blob: f48d1b6376a42d5429ab738e1563c5b588d6776c [file] [log] [blame]
Eric W. Biedermandbec2842016-07-30 13:58:49 -05001/*
2 * This program is free software; you can redistribute it and/or
3 * modify it under the terms of the GNU General Public License as
4 * published by the Free Software Foundation, version 2 of the
5 * License.
6 */
7
8#include <linux/stat.h>
9#include <linux/sysctl.h>
10#include <linux/slab.h>
Ingo Molnar5b825c32017-02-02 17:54:15 +010011#include <linux/cred.h>
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -050012#include <linux/hash.h>
Randy Dunlap514c6032018-04-05 16:25:34 -070013#include <linux/kmemleak.h>
Eric W. Biedermandbec2842016-07-30 13:58:49 -050014#include <linux/user_namespace.h>
15
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -050016#define UCOUNTS_HASHTABLE_BITS 10
17static struct hlist_head ucounts_hashtable[(1 << UCOUNTS_HASHTABLE_BITS)];
18static DEFINE_SPINLOCK(ucounts_lock);
19
20#define ucounts_hashfn(ns, uid) \
21 hash_long((unsigned long)__kuid_val(uid) + (unsigned long)(ns), \
22 UCOUNTS_HASHTABLE_BITS)
23#define ucounts_hashentry(ns, uid) \
24 (ucounts_hashtable + ucounts_hashfn(ns, uid))
25
26
Eric W. Biedermandbec2842016-07-30 13:58:49 -050027#ifdef CONFIG_SYSCTL
28static struct ctl_table_set *
29set_lookup(struct ctl_table_root *root)
30{
31 return &current_user_ns()->set;
32}
33
34static int set_is_seen(struct ctl_table_set *set)
35{
36 return &current_user_ns()->set == set;
37}
38
39static int set_permissions(struct ctl_table_header *head,
40 struct ctl_table *table)
41{
42 struct user_namespace *user_ns =
43 container_of(head->set, struct user_namespace, set);
44 int mode;
45
46 /* Allow users with CAP_SYS_RESOURCE unrestrained access */
47 if (ns_capable(user_ns, CAP_SYS_RESOURCE))
48 mode = (table->mode & S_IRWXU) >> 6;
49 else
50 /* Allow all others at most read-only access */
51 mode = table->mode & S_IROTH;
52 return (mode << 6) | (mode << 3) | mode;
53}
54
55static struct ctl_table_root set_root = {
56 .lookup = set_lookup,
57 .permissions = set_permissions,
58};
59
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -050060static int zero = 0;
61static int int_max = INT_MAX;
Nikolay Borisov1cce1ee2016-12-14 15:56:33 +020062#define UCOUNT_ENTRY(name) \
Eric W. Biederman25f9c082016-08-08 14:41:52 -050063 { \
64 .procname = name, \
65 .maxlen = sizeof(int), \
66 .mode = 0644, \
67 .proc_handler = proc_dointvec_minmax, \
68 .extra1 = &zero, \
69 .extra2 = &int_max, \
70 }
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -050071static struct ctl_table user_table[] = {
Eric W. Biederman25f9c082016-08-08 14:41:52 -050072 UCOUNT_ENTRY("max_user_namespaces"),
Eric W. Biedermanf333c702016-08-08 14:08:36 -050073 UCOUNT_ENTRY("max_pid_namespaces"),
Eric W. Biedermanf7af3d12016-08-08 14:11:25 -050074 UCOUNT_ENTRY("max_uts_namespaces"),
Eric W. Biedermanaba35662016-08-08 14:20:23 -050075 UCOUNT_ENTRY("max_ipc_namespaces"),
Eric W. Biederman70328662016-08-08 14:33:23 -050076 UCOUNT_ENTRY("max_net_namespaces"),
Eric W. Biederman537f7cc2016-08-08 14:37:37 -050077 UCOUNT_ENTRY("max_mnt_namespaces"),
Eric W. Biedermand08311d2016-08-08 14:25:30 -050078 UCOUNT_ENTRY("max_cgroup_namespaces"),
Nikolay Borisov1cce1ee2016-12-14 15:56:33 +020079#ifdef CONFIG_INOTIFY_USER
80 UCOUNT_ENTRY("max_inotify_instances"),
81 UCOUNT_ENTRY("max_inotify_watches"),
82#endif
Eric W. Biedermandbec2842016-07-30 13:58:49 -050083 { }
84};
85#endif /* CONFIG_SYSCTL */
86
87bool setup_userns_sysctls(struct user_namespace *ns)
88{
89#ifdef CONFIG_SYSCTL
90 struct ctl_table *tbl;
91 setup_sysctl_set(&ns->set, &set_root, set_is_seen);
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -050092 tbl = kmemdup(user_table, sizeof(user_table), GFP_KERNEL);
Eric W. Biedermandbec2842016-07-30 13:58:49 -050093 if (tbl) {
Eric W. Biederman25f9c082016-08-08 14:41:52 -050094 int i;
95 for (i = 0; i < UCOUNT_COUNTS; i++) {
96 tbl[i].data = &ns->ucount_max[i];
97 }
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -050098 ns->sysctls = __register_sysctl_table(&ns->set, "user", tbl);
Eric W. Biedermandbec2842016-07-30 13:58:49 -050099 }
100 if (!ns->sysctls) {
101 kfree(tbl);
102 retire_sysctl_set(&ns->set);
103 return false;
104 }
105#endif
106 return true;
107}
108
109void retire_userns_sysctls(struct user_namespace *ns)
110{
111#ifdef CONFIG_SYSCTL
112 struct ctl_table *tbl;
113
114 tbl = ns->sysctls->ctl_table_arg;
115 unregister_sysctl_table(ns->sysctls);
116 retire_sysctl_set(&ns->set);
117 kfree(tbl);
118#endif
119}
120
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500121static struct ucounts *find_ucounts(struct user_namespace *ns, kuid_t uid, struct hlist_head *hashent)
122{
123 struct ucounts *ucounts;
124
125 hlist_for_each_entry(ucounts, hashent, node) {
126 if (uid_eq(ucounts->uid, uid) && (ucounts->ns == ns))
127 return ucounts;
128 }
129 return NULL;
130}
131
132static struct ucounts *get_ucounts(struct user_namespace *ns, kuid_t uid)
133{
134 struct hlist_head *hashent = ucounts_hashentry(ns, uid);
135 struct ucounts *ucounts, *new;
136
Nikolay Borisov880a3852017-01-20 15:21:35 +0200137 spin_lock_irq(&ucounts_lock);
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500138 ucounts = find_ucounts(ns, uid, hashent);
139 if (!ucounts) {
Nikolay Borisov880a3852017-01-20 15:21:35 +0200140 spin_unlock_irq(&ucounts_lock);
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500141
142 new = kzalloc(sizeof(*new), GFP_KERNEL);
143 if (!new)
144 return NULL;
145
146 new->ns = ns;
147 new->uid = uid;
Eric W. Biederman040757f2017-03-05 15:03:22 -0600148 new->count = 0;
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500149
Nikolay Borisov880a3852017-01-20 15:21:35 +0200150 spin_lock_irq(&ucounts_lock);
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500151 ucounts = find_ucounts(ns, uid, hashent);
152 if (ucounts) {
153 kfree(new);
154 } else {
155 hlist_add_head(&new->node, hashent);
156 ucounts = new;
157 }
158 }
Eric W. Biederman040757f2017-03-05 15:03:22 -0600159 if (ucounts->count == INT_MAX)
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500160 ucounts = NULL;
Eric W. Biederman040757f2017-03-05 15:03:22 -0600161 else
162 ucounts->count += 1;
Nikolay Borisov880a3852017-01-20 15:21:35 +0200163 spin_unlock_irq(&ucounts_lock);
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500164 return ucounts;
165}
166
167static void put_ucounts(struct ucounts *ucounts)
168{
Nikolay Borisov880a3852017-01-20 15:21:35 +0200169 unsigned long flags;
170
Eric W. Biederman040757f2017-03-05 15:03:22 -0600171 spin_lock_irqsave(&ucounts_lock, flags);
172 ucounts->count -= 1;
173 if (!ucounts->count)
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500174 hlist_del_init(&ucounts->node);
Eric W. Biederman040757f2017-03-05 15:03:22 -0600175 else
176 ucounts = NULL;
177 spin_unlock_irqrestore(&ucounts_lock, flags);
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500178
Eric W. Biederman040757f2017-03-05 15:03:22 -0600179 kfree(ucounts);
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500180}
181
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500182static inline bool atomic_inc_below(atomic_t *v, int u)
183{
184 int c, old;
185 c = atomic_read(v);
186 for (;;) {
187 if (unlikely(c >= u))
188 return false;
189 old = atomic_cmpxchg(v, c, c+1);
190 if (likely(old == c))
191 return true;
192 c = old;
193 }
194}
195
Eric W. Biederman25f9c082016-08-08 14:41:52 -0500196struct ucounts *inc_ucount(struct user_namespace *ns, kuid_t uid,
197 enum ucount_type type)
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500198{
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500199 struct ucounts *ucounts, *iter, *bad;
200 struct user_namespace *tns;
201 ucounts = get_ucounts(ns, uid);
202 for (iter = ucounts; iter; iter = tns->ucounts) {
203 int max;
204 tns = iter->ns;
Eric W. Biederman25f9c082016-08-08 14:41:52 -0500205 max = READ_ONCE(tns->ucount_max[type]);
206 if (!atomic_inc_below(&iter->ucount[type], max))
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500207 goto fail;
208 }
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500209 return ucounts;
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500210fail:
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500211 bad = iter;
212 for (iter = ucounts; iter != bad; iter = iter->ns->ucounts)
Eric W. Biederman25f9c082016-08-08 14:41:52 -0500213 atomic_dec(&iter->ucount[type]);
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500214
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500215 put_ucounts(ucounts);
216 return NULL;
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500217}
218
Eric W. Biederman25f9c082016-08-08 14:41:52 -0500219void dec_ucount(struct ucounts *ucounts, enum ucount_type type)
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500220{
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500221 struct ucounts *iter;
222 for (iter = ucounts; iter; iter = iter->ns->ucounts) {
Eric W. Biederman25f9c082016-08-08 14:41:52 -0500223 int dec = atomic_dec_if_positive(&iter->ucount[type]);
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500224 WARN_ON_ONCE(dec < 0);
225 }
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500226 put_ucounts(ucounts);
Eric W. Biedermanb376c3e2016-08-08 13:41:24 -0500227}
228
Eric W. Biedermandbec2842016-07-30 13:58:49 -0500229static __init int user_namespace_sysctl_init(void)
230{
231#ifdef CONFIG_SYSCTL
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500232 static struct ctl_table_header *user_header;
Eric W. Biedermandbec2842016-07-30 13:58:49 -0500233 static struct ctl_table empty[1];
234 /*
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500235 * It is necessary to register the user directory in the
Eric W. Biedermandbec2842016-07-30 13:58:49 -0500236 * default set so that registrations in the child sets work
237 * properly.
238 */
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500239 user_header = register_sysctl("user", empty);
Luis R. Rodriguezed5bd7d2017-02-08 14:30:50 -0800240 kmemleak_ignore(user_header);
Eric W. Biedermanf6b2db1a2016-08-08 13:54:50 -0500241 BUG_ON(!user_header);
Eric W. Biedermandbec2842016-07-30 13:58:49 -0500242 BUG_ON(!setup_userns_sysctls(&init_user_ns));
243#endif
244 return 0;
245}
246subsys_initcall(user_namespace_sysctl_init);