blob: 65b31b05f55629d4259913c71f3cc2662aa56016 [file] [log] [blame]
b.liue9582032025-04-17 19:18:16 +08001From 0000000000000000000000000000000000000000 Mon Sep 17 00:00:00 2001
2From: "Jason A. Donenfeld" <Jason@zx2c4.com>
3Date: Fri, 4 Jun 2021 17:17:37 +0200
4Subject: [PATCH] wireguard: allowedips: allocate nodes in kmem_cache
5
6commit dc680de28ca849dfe589dc15ac56d22505f0ef11 upstream.
7
8The previous commit moved from O(n) to O(1) for removal, but in the
9process introduced an additional pointer member to a struct that
10increased the size from 60 to 68 bytes, putting nodes in the 128-byte
11slab. With deployed systems having as many as 2 million nodes, this
12represents a significant doubling in memory usage (128 MiB -> 256 MiB).
13Fix this by using our own kmem_cache, that's sized exactly right. This
14also makes wireguard's memory usage more transparent in tools like
15slabtop and /proc/slabinfo.
16
17Fixes: e7096c131e51 ("net: WireGuard secure network tunnel")
18Suggested-by: Arnd Bergmann <arnd@arndb.de>
19Suggested-by: Matthew Wilcox <willy@infradead.org>
20Cc: stable@vger.kernel.org
21Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
22Signed-off-by: David S. Miller <davem@davemloft.net>
23Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
24---
25 drivers/net/wireguard/allowedips.c | 31 ++++++++++++++++++++++++------
26 drivers/net/wireguard/allowedips.h | 5 ++++-
27 drivers/net/wireguard/main.c | 10 +++++++++-
28 3 files changed, 38 insertions(+), 8 deletions(-)
29
30--- a/drivers/net/wireguard/allowedips.c
31+++ b/drivers/net/wireguard/allowedips.c
32@@ -6,6 +6,8 @@
33 #include "allowedips.h"
34 #include "peer.h"
35
36+static struct kmem_cache *node_cache;
37+
38 static void swap_endian(u8 *dst, const u8 *src, u8 bits)
39 {
40 if (bits == 32) {
41@@ -40,6 +42,11 @@ static void push_rcu(struct allowedips_n
42 }
43 }
44
45+static void node_free_rcu(struct rcu_head *rcu)
46+{
47+ kmem_cache_free(node_cache, container_of(rcu, struct allowedips_node, rcu));
48+}
49+
50 static void root_free_rcu(struct rcu_head *rcu)
51 {
52 struct allowedips_node *node, *stack[128] = {
53@@ -49,7 +56,7 @@ static void root_free_rcu(struct rcu_hea
54 while (len > 0 && (node = stack[--len])) {
55 push_rcu(stack, node->bit[0], &len);
56 push_rcu(stack, node->bit[1], &len);
57- kfree(node);
58+ kmem_cache_free(node_cache, node);
59 }
60 }
61
62@@ -164,7 +171,7 @@ static int add(struct allowedips_node __
63 return -EINVAL;
64
65 if (!rcu_access_pointer(*trie)) {
66- node = kzalloc(sizeof(*node), GFP_KERNEL);
67+ node = kmem_cache_zalloc(node_cache, GFP_KERNEL);
68 if (unlikely(!node))
69 return -ENOMEM;
70 RCU_INIT_POINTER(node->peer, peer);
71@@ -180,7 +187,7 @@ static int add(struct allowedips_node __
72 return 0;
73 }
74
75- newnode = kzalloc(sizeof(*newnode), GFP_KERNEL);
76+ newnode = kmem_cache_zalloc(node_cache, GFP_KERNEL);
77 if (unlikely(!newnode))
78 return -ENOMEM;
79 RCU_INIT_POINTER(newnode->peer, peer);
80@@ -213,10 +220,10 @@ static int add(struct allowedips_node __
81 return 0;
82 }
83
84- node = kzalloc(sizeof(*node), GFP_KERNEL);
85+ node = kmem_cache_zalloc(node_cache, GFP_KERNEL);
86 if (unlikely(!node)) {
87 list_del(&newnode->peer_list);
88- kfree(newnode);
89+ kmem_cache_free(node_cache, newnode);
90 return -ENOMEM;
91 }
92 INIT_LIST_HEAD(&node->peer_list);
93@@ -306,7 +313,7 @@ void wg_allowedips_remove_by_peer(struct
94 if (child)
95 child->parent_bit = node->parent_bit;
96 *rcu_dereference_protected(node->parent_bit, lockdep_is_held(lock)) = child;
97- kfree_rcu(node, rcu);
98+ call_rcu(&node->rcu, node_free_rcu);
99
100 /* TODO: Note that we currently don't walk up and down in order to
101 * free any potential filler nodes. This means that this function
102@@ -350,4 +357,16 @@ struct wg_peer *wg_allowedips_lookup_src
103 return NULL;
104 }
105
106+int __init wg_allowedips_slab_init(void)
107+{
108+ node_cache = KMEM_CACHE(allowedips_node, 0);
109+ return node_cache ? 0 : -ENOMEM;
110+}
111+
112+void wg_allowedips_slab_uninit(void)
113+{
114+ rcu_barrier();
115+ kmem_cache_destroy(node_cache);
116+}
117+
118 #include "selftest/allowedips.c"
119--- a/drivers/net/wireguard/allowedips.h
120+++ b/drivers/net/wireguard/allowedips.h
121@@ -19,7 +19,7 @@ struct allowedips_node {
122 u8 bits[16] __aligned(__alignof(u64));
123
124 /* Keep rarely used members at bottom to be beyond cache line. */
125- struct allowedips_node *__rcu *parent_bit; /* XXX: this puts us at 68->128 bytes instead of 60->64 bytes!! */
126+ struct allowedips_node *__rcu *parent_bit;
127 union {
128 struct list_head peer_list;
129 struct rcu_head rcu;
130@@ -53,4 +53,7 @@ struct wg_peer *wg_allowedips_lookup_src
131 bool wg_allowedips_selftest(void);
132 #endif
133
134+int wg_allowedips_slab_init(void);
135+void wg_allowedips_slab_uninit(void);
136+
137 #endif /* _WG_ALLOWEDIPS_H */
138--- a/drivers/net/wireguard/main.c
139+++ b/drivers/net/wireguard/main.c
140@@ -21,10 +21,15 @@ static int __init mod_init(void)
141 {
142 int ret;
143
144+ ret = wg_allowedips_slab_init();
145+ if (ret < 0)
146+ goto err_allowedips;
147+
148 #ifdef DEBUG
149+ ret = -ENOTRECOVERABLE;
150 if (!wg_allowedips_selftest() || !wg_packet_counter_selftest() ||
151 !wg_ratelimiter_selftest())
152- return -ENOTRECOVERABLE;
153+ goto err_peer;
154 #endif
155 wg_noise_init();
156
157@@ -50,6 +55,8 @@ err_netlink:
158 err_device:
159 wg_peer_uninit();
160 err_peer:
161+ wg_allowedips_slab_uninit();
162+err_allowedips:
163 return ret;
164 }
165
166@@ -58,6 +65,7 @@ static void __exit mod_exit(void)
167 wg_genetlink_uninit();
168 wg_device_uninit();
169 wg_peer_uninit();
170+ wg_allowedips_slab_uninit();
171 }
172
173 module_init(mod_init);