
+struct xfrm_state *xfrm_input_state_lookup(struct net *net, u32 mark,
+					   const xfrm_address_t *daddr,
+					   __be32 spi, u8 proto,
+					   unsigned short family)
+{
+	struct hlist_head *state_cache_input;
+	struct xfrm_state *x = NULL;
+	int cpu = get_cpu();
+
+	state_cache_input =  per_cpu_ptr(net->xfrm.state_cache_input, cpu);
+
+	rcu_read_lock();
+	hlist_for_each_entry_rcu(x, state_cache_input, state_cache_input) {
+		if (x->props.family != family ||
+		    x->id.spi       != spi ||
+		    x->id.proto     != proto ||
+		    !xfrm_addr_equal(&x->id.daddr, daddr, family))
+			continue;
+
+		if ((mark & x->mark.m) != x->mark.v)
+			continue;
+		if (!xfrm_state_hold_rcu(x))
+			continue;
+		goto out;
+	}
+
+	x = __xfrm_state_lookup(net, mark, daddr, spi, proto, family);
+
+	if (x && x->km.state == XFRM_STATE_VALID) {
+		spin_lock_bh(&net->xfrm.xfrm_state_lock);
+		if (hlist_unhashed(&x->state_cache_input)) {
+			hlist_add_head_rcu(&x->state_cache_input, state_cache_input);
+		} else {
+			hlist_del_rcu(&x->state_cache_input);
+			hlist_add_head_rcu(&x->state_cache_input, state_cache_input);
+		}
+		spin_unlock_bh(&net->xfrm.xfrm_state_lock);
+	}
+
+out:
+	rcu_read_unlock();
+	put_cpu();
+	return x;
+}
+EXPORT_SYMBOL(xfrm_input_state_lookup);
 
 static struct xfrm_state *__xfrm_state_lookup_byaddr(struct net *net, u32 mark,
 						     const xfrm_address_t *daddr,
···
 	net->xfrm.state_byseq = xfrm_hash_alloc(sz);
 	if (!net->xfrm.state_byseq)
 		goto out_byseq;
+
+	net->xfrm.state_cache_input = alloc_percpu(struct hlist_head);
+	if (!net->xfrm.state_cache_input)
+		goto out_state_cache_input;
+
 	net->xfrm.state_hmask = ((sz / sizeof(struct hlist_head)) - 1);
 
 	net->xfrm.state_num = 0;
···
 			       &net->xfrm.xfrm_state_lock);
 	return 0;
 
+out_state_cache_input:
+	xfrm_hash_free(net->xfrm.state_byseq, sz);
 out_byseq:
 	xfrm_hash_free(net->xfrm.state_byspi, sz);
 out_byspi:
···
 	xfrm_hash_free(net->xfrm.state_bysrc, sz);
 	WARN_ON(!hlist_empty(net->xfrm.state_bydst));
 	xfrm_hash_free(net->xfrm.state_bydst, sz);
+	free_percpu(net->xfrm.state_cache_input);
 }
 
 #ifdef CONFIG_AUDITSYSCALL