peer: allocate in kmem_cache

With deployments having upwards of 600k peers now, this somewhat heavy
structure could benefit from more fine-grained allocations.
Specifically, instead of using a 2048-byte slab for a 1544-byte object,
we can now use 1544-byte objects directly, thus saving almost 25%
per-peer, or with 600k peers, that's a savings of 303 MiB. This also
makes wireguard's memory usage more transparent in tools like slabtop
and /proc/slabinfo.

Suggested-by: Arnd Bergmann <arnd@arndb.de>
Suggested-by: Matthew Wilcox <willy@infradead.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>

Author: zx2c4

src/main.c

@@ -33,6 +33,10 @@ static int __init mod_init(void)
 #endif
 	wg_noise_init();
 
+	ret = wg_peer_init();
+	if (ret < 0)
+		goto err_peer;
+
 	ret = wg_device_init();
 	if (ret < 0)
 		goto err_device;
@@ -49,13 +53,16 @@ static int __init mod_init(void)
 err_netlink:
 	wg_device_uninit();
 err_device:
+	wg_peer_uninit();
+err_peer:
 	return ret;
 }
 
 static void __exit mod_exit(void)
 {
 	wg_genetlink_uninit();
 	wg_device_uninit();
+	wg_peer_uninit();
 }
 
 module_init(mod_init);

src/peer.c

@@ -15,6 +15,7 @@
 #include <linux/rcupdate.h>
 #include <linux/list.h>
 
+static struct kmem_cache *peer_cache;
 static atomic64_t peer_counter = ATOMIC64_INIT(0);
 
 struct wg_peer *wg_peer_create(struct wg_device *wg,
@@ -29,10 +30,10 @@ struct wg_peer *wg_peer_create(struct wg_device *wg,
 	if (wg->num_peers >= MAX_PEERS_PER_DEVICE)
 		return ERR_PTR(ret);
 
-	peer = kzalloc(sizeof(*peer), GFP_KERNEL);
+	peer = kmem_cache_zalloc(peer_cache, GFP_KERNEL);
 	if (unlikely(!peer))
 		return ERR_PTR(ret);
-	if (dst_cache_init(&peer->endpoint_cache, GFP_KERNEL))
+	if (unlikely(dst_cache_init(&peer->endpoint_cache, GFP_KERNEL)))
 		goto err;
 
 	peer->device = wg;
@@ -64,7 +65,7 @@ struct wg_peer *wg_peer_create(struct wg_device *wg,
 	return peer;
 
 err:
-	kfree(peer);
+	kmem_cache_free(peer_cache, peer);
 	return ERR_PTR(ret);
 }
 
@@ -193,7 +194,8 @@ static void rcu_release(struct rcu_head *rcu)
 	/* The final zeroing takes care of clearing any remaining handshake key
 	 * material and other potentially sensitive information.
 	 */
-	kfree_sensitive(peer);
+	memzero_explicit(peer, sizeof(*peer));
+	kmem_cache_free(peer_cache, peer);
 }
 
 static void kref_release(struct kref *refcount)
@@ -225,3 +227,14 @@ void wg_peer_put(struct wg_peer *peer)
 		return;
 	kref_put(&peer->refcount, kref_release);
 }
+
+int __init wg_peer_init(void)
+{
+	peer_cache = KMEM_CACHE(wg_peer, 0);
+	return peer_cache ? 0 : -ENOMEM;
+}
+
+void wg_peer_uninit(void)
+{
+	kmem_cache_destroy(peer_cache);
+}

src/peer.h

@@ -80,4 +80,7 @@ void wg_peer_put(struct wg_peer *peer);
 void wg_peer_remove(struct wg_peer *peer);
 void wg_peer_remove_all(struct wg_device *wg);
 
+int wg_peer_init(void);
+void wg_peer_uninit(void);
+
 #endif /* _WG_PEER_H */