Skip to contentSkip to conversation

capabilites: allow the application of capability limits to usermode helpers

PATCH In review Eric Paris 5 participants 8 messages Mar 9, 2011 – Mar 12, 2011

Subsystems (2)
lore
Share & tools
lore

Eric Paris

author
#0 lore raw

PATCH 1/1 capabilites: allow the application of capability limits to usermode helpers

There is no way to limit the capabilities of usermodehelpers. This problem reared its head recently when someone complained that any user with cap_net_admin was able to load arbitrary kernel modules, even though the user didn't have cap_sys_module. The reason is because the actual load is done by a usermode helper and those always have the full cap set. This patch addes new sysctls which allow us to bound the permissions of usermode helpers.

/proc/sys/kernel/usermodehelper/bset
/proc/sys/kernel/usermodehelper/inheritable

You must have CAP_SYS_MODULE to change these (changes are &= ONLY). When the kernel launches a usermodehelper it will do so with these as the bset and pI.

-v2: make globals static
create spinlock to protect globals

No-objection-from: Serge E. Hallyn <[email protected]>
---

 include/linux/kmod.h |    3 ++
 kernel/kmod.c        |   99 ++++++++++++++++++++++++++++++++++++++++++++++++++
 kernel/sysctl.c      |    6 +++
 3 files changed, 108 insertions(+), 0 deletions(-)
include/linux/kmod.h +3 -0
index 6efd7a7..79bb98d 100644
@@ -24,6 +24,7 @@
 #include <linux/errno.h>
 #include <linux/compiler.h>
 #include <linux/workqueue.h>
+#include <linux/sysctl.h>
 
 #define KMOD_PATH_LEN 256
 
@@ -109,6 +110,8 @@ call_usermodehelper(char *path, char **argv, char **envp, enum umh_wait wait)
 				       NULL, NULL, NULL);
 }
 
+extern struct ctl_table usermodehelper_table[];
+
 extern void usermodehelper_init(void);
 
 extern int usermodehelper_disable(void);
kernel/kmod.c +99 -0

129 lines not on this page · raw message

kernel/sysctl.c +6 -0
index 79268ee..f8ecc96 100644
@@ -56,6 +56,7 @@
 #include <linux/kprobes.h>
 #include <linux/pipe_fs_i.h>
 #include <linux/oom.h>
+#include <linux/kmod.h>
 
 #include <asm/uaccess.h>
 #include <asm/processor.h>
@@ -610,6 +611,11 @@ static struct ctl_table kern_table[] = {
 		.child		= random_table,
 	},
 	{
+		.procname	= "usermodehelper",
+		.mode		= 0555,
+		.child		= usermodehelper_table,
+	},
+	{
 		.procname	= "overflowuid",
 		.data		= &overflowuid,
 		.maxlen		= sizeof(int),
#1 lore raw

Eric,

Quoted 4 lines · Eric Paris

someone complained that any user with
cap_net_admin was able to load arbitrary kernel modules, even though the user didn't have cap_sys_module. The reason is because the actual load is done by a usermode helper and those always have the full cap set.

AFAIU, your patch sets system-wide caps for _all_ usermode helpers, right? Then it does nothing with cap_net_admin's problem as it should restrict caps of specific helpers spawned from specific networking code, but not touching anything related to another helpers.

Quoted 8 lines

sysctls which allow us to bound the permissions of usermode helpers.

/proc/sys/kernel/usermodehelper/bset
/proc/sys/kernel/usermodehelper/inheritable

You must have CAP_SYS_MODULE to change these (changes are &= ONLY). When the kernel launches a usermodehelper it will do so with these as the bset and pI.

Thanks,

Signature · 1 line
Vasiliy
author
#2 lore raw
Quoted 8 lines · Vasiliy Kulikov

Eric,

Quoted 4 lines · Eric Paris

someone complained that any user with
cap_net_admin was able to load arbitrary kernel modules, even though the user didn't have cap_sys_module. The reason is because the actual load is done by a usermode helper and those always have the full cap set.

Quoted 4 lines

AFAIU, your patch sets system-wide caps for _all_ usermode helpers, right? Then it does nothing with cap_net_admin's problem as it should restrict caps of specific helpers spawned from specific networking code, but not touching anything related to another helpers.

I'm actually solving 2 problems at once and it just so happens the the CAP_NET_ADMIN brew-ha-ha came up more recently. The original problem, and reason I wrote this patch, is because it's impossible on a modern Linux system to permanently and irrevocably drop capabilities system wide. You can get very close by dropping capabilities from the bset before init is launched which means everything launched by userspace can't have the dropped capabilities. But, khelper is still going to have the full set and thus usermodehelpers launched by the kernel will have the full set. This patch allows one to control usermode helpers. How one chooses to use that, is up to them.

-Eric

Acked-by on v1 1/1
#3 lore raw
In reply to Eric Paris
Quoted 14 lines · Eric Paris

There is no way to limit the capabilities of usermodehelpers. This problem reared its head recently when someone complained that any user with cap_net_admin was able to load arbitrary kernel modules, even though the user didn't have cap_sys_module. The reason is because the actual load is done by a usermode helper and those always have the full cap set. This patch addes new sysctls which allow us to bound the permissions of usermode helpers.

/proc/sys/kernel/usermodehelper/bset
/proc/sys/kernel/usermodehelper/inheritable

You must have CAP_SYS_MODULE to change these (changes are &= ONLY). When the kernel launches a usermodehelper it will do so with these as the bset and pI.

Quoted 5 lines

-v2: make globals static
create spinlock to protect globals

No-objection-from: Serge E. Hallyn <[email protected]>

#4 lore raw
In reply to Eric Paris
Quoted 9 lines · Eric Paris

There is no way to limit the capabilities of usermodehelpers. This problem reared its head recently when someone complained that any user with cap_net_admin was able to load arbitrary kernel modules, even though the user didn't have cap_sys_module. The reason is because the actual load is done by a usermode helper and those always have the full cap set. This patch addes new sysctls which allow us to bound the permissions of usermode helpers.

/proc/sys/kernel/usermodehelper/bset
/proc/sys/kernel/usermodehelper/inheritable

Shouldn't these be documented somewhere? Documentation/ABI?

Quoted 1 line

You must have CAP_SYS_MODULE to change these (changes are &= ONLY).

Why that permission? Just because 'modprobe' is usually run from this callback? Or some other reason?

Quoted 2 lines

When the kernel launches a usermodehelper it will do so with these as the bset and pI.

Shouldn't the caller of these functions be the ones dictating the capabilities it should be run with?

thanks,

greg k-h

author
#5 lore raw
Quoted 8 lines · Greg KH
Quoted 7 lines · Eric Paris

There is no way to limit the capabilities of usermodehelpers. This problem reared its head recently when someone complained that any user with cap_net_admin was able to load arbitrary kernel modules, even though the user didn't have cap_sys_module. The reason is because the actual load is done by a usermode helper and those always have the full cap set. This patch addes new sysctls which allow us to bound the permissions of usermode helpers.

Quoted 4 lines
Quoted 2 lines

/proc/sys/kernel/usermodehelper/bset
/proc/sys/kernel/usermodehelper/inheritable

Shouldn't these be documented somewhere? Documentation/ABI?

Yes, they should. Will do.

Quoted 4 lines
Quoted 1 line

You must have CAP_SYS_MODULE to change these (changes are &= ONLY).

Why that permission? Just because 'modprobe' is usually run from this callback? Or some other reason?

I don't have a good answer for this. I was originally looking at this thinking about modprobe. Since it's a purely restrictive interface I probably don't need something quite so strong. I'm trying to decide what the implications are of usemode helpers being forced to run with reduced permissions. Andrew, do you have thoughts?

Quoted 5 lines
Quoted 2 lines

When the kernel launches a usermodehelper it will do so with these as the bset and pI.

Shouldn't the caller of these functions be the ones dictating the capabilities it should be run with?

Yes. And no. It depends what you mean. The caps of the 'caller' task are irrelevant. If I run ifconfig ipv6 I need CAP_NET_ADMIN but the upcall needs CAP_SYS_MODULE. If I plug in a USB drive there is no 'caller' task which makes sense.

Now if by 'caller' you mean 'call site' then yes, we could probably launch usermodehelpers with reduced privileged sets. We know in the code when we are asking to launch modprobe that we are going to need CAP_SYS_MODULE and don't need caps like CAP_SYS_RAWIO and CAP_MAC_ADMIN. We know when we upcall to hotplug we don't really need any priv, since it's another task that is going to do the real work. So yeah, there might be some value in another patch to address this....

But neither solves the problem of being able to eliminate capabilities from a machine globally. In olden times we had a global cap-bound but it was dropped in favor of an inheritance from init type mechanism. Since kthreads don't inherit from init we still end up with this patch.

#6 lore raw
Quoted 12 lines · Eric Paris
Quoted 5 lines · Greg KH
Quoted 2 lines

When the kernel launches a usermodehelper it will do so with these as the bset and pI.

Shouldn't the caller of these functions be the ones dictating the capabilities it should be run with?

Yes. And no. It depends what you mean. The caps of the 'caller' task are irrelevant. If I run ifconfig ipv6 I need CAP_NET_ADMIN but the upcall needs CAP_SYS_MODULE. If I plug in a USB drive there is no 'caller' task which makes sense.

Quoted 7 lines

Now if by 'caller' you mean 'call site' then yes, we could probably launch usermodehelpers with reduced privileged sets. We know in the code when we are asking to launch modprobe that we are going to need CAP_SYS_MODULE and don't need caps like CAP_SYS_RAWIO and CAP_MAC_ADMIN. We know when we upcall to hotplug we don't really need any priv, since it's another task that is going to do the real work. So yeah, there might be some value in another patch to address this....

Yes, that is what I was referring to.

Quoted 4 lines

But neither solves the problem of being able to eliminate capabilities from a machine globally. In olden times we had a global cap-bound but it was dropped in favor of an inheritance from init type mechanism. Since kthreads don't inherit from init we still end up with this patch.

I'm not objecting to the patch, or the idea, just want to make sure it is correct.

thanks,

greg k-h

#7 lore raw

I'd prefer it if the invoker had both CAP_SETPCAP and CAP_SYS_MODULE in its pE set.

Also, have you considered an interface for setting the default values that is something like a "clone mine"? That is, cache the invoker's cap_inheritable, and cap_bound? (But not cap_permitted and cap_effective!) This would take care of atomicity and also allow for potentially including more state when folk find its logically necessary as the kernel evolves.

Cheers

Andrew

Quoted 35 lines · Greg KH
Quoted 12 lines · Eric Paris
Quoted 5 lines · Greg KH
Quoted 2 lines

When the kernel launches a usermodehelper it will do so with these as the bset and pI.

Shouldn't the caller of these functions be the ones dictating the capabilities it should be run with?

Yes. And no.  It depends what you mean.  The caps of the 'caller' task are irrelevant.  If I run ifconfig ipv6 I need CAP_NET_ADMIN but the upcall needs CAP_SYS_MODULE.  If I plug in a USB drive there is no 'caller' task which makes sense.

Quoted 7 lines

Now if by 'caller' you mean 'call site' then yes, we could probably launch usermodehelpers with reduced privileged sets.  We know in the code when we are asking to launch modprobe that we are going to need CAP_SYS_MODULE and don't need caps like CAP_SYS_RAWIO and CAP_MAC_ADMIN. We know when we upcall to hotplug we don't really need any priv, since it's another task that is going to do the real work.  So yeah, there might be some value in another patch to address this....

Yes, that is what I was referring to.

Quoted 4 lines

But neither solves the problem of being able to eliminate capabilities from a machine globally.  In olden times we had a global cap-bound but it was dropped in favor of an inheritance from init type mechanism. Since kthreads don't inherit from init we still end up with this patch.

I'm not objecting to the patch, or the idea, just want to make sure it is correct.

thanks,

greg k-h