In environments that require the kexec kernel image to be signed, prevent
using the kexec_load syscall. In order for LSMs and IMA to differentiate
between kexec_load and kexec_file_load syscalls, this patch set adds a
call to security_kernel_read_file() in kexec_load_check().
Signed-off-by: Mimi Zohar <[email protected]>
Mimi Zohar (3):
ima: based on the "secure_boot" policy limit syscalls
kexec: call LSM hook for kexec_load syscall
ima: based on policy require signed kexec kernel images
kernel/kexec.c | 11 +++++++++++
security/integrity/ima/ima.h | 1 +
security/integrity/ima/ima_main.c | 9 +++++++++
security/integrity/ima/ima_policy.c | 27 ++++++++++++++++++++-------
4 files changed, 41 insertions(+), 7 deletions(-)
--
2.7.5
Allow LSMs and IMA to differentiate between the kexec_load and
kexec_file_load syscalls by adding an "unnecessary" call to
security_kernel_read_file() in kexec_load. This would be similar to the
existing init_module syscall calling security_kernel_read_file().
Signed-off-by: Mimi Zohar <[email protected]>
---
kernel/kexec.c | 11 +++++++++++
1 file changed, 11 insertions(+)
diff --git a/kernel/kexec.c b/kernel/kexec.c
index aed8fb2564b3..d1386cfc6796 100644
--- a/kernel/kexec.c
+++ b/kernel/kexec.c
@@ -11,6 +11,7 @@
#include <linux/capability.h>
#include <linux/mm.h>
#include <linux/file.h>
+#include <linux/security.h>
#include <linux/kexec.h>
#include <linux/mutex.h>
#include <linux/list.h>
@@ -195,11 +196,21 @@ static int do_kexec_load(unsigned long entry, unsigned long nr_segments,
static inline int kexec_load_check(unsigned long nr_segments,
unsigned long flags)
{
+ int result;
+
/* We only trust the superuser with rebooting the system. */
if (!capable(CAP_SYS_BOOT) || kexec_load_disabled)
return -EPERM;
/*
+ * Allow LSMs and IMA to differentiate between kexec_load and
+ * kexec_file_load syscalls.
+ */
+ result = security_kernel_read_file(NULL, READING_KEXEC_IMAGE);
+ if (result < 0)
+ return result;
+
+ /*
* Verify we have a legal set of flags
* This leaves us room for future extensions.
*/
--
2.7.5
The builtin "secure_boot" policy adds IMA appraisal rules requiring kernel
modules (finit_module syscall), direct firmware load, kexec kernel image
(kexec_file_load syscall), and the IMA policy to be signed, but did not
prevent the other syscalls/methods from working. Loading an equivalent
custom policy containing these same rules would have prevented the other
syscalls/methods from working.
This patch refactors the code to load custom policies, defining a new
function named ima_appraise_flag(). The new function is called either
when loading the builtin "secure_boot" or custom policies.
Fixes: 503ceaef8e2e ("ima: define a set of appraisal rules requiring file signatures")
Signed-off-by: Mimi Zohar <[email protected]>
---
security/integrity/ima/ima_policy.c | 25 ++++++++++++++++++-------
1 file changed, 18 insertions(+), 7 deletions(-)
diff --git a/security/integrity/ima/ima_policy.c b/security/integrity/ima/ima_policy.c
index d89bebf85421..1bdb5bc57568 100644
--- a/security/integrity/ima/ima_policy.c
+++ b/security/integrity/ima/ima_policy.c
@@ -435,6 +435,17 @@ void ima_update_policy_flag(void)
ima_policy_flag &= ~IMA_APPRAISE;
}
+static int ima_appraise_flag(enum ima_hooks func)
+{
+ if (func == MODULE_CHECK)
+ return IMA_APPRAISE_MODULES;
+ else if (func == FIRMWARE_CHECK)
+ return IMA_APPRAISE_FIRMWARE;
+ else if (func == POLICY_CHECK)
+ return IMA_APPRAISE_POLICY;
+ return 0;
+}
+
/**
* ima_init_policy - initialize the default measure rules.
*
@@ -473,9 +484,12 @@ void __init ima_init_policy(void)
* Insert the appraise rules requiring file signatures, prior to
* any other appraise rules.
*/
- for (i = 0; i < secure_boot_entries; i++)
+ for (i = 0; i < secure_boot_entries; i++) {
list_add_tail(&secure_boot_rules[i].list,
&ima_default_rules);
+ temp_ima_appraise |=
+ ima_appraise_flag(secure_boot_rules[i].func);
+ }
for (i = 0; i < appraise_entries; i++) {
list_add_tail(&default_appraise_rules[i].list,
@@ -917,12 +931,9 @@ static int ima_parse_rule(char *rule, struct ima_rule_entry *entry)
}
if (!result && (entry->action == UNKNOWN))
result = -EINVAL;
- else if (entry->func == MODULE_CHECK)
- temp_ima_appraise |= IMA_APPRAISE_MODULES;
- else if (entry->func == FIRMWARE_CHECK)
- temp_ima_appraise |= IMA_APPRAISE_FIRMWARE;
- else if (entry->func == POLICY_CHECK)
- temp_ima_appraise |= IMA_APPRAISE_POLICY;
+ else if (entry->action == APPRAISE)
+ temp_ima_appraise |= ima_appraise_flag(entry->func);
+
audit_log_format(ab, "res=%d", !result);
audit_log_end(ab);
return result;
--
2.7.5
The original kexec_load syscall can not verify file signatures. This
patch differentiates between the kexec_load and kexec_file_load
syscalls.
Signed-off-by: Mimi Zohar <[email protected]>
---
security/integrity/ima/ima.h | 1 +
security/integrity/ima/ima_main.c | 9 +++++++++
security/integrity/ima/ima_policy.c | 2 ++
3 files changed, 12 insertions(+)
diff --git a/security/integrity/ima/ima.h b/security/integrity/ima/ima.h
index 35fe91aa1fc9..03c9c37ee345 100644
--- a/security/integrity/ima/ima.h
+++ b/security/integrity/ima/ima.h
@@ -233,6 +233,7 @@ int ima_policy_show(struct seq_file *m, void *v);
#define IMA_APPRAISE_MODULES 0x08
#define IMA_APPRAISE_FIRMWARE 0x10
#define IMA_APPRAISE_POLICY 0x20
+#define IMA_APPRAISE_KEXEC 0x40
#ifdef CONFIG_IMA_APPRAISE
int ima_appraise_measurement(enum ima_hooks func,
diff --git a/security/integrity/ima/ima_main.c b/security/integrity/ima/ima_main.c
index 74d0bd7e76d7..754ece08e1c6 100644
--- a/security/integrity/ima/ima_main.c
+++ b/security/integrity/ima/ima_main.c
@@ -444,6 +444,15 @@ int ima_read_file(struct file *file, enum kernel_read_file_id read_id)
}
return 0; /* We rely on module signature checking */
}
+
+ if (!file && read_id == READING_KEXEC_IMAGE) {
+ if ((ima_appraise & IMA_APPRAISE_KEXEC) &&
+ (ima_appraise & IMA_APPRAISE_ENFORCE)) {
+ pr_err("impossible to appraise a kernel image without a file descriptor; try using kexec_file syscall.\n");
+ return -EACCES; /* INTEGRITY_UNKNOWN */
+ }
+ return 0;
+ }
return 0;
}
diff --git a/security/integrity/ima/ima_policy.c b/security/integrity/ima/ima_policy.c
index 1bdb5bc57568..3444352caf80 100644
--- a/security/integrity/ima/ima_policy.c
+++ b/security/integrity/ima/ima_policy.c
@@ -443,6 +443,8 @@ static int ima_appraise_flag(enum ima_hooks func)
return IMA_APPRAISE_FIRMWARE;
else if (func == POLICY_CHECK)
return IMA_APPRAISE_POLICY;
+ else if (func == KEXEC_KERNEL_CHECK)
+ return IMA_APPRAISE_KEXEC;
return 0;
}
--
2.7.5
Mimi Zohar <[email protected]> writes:
> Allow LSMs and IMA to differentiate between the kexec_load and
> kexec_file_load syscalls by adding an "unnecessary" call to
> security_kernel_read_file() in kexec_load. This would be similar to the
> existing init_module syscall calling security_kernel_read_file().
Given the reasonable desire to load a policy that ensures everything
has a signature I don't have fundamental objections.
security_kernel_read_file as a hook seems an odd choice. At the very
least it has a bad name because there is no file reading going on here.
I am concerned that I don't see CONFIG_KEXEC_VERIFY_SIG being tested
anywhere. Which means I could have a kernel compiled without that and I
would be allowed to use kexec_file_load without signature checking.
While kexec_load would be denied.
Am I missing something here?
Eric
> Signed-off-by: Mimi Zohar <[email protected]>
> ---
> kernel/kexec.c | 11 +++++++++++
> 1 file changed, 11 insertions(+)
>
> diff --git a/kernel/kexec.c b/kernel/kexec.c
> index aed8fb2564b3..d1386cfc6796 100644
> --- a/kernel/kexec.c
> +++ b/kernel/kexec.c
> @@ -11,6 +11,7 @@
> #include <linux/capability.h>
> #include <linux/mm.h>
> #include <linux/file.h>
> +#include <linux/security.h>
> #include <linux/kexec.h>
> #include <linux/mutex.h>
> #include <linux/list.h>
> @@ -195,11 +196,21 @@ static int do_kexec_load(unsigned long entry, unsigned long nr_segments,
> static inline int kexec_load_check(unsigned long nr_segments,
> unsigned long flags)
> {
> + int result;
> +
> /* We only trust the superuser with rebooting the system. */
> if (!capable(CAP_SYS_BOOT) || kexec_load_disabled)
> return -EPERM;
>
> /*
> + * Allow LSMs and IMA to differentiate between kexec_load and
> + * kexec_file_load syscalls.
> + */
> + result = security_kernel_read_file(NULL, READING_KEXEC_IMAGE);
> + if (result < 0)
> + return result;
> +
> + /*
> * Verify we have a legal set of flags
> * This leaves us room for future extensions.
> */
On Wed, 2018-05-02 at 09:45 -0500, Eric W. Biederman wrote:
> Mimi Zohar <[email protected]> writes:
>
> > Allow LSMs and IMA to differentiate between the kexec_load and
> > kexec_file_load syscalls by adding an "unnecessary" call to
> > security_kernel_read_file() in kexec_load. This would be similar to the
> > existing init_module syscall calling security_kernel_read_file().
>
> Given the reasonable desire to load a policy that ensures everything
> has a signature I don't have fundamental objections.
>
> security_kernel_read_file as a hook seems an odd choice. At the very
> least it has a bad name because there is no file reading going on here.
>
> I am concerned that I don't see CONFIG_KEXEC_VERIFY_SIG being tested
> anywhere. Which means I could have a kernel compiled without that and I
> would be allowed to use kexec_file_load without signature checking.
> While kexec_load would be denied.
>
> Am I missing something here?
The kexec_file_load() calls kernel_read_file_from_fd(), which in turn
calls security_kernel_read_file(). So kexec_file_load and kexec_load
syscall would be using the same method for enforcing signature
verification.
This is independent of the architecture specific method for verifying
signatures. The coordination between these two methods was included
in the lockdown patch set, but is being removed, as well the gating of
kexec_load syscall. Instead of being based on the lockdown flag, I
assume the coordination between the two methods will reappear based on
a secure boot flag of some sort.
Mimi
>
> > Signed-off-by: Mimi Zohar <[email protected]>
> > ---
> > kernel/kexec.c | 11 +++++++++++
> > 1 file changed, 11 insertions(+)
> >
> > diff --git a/kernel/kexec.c b/kernel/kexec.c
> > index aed8fb2564b3..d1386cfc6796 100644
> > --- a/kernel/kexec.c
> > +++ b/kernel/kexec.c
> > @@ -11,6 +11,7 @@
> > #include <linux/capability.h>
> > #include <linux/mm.h>
> > #include <linux/file.h>
> > +#include <linux/security.h>
> > #include <linux/kexec.h>
> > #include <linux/mutex.h>
> > #include <linux/list.h>
> > @@ -195,11 +196,21 @@ static int do_kexec_load(unsigned long entry, unsigned long nr_segments,
> > static inline int kexec_load_check(unsigned long nr_segments,
> > unsigned long flags)
> > {
> > + int result;
> > +
> > /* We only trust the superuser with rebooting the system. */
> > if (!capable(CAP_SYS_BOOT) || kexec_load_disabled)
> > return -EPERM;
> >
> > /*
> > + * Allow LSMs and IMA to differentiate between kexec_load and
> > + * kexec_file_load syscalls.
> > + */
> > + result = security_kernel_read_file(NULL, READING_KEXEC_IMAGE);
> > + if (result < 0)
> > + return result;
> > +
> > + /*
> > * Verify we have a legal set of flags
> > * This leaves us room for future extensions.
> > */
>
Mimi Zohar <[email protected]> writes:
> On Wed, 2018-05-02 at 09:45 -0500, Eric W. Biederman wrote:
>> Mimi Zohar <[email protected]> writes:
>>
>> > Allow LSMs and IMA to differentiate between the kexec_load and
>> > kexec_file_load syscalls by adding an "unnecessary" call to
>> > security_kernel_read_file() in kexec_load. This would be similar to the
>> > existing init_module syscall calling security_kernel_read_file().
>>
>> Given the reasonable desire to load a policy that ensures everything
>> has a signature I don't have fundamental objections.
>>
>> security_kernel_read_file as a hook seems an odd choice. At the very
>> least it has a bad name because there is no file reading going on here.
>>
>> I am concerned that I don't see CONFIG_KEXEC_VERIFY_SIG being tested
>> anywhere. Which means I could have a kernel compiled without that and I
>> would be allowed to use kexec_file_load without signature checking.
>> While kexec_load would be denied.
>>
>> Am I missing something here?
>
> The kexec_file_load() calls kernel_read_file_from_fd(), which in turn
> calls security_kernel_read_file(). So kexec_file_load and kexec_load
> syscall would be using the same method for enforcing signature
> verification.
Having looked at your patches and the kernel a little more I think
this should be a separate security hook that does not take a file
parameter.
Right now every other security module assumes !file is init_module.
So I think this change has the potential to confuse other security
modules, with the result of unintended policy being applied.
So just for good security module hygeine I think this needs a dedicated
kexec_load security hook.
> This is independent of the architecture specific method for verifying
> signatures. The coordination between these two methods was included
> in the lockdown patch set, but is being removed, as well the gating of
> kexec_load syscall. Instead of being based on the lockdown flag, I
> assume the coordination between the two methods will reappear based on
> a secure boot flag of some sort.
I was blind there for a moment. Yes this is all about the ima xattrs
allowing a file to be loaded.
Eric
On 5/3/2018 8:51 AM, Eric W. Biederman wrote:
> Mimi Zohar <[email protected]> writes:
>
>> On Wed, 2018-05-02 at 09:45 -0500, Eric W. Biederman wrote:
>>> Mimi Zohar <[email protected]> writes:
>>>
>>>> Allow LSMs and IMA to differentiate between the kexec_load and
>>>> kexec_file_load syscalls by adding an "unnecessary" call to
>>>> security_kernel_read_file() in kexec_load. This would be similar to the
>>>> existing init_module syscall calling security_kernel_read_file().
>>> Given the reasonable desire to load a policy that ensures everything
>>> has a signature I don't have fundamental objections.
>>>
>>> security_kernel_read_file as a hook seems an odd choice. At the very
>>> least it has a bad name because there is no file reading going on here.
>>>
>>> I am concerned that I don't see CONFIG_KEXEC_VERIFY_SIG being tested
>>> anywhere. Which means I could have a kernel compiled without that and I
>>> would be allowed to use kexec_file_load without signature checking.
>>> While kexec_load would be denied.
>>>
>>> Am I missing something here?
>> The kexec_file_load() calls kernel_read_file_from_fd(), which in turn
>> calls security_kernel_read_file(). So kexec_file_load and kexec_load
>> syscall would be using the same method for enforcing signature
>> verification.
> Having looked at your patches and the kernel a little more I think
> this should be a separate security hook that does not take a file
> parameter.
>
> Right now every other security module assumes !file is init_module.
> So I think this change has the potential to confuse other security
> modules, with the result of unintended policy being applied.
>
> So just for good security module hygeine I think this needs a dedicated
> kexec_load security hook.
I would rather see the existing modules updated than a new
hook added. Too many hooks spoil the broth. Two hooks with
trivial differences just add to the clutter and make it harder
for non-lsm developers to figure out what to use in their
code.
Casey Schaufler <[email protected]> writes:
> On 5/3/2018 8:51 AM, Eric W. Biederman wrote:
>> Mimi Zohar <[email protected]> writes:
>>
>>> On Wed, 2018-05-02 at 09:45 -0500, Eric W. Biederman wrote:
>>>> Mimi Zohar <[email protected]> writes:
>>>>
>>>>> Allow LSMs and IMA to differentiate between the kexec_load and
>>>>> kexec_file_load syscalls by adding an "unnecessary" call to
>>>>> security_kernel_read_file() in kexec_load. This would be similar to the
>>>>> existing init_module syscall calling security_kernel_read_file().
>>>> Given the reasonable desire to load a policy that ensures everything
>>>> has a signature I don't have fundamental objections.
>>>>
>>>> security_kernel_read_file as a hook seems an odd choice. At the very
>>>> least it has a bad name because there is no file reading going on here.
>>>>
>>>> I am concerned that I don't see CONFIG_KEXEC_VERIFY_SIG being tested
>>>> anywhere. Which means I could have a kernel compiled without that and I
>>>> would be allowed to use kexec_file_load without signature checking.
>>>> While kexec_load would be denied.
>>>>
>>>> Am I missing something here?
>>> The kexec_file_load() calls kernel_read_file_from_fd(), which in turn
>>> calls security_kernel_read_file(). So kexec_file_load and kexec_load
>>> syscall would be using the same method for enforcing signature
>>> verification.
>> Having looked at your patches and the kernel a little more I think
>> this should be a separate security hook that does not take a file
>> parameter.
>>
>> Right now every other security module assumes !file is init_module.
>> So I think this change has the potential to confuse other security
>> modules, with the result of unintended policy being applied.
>>
>> So just for good security module hygeine I think this needs a dedicated
>> kexec_load security hook.
>
> I would rather see the existing modules updated than a new
> hook added. Too many hooks spoil the broth. Two hooks with
> trivial differences just add to the clutter and make it harder
> for non-lsm developers to figure out what to use in their
> code.
These are not non-trivial differences. There is absolutely nothing
file related about kexec_load. Nor for init_module for that matter.
If something is called security_kernel_read_file I think it is wholly
appropriate for code that processes such a hook to assume file is
non-NULL.
When you have to dance a jig (which is what I see the security modules
doing) to figure out who is calling a lsm hook for what purpose I think
it is a maintenance problem waiting to happen and that the hook is badly
designed.
At this point I don't care what the lsm's do with the hooks but the
hooks need to make sense for people outside of the lsm's and something
about reading a file in a syscall that doesn't read files is complete
and utter nonsense.
Eric
Mimi Zohar <[email protected]> writes:
> In environments that require the kexec kernel image to be signed, prevent
> using the kexec_load syscall. In order for LSMs and IMA to differentiate
> between kexec_load and kexec_file_load syscalls, this patch set adds a
> call to security_kernel_read_file() in kexec_load_check().
Having thought about it some more this justification for these changes
does not work. The functionality of kexec_load is already root-only.
So in environments that require the kernel image to be signed just don't
use kexec_load. Possibly even compile kexec_load out to save space
because you will never need it. You don't need a new security hook to
do any of that. Userspace is a very fine mechanism for being the
instrument of policy.
If you don't trust userspace that needs to be spelled out very clearly.
You need to talk about what your threat models are.
If the only justification is so that that we can't boot windows if
someone hacks into userspace it has my nack because that is another kind
of complete non-sense.
Given that you are not trusting userspace this changeset also probably
needs to have the kernel-hardening list cc'd. Because the only possible
justification I can imagine for something like this is kernel-hardening.
Eric
On Thu, May 3, 2018 at 1:13 PM Eric W. Biederman <[email protected]>
wrote:
> Mimi Zohar <[email protected]> writes:
> > In environments that require the kexec kernel image to be signed,
prevent
> > using the kexec_load syscall. In order for LSMs and IMA to
differentiate
> > between kexec_load and kexec_file_load syscalls, this patch set adds a
> > call to security_kernel_read_file() in kexec_load_check().
> Having thought about it some more this justification for these changes
> does not work. The functionality of kexec_load is already root-only.
> So in environments that require the kernel image to be signed just don't
> use kexec_load. Possibly even compile kexec_load out to save space
> because you will never need it. You don't need a new security hook to
> do any of that. Userspace is a very fine mechanism for being the
> instrument of policy.
> If you don't trust userspace that needs to be spelled out very clearly.
> You need to talk about what your threat models are.
kexec_load gives root arbitrary power to modify the running kernel image,
including the ability to disable enforcement of module signatures. Given
that it weakens other security mechanisms that are designed to prevent root
from disabling them, it makes sense to allow the imposition of an
equivalent restriction.
On Thu, 2018-05-03 at 11:42 -0500, Eric W. Biederman wrote:
> Casey Schaufler <[email protected]> writes:
>
> > On 5/3/2018 8:51 AM, Eric W. Biederman wrote:
> >> Mimi Zohar <[email protected]> writes:
> >>
> >>> On Wed, 2018-05-02 at 09:45 -0500, Eric W. Biederman wrote:
> >>>> Mimi Zohar <[email protected]> writes:
> >>>>
> >>>>> Allow LSMs and IMA to differentiate between the kexec_load and
> >>>>> kexec_file_load syscalls by adding an "unnecessary" call to
> >>>>> security_kernel_read_file() in kexec_load. This would be similar to the
> >>>>> existing init_module syscall calling security_kernel_read_file().
> >>>> Given the reasonable desire to load a policy that ensures everything
> >>>> has a signature I don't have fundamental objections.
> >>>>
> >>>> security_kernel_read_file as a hook seems an odd choice. At the very
> >>>> least it has a bad name because there is no file reading going on here.
> >>>>
> >>>> I am concerned that I don't see CONFIG_KEXEC_VERIFY_SIG being tested
> >>>> anywhere. Which means I could have a kernel compiled without that and I
> >>>> would be allowed to use kexec_file_load without signature checking.
> >>>> While kexec_load would be denied.
> >>>>
> >>>> Am I missing something here?
> >>> The kexec_file_load() calls kernel_read_file_from_fd(), which in turn
> >>> calls security_kernel_read_file(). So kexec_file_load and kexec_load
> >>> syscall would be using the same method for enforcing signature
> >>> verification.
> >> Having looked at your patches and the kernel a little more I think
> >> this should be a separate security hook that does not take a file
> >> parameter.
> >>
> >> Right now every other security module assumes !file is init_module.
> >> So I think this change has the potential to confuse other security
> >> modules, with the result of unintended policy being applied.
> >>
> >> So just for good security module hygeine I think this needs a dedicated
> >> kexec_load security hook.
> >
> > I would rather see the existing modules updated than a new
> > hook added. Too many hooks spoil the broth. Two hooks with
> > trivial differences just add to the clutter and make it harder
> > for non-lsm developers to figure out what to use in their
> > code.
>
> These are not non-trivial differences. There is absolutely nothing
> file related about kexec_load. Nor for init_module for that matter.
>
> If something is called security_kernel_read_file I think it is wholly
> appropriate for code that processes such a hook to assume file is
> non-NULL.
>
> When you have to dance a jig (which is what I see the security modules
> doing) to figure out who is calling a lsm hook for what purpose I think
> it is a maintenance problem waiting to happen and that the hook is badly
> designed.
>
> At this point I don't care what the lsm's do with the hooks but the
> hooks need to make sense for people outside of the lsm's and something
> about reading a file in a syscall that doesn't read files is complete
> and utter nonsense.
Sure, we can define a wrapper around the security_kernel_read_file()
hook, calling it security_non-fd_syscall() or even
security_old_syscall().
Mimi
[Cc'ing Kees and kernel-hardening]
On Thu, 2018-05-03 at 15:13 -0500, Eric W. Biederman wrote:
> Mimi Zohar <[email protected]> writes:
>
> > In environments that require the kexec kernel image to be signed, prevent
> > using the kexec_load syscall. In order for LSMs and IMA to differentiate
> > between kexec_load and kexec_file_load syscalls, this patch set adds a
> > call to security_kernel_read_file() in kexec_load_check().
>
> Having thought about it some more this justification for these changes
> does not work. The functionality of kexec_load is already root-only.
> So in environments that require the kernel image to be signed just don't
> use kexec_load. Possibly even compile kexec_load out to save space
> because you will never need it. You don't need a new security hook to
> do any of that. Userspace is a very fine mechanism for being the
> instrument of policy.
True, for those building their own kernel, they can disable the old
syscalls. The concern is not for those building their own kernels,
but for those using stock kernels.
By adding an LSM hook here in the kexec_load syscall, as opposed to an
IMA specific hook, other LSMs can piggy back on top of it. Currently,
both load_pin and SELinux are gating the kernel module syscalls based
on security_kernel_read_file.
If there was a similar option for the kernel image, I'm pretty sure
other LSMs would use it.
From an IMA perspective, there needs to be some method for only
allowing signed code to be loaded, executed, etc. - kernel modules,
kernel image/initramfs, firmware, policies.
> If you don't trust userspace that needs to be spelled out very clearly.
> You need to talk about what your threat models are.
>
> If the only justification is so that that we can't boot windows if
> someone hacks into userspace it has my nack because that is another kind
> of complete non-sense.
The usecase is the ability to gate the kexec_load usage in stock
kernels.
>
> Given that you are not trusting userspace this changeset also probably
> needs to have the kernel-hardening list cc'd. Because the only possible
> justification I can imagine for something like this is kernel-hardening.
Sure, Cc'ing linux-hardening and Kees.
Mimi
Mimi Zohar <[email protected]> writes:
> On Thu, 2018-05-03 at 11:42 -0500, Eric W. Biederman wrote:
>> Casey Schaufler <[email protected]> writes:
>>
>> > On 5/3/2018 8:51 AM, Eric W. Biederman wrote:
>> >> Mimi Zohar <[email protected]> writes:
>> >>
>> >>> On Wed, 2018-05-02 at 09:45 -0500, Eric W. Biederman wrote:
>> >>>> Mimi Zohar <[email protected]> writes:
>> >>>>
>> >>>>> Allow LSMs and IMA to differentiate between the kexec_load and
>> >>>>> kexec_file_load syscalls by adding an "unnecessary" call to
>> >>>>> security_kernel_read_file() in kexec_load. This would be similar to the
>> >>>>> existing init_module syscall calling security_kernel_read_file().
>> >>>> Given the reasonable desire to load a policy that ensures everything
>> >>>> has a signature I don't have fundamental objections.
>> >>>>
>> >>>> security_kernel_read_file as a hook seems an odd choice. At the very
>> >>>> least it has a bad name because there is no file reading going on here.
>> >>>>
>> >>>> I am concerned that I don't see CONFIG_KEXEC_VERIFY_SIG being tested
>> >>>> anywhere. Which means I could have a kernel compiled without that and I
>> >>>> would be allowed to use kexec_file_load without signature checking.
>> >>>> While kexec_load would be denied.
>> >>>>
>> >>>> Am I missing something here?
>> >>> The kexec_file_load() calls kernel_read_file_from_fd(), which in turn
>> >>> calls security_kernel_read_file(). So kexec_file_load and kexec_load
>> >>> syscall would be using the same method for enforcing signature
>> >>> verification.
>> >> Having looked at your patches and the kernel a little more I think
>> >> this should be a separate security hook that does not take a file
>> >> parameter.
>> >>
>> >> Right now every other security module assumes !file is init_module.
>> >> So I think this change has the potential to confuse other security
>> >> modules, with the result of unintended policy being applied.
>> >>
>> >> So just for good security module hygeine I think this needs a dedicated
>> >> kexec_load security hook.
>> >
>> > I would rather see the existing modules updated than a new
>> > hook added. Too many hooks spoil the broth. Two hooks with
>> > trivial differences just add to the clutter and make it harder
>> > for non-lsm developers to figure out what to use in their
>> > code.
>>
>> These are not non-trivial differences. There is absolutely nothing
>> file related about kexec_load. Nor for init_module for that matter.
>>
>> If something is called security_kernel_read_file I think it is wholly
>> appropriate for code that processes such a hook to assume file is
>> non-NULL.
>>
>> When you have to dance a jig (which is what I see the security modules
>> doing) to figure out who is calling a lsm hook for what purpose I think
>> it is a maintenance problem waiting to happen and that the hook is badly
>> designed.
>>
>> At this point I don't care what the lsm's do with the hooks but the
>> hooks need to make sense for people outside of the lsm's and something
>> about reading a file in a syscall that doesn't read files is complete
>> and utter nonsense.
>
> Sure, we can define a wrapper around the security_kernel_read_file()
> hook, calling it security_non-fd_syscall() or even
> security_old_syscall().
I really don't see why you want to use the same hook.
I just read through the code of all three users. None of them.
Especially IMA shares any significant code between the !file case and
the file case.
Eric
Mimi Zohar <[email protected]> writes:
> [Cc'ing Kees and kernel-hardening]
>
> On Thu, 2018-05-03 at 15:13 -0500, Eric W. Biederman wrote:
>> Mimi Zohar <[email protected]> writes:
>>
>> > In environments that require the kexec kernel image to be signed, prevent
>> > using the kexec_load syscall. In order for LSMs and IMA to differentiate
>> > between kexec_load and kexec_file_load syscalls, this patch set adds a
>> > call to security_kernel_read_file() in kexec_load_check().
>>
>> Having thought about it some more this justification for these changes
>> does not work. The functionality of kexec_load is already root-only.
>> So in environments that require the kernel image to be signed just don't
>> use kexec_load. Possibly even compile kexec_load out to save space
>> because you will never need it. You don't need a new security hook to
>> do any of that. Userspace is a very fine mechanism for being the
>> instrument of policy.
>
> True, for those building their own kernel, they can disable the old
> syscalls. The concern is not for those building their own kernels,
> but for those using stock kernels.
>
> By adding an LSM hook here in the kexec_load syscall, as opposed to an
> IMA specific hook, other LSMs can piggy back on top of it. Currently,
> both load_pin and SELinux are gating the kernel module syscalls based
> on security_kernel_read_file.
>
> If there was a similar option for the kernel image, I'm pretty sure
> other LSMs would use it.
>
> From an IMA perspective, there needs to be some method for only
> allowing signed code to be loaded, executed, etc. - kernel modules,
> kernel image/initramfs, firmware, policies.
What is the IMA perspective. Why can't IMA trust appropriately
authorized userspace?
>> If you don't trust userspace that needs to be spelled out very clearly.
>> You need to talk about what your threat models are.
>>
>> If the only justification is so that that we can't boot windows if
>> someone hacks into userspace it has my nack because that is another kind
>> of complete non-sense.
>
> The usecase is the ability to gate the kexec_load usage in stock
> kernels.
But kexec_load is already gated. It requires CAP_SYS_BOOT.
>> Given that you are not trusting userspace this changeset also probably
>> needs to have the kernel-hardening list cc'd. Because the only possible
>> justification I can imagine for something like this is kernel-hardening.
>
> Sure, Cc'ing linux-hardening and Kees.
>
> Mimi
On Thu, 2018-05-03 at 16:38 -0500, Eric W. Biederman wrote:
> Mimi Zohar <[email protected]> writes:
>
> > [Cc'ing Kees and kernel-hardening]
> >
> > On Thu, 2018-05-03 at 15:13 -0500, Eric W. Biederman wrote:
> >> Mimi Zohar <[email protected]> writes:
> >>
> >> > In environments that require the kexec kernel image to be signed, prevent
> >> > using the kexec_load syscall. In order for LSMs and IMA to differentiate
> >> > between kexec_load and kexec_file_load syscalls, this patch set adds a
> >> > call to security_kernel_read_file() in kexec_load_check().
> >>
> >> Having thought about it some more this justification for these changes
> >> does not work. The functionality of kexec_load is already root-only.
> >> So in environments that require the kernel image to be signed just don't
> >> use kexec_load. Possibly even compile kexec_load out to save space
> >> because you will never need it. You don't need a new security hook to
> >> do any of that. Userspace is a very fine mechanism for being the
> >> instrument of policy.
> >
> > True, for those building their own kernel, they can disable the old
> > syscalls. The concern is not for those building their own kernels,
> > but for those using stock kernels.
> >
> > By adding an LSM hook here in the kexec_load syscall, as opposed to an
> > IMA specific hook, other LSMs can piggy back on top of it. Currently,
> > both load_pin and SELinux are gating the kernel module syscalls based
> > on security_kernel_read_file.
> >
> > If there was a similar option for the kernel image, I'm pretty sure
> > other LSMs would use it.
> >
> > From an IMA perspective, there needs to be some method for only
> > allowing signed code to be loaded, executed, etc. - kernel modules,
> > kernel image/initramfs, firmware, policies.
>
> What is the IMA perspective. Why can't IMA trust appropriately
> authorized userspace?
Suppose a system owner wants to define a system wide policy that
requires all code be signed - kernel modules, firmware, kexec image &
initramfs, executables, mmapped files, etc - without having to rebuild
the kernel. Without a call in kexec_load that isn't possible.
>
> >> If you don't trust userspace that needs to be spelled out very clearly.
> >> You need to talk about what your threat models are.
> >>
> >> If the only justification is so that that we can't boot windows if
> >> someone hacks into userspace it has my nack because that is another kind
> >> of complete non-sense.
> >
> > The usecase is the ability to gate the kexec_load usage in stock
> > kernels.
>
> But kexec_load is already gated. It requires CAP_SYS_BOOT.
It isn't a matter of kexec_load already being gated, but of wanting a
single place for defining a system wide policy, as described above.
Mimi
>
> >> Given that you are not trusting userspace this changeset also probably
> >> needs to have the kernel-hardening list cc'd. Because the only possible
> >> justification I can imagine for something like this is kernel-hardening.
> >
> > Sure, Cc'ing linux-hardening and Kees.
> >
> > Mimi
>
Matthew Garrett <[email protected]> writes:
> On Thu, May 3, 2018 at 1:13 PM Eric W. Biederman <[email protected]>
> wrote:
>
>> Mimi Zohar <[email protected]> writes:
>
>> > In environments that require the kexec kernel image to be signed,
> prevent
>> > using the kexec_load syscall. In order for LSMs and IMA to
> differentiate
>> > between kexec_load and kexec_file_load syscalls, this patch set adds a
>> > call to security_kernel_read_file() in kexec_load_check().
>
>> Having thought about it some more this justification for these changes
>> does not work. The functionality of kexec_load is already root-only.
>> So in environments that require the kernel image to be signed just don't
>> use kexec_load. Possibly even compile kexec_load out to save space
>> because you will never need it. You don't need a new security hook to
>> do any of that. Userspace is a very fine mechanism for being the
>> instrument of policy.
>
>> If you don't trust userspace that needs to be spelled out very clearly.
>> You need to talk about what your threat models are.
>
> kexec_load gives root arbitrary power to modify the running kernel image,
> including the ability to disable enforcement of module signatures.
No. It does absolutely nothing to the running kernel image.
Combined with reboot(..., LINUX_REBOOT_CMD_KEXE, ...) it does allow
booting something different. It is argubably a little more efficient
than writing to a file to direct the bootloader to boot something
different and then calling reboot. But it is not fundamentally
different.
> Given
> that it weakens other security mechanisms that are designed to prevent root
> from disabling them, it makes sense to allow the imposition of an
> equivalent restriction.
Say what. You are saying a lot of words without any specifics. Not a
specific threat mode. Not which security mecahnisms you are worried
about weakening. Not what classes of problems you are trying to defend
against.
I absolutely hate this nonsense. I thought you already went 20 rounds
with Linus and learned you need to be upfront with what you are
concerned about.
I believe reasonable situations can be constructed. But I am not seeing
that happen here.
My hand wavy argument to go with yours is that code paths that are root
only are not audited for security properties. As such the number of
exploitable bus you can find in them is larger than normal. It might be
a little harder to mount xfs or another filesystem with an exploitable
file system image but I expect it exists.
Further nothing I have seen you involved with has been about truly
hardening the system against a hostile root. I have for the last
several years been chipping away at that and you have been nowhere to be
found.
So please be specific. Talk about which threat you are worried about.
Because so far this looks like someones effort to look like they were
doing something without actually caring about real world threats.
Eric
On Thu, May 3, 2018 at 2:59 PM Eric W. Biederman <[email protected]>
wrote:
> Matthew Garrett <[email protected]> writes:
> > kexec_load gives root arbitrary power to modify the running kernel
image,
> > including the ability to disable enforcement of module signatures.
> No. It does absolutely nothing to the running kernel image.
> Combined with reboot(..., LINUX_REBOOT_CMD_KEXE, ...) it does allow
> booting something different. It is argubably a little more efficient
> than writing to a file to direct the bootloader to boot something
> different and then calling reboot. But it is not fundamentally
> different.
It absolutely does - https://mjg59.dreamwidth.org/28746.html gives an
example. The payload just needs to return.
> > Given
> > that it weakens other security mechanisms that are designed to prevent
root
> > from disabling them, it makes sense to allow the imposition of an
> > equivalent restriction.
> Say what. You are saying a lot of words without any specifics. Not a
> specific threat mode. Not which security mecahnisms you are worried
> about weakening. Not what classes of problems you are trying to defend
> against.
I have a kernel configured with module.sig_enforce enabled - root is unable
to load kernel modules that are unsigned, and since sig_enforce is
bool_enable_only, root is unable to flip that back. Any number of security
models may be implemented with that assumption. However, root still has
access to kexec_load(), and can therefore kexec into a dummy payload that
flips that byte back to 0 and permits loading unsigned module code.
There may well be other mechanisms that permit root to gain arbitrary
ability to modify kernel code. My argument is that we should treat those as
bugs, not use their existence as a justification for leaving open known
cases.
Mimi Zohar <[email protected]> writes:
> On Thu, 2018-05-03 at 16:38 -0500, Eric W. Biederman wrote:
>> Mimi Zohar <[email protected]> writes:
>>
>> > [Cc'ing Kees and kernel-hardening]
>> >
>> > On Thu, 2018-05-03 at 15:13 -0500, Eric W. Biederman wrote:
>> >> Mimi Zohar <[email protected]> writes:
>> >>
>> >> > In environments that require the kexec kernel image to be signed, prevent
>> >> > using the kexec_load syscall. In order for LSMs and IMA to differentiate
>> >> > between kexec_load and kexec_file_load syscalls, this patch set adds a
>> >> > call to security_kernel_read_file() in kexec_load_check().
>> >>
>> >> Having thought about it some more this justification for these changes
>> >> does not work. The functionality of kexec_load is already root-only.
>> >> So in environments that require the kernel image to be signed just don't
>> >> use kexec_load. Possibly even compile kexec_load out to save space
>> >> because you will never need it. You don't need a new security hook to
>> >> do any of that. Userspace is a very fine mechanism for being the
>> >> instrument of policy.
>> >
>> > True, for those building their own kernel, they can disable the old
>> > syscalls. The concern is not for those building their own kernels,
>> > but for those using stock kernels.
>> >
>> > By adding an LSM hook here in the kexec_load syscall, as opposed to an
>> > IMA specific hook, other LSMs can piggy back on top of it. Currently,
>> > both load_pin and SELinux are gating the kernel module syscalls based
>> > on security_kernel_read_file.
>> >
>> > If there was a similar option for the kernel image, I'm pretty sure
>> > other LSMs would use it.
>> >
>> > From an IMA perspective, there needs to be some method for only
>> > allowing signed code to be loaded, executed, etc. - kernel modules,
>> > kernel image/initramfs, firmware, policies.
>>
>> What is the IMA perspective. Why can't IMA trust appropriately
>> authorized userspace?
>
> Suppose a system owner wants to define a system wide policy that
> requires all code be signed - kernel modules, firmware, kexec image &
> initramfs, executables, mmapped files, etc - without having to rebuild
> the kernel. Without a call in kexec_load that isn't possible.
Of course it is. You just make it a requirement that before an
executable will be signed it will be audited to see that it doesn't
call sys_kexec_load. Signing presumably means something. So it should
not be hard to enforce a policy like that on a specialty system call
that most applications will never call.
>> >> If you don't trust userspace that needs to be spelled out very clearly.
>> >> You need to talk about what your threat models are.
>> >>
>> >> If the only justification is so that that we can't boot windows if
>> >> someone hacks into userspace it has my nack because that is another kind
>> >> of complete non-sense.
>> >
>> > The usecase is the ability to gate the kexec_load usage in stock
>> > kernels.
>>
>> But kexec_load is already gated. It requires CAP_SYS_BOOT.
>
> It isn't a matter of kexec_load already being gated, but of wanting a
> single place for defining a system wide policy, as described above.
Signing is only a tool to enforce a policy. Signing by itself is not a
policy. Enforcing any quality controls in the signed executables should
trivially prevent kexec_load from being used.
Eric
On Thu, 2018-05-03 at 18:03 -0500, Eric W. Biederman wrote:
> Mimi Zohar <[email protected]> writes:
>
> > On Thu, 2018-05-03 at 16:38 -0500, Eric W. Biederman wrote:
> >> Mimi Zohar <[email protected]> writes:
> >>
> >> > [Cc'ing Kees and kernel-hardening]
> >> >
> >> > On Thu, 2018-05-03 at 15:13 -0500, Eric W. Biederman wrote:
> >> >> Mimi Zohar <[email protected]> writes:
> >> >>
> >> >> > In environments that require the kexec kernel image to be signed, prevent
> >> >> > using the kexec_load syscall. In order for LSMs and IMA to differentiate
> >> >> > between kexec_load and kexec_file_load syscalls, this patch set adds a
> >> >> > call to security_kernel_read_file() in kexec_load_check().
> >> >>
> >> >> Having thought about it some more this justification for these changes
> >> >> does not work. The functionality of kexec_load is already root-only.
> >> >> So in environments that require the kernel image to be signed just don't
> >> >> use kexec_load. Possibly even compile kexec_load out to save space
> >> >> because you will never need it. You don't need a new security hook to
> >> >> do any of that. Userspace is a very fine mechanism for being the
> >> >> instrument of policy.
> >> >
> >> > True, for those building their own kernel, they can disable the old
> >> > syscalls. The concern is not for those building their own kernels,
> >> > but for those using stock kernels.
> >> >
> >> > By adding an LSM hook here in the kexec_load syscall, as opposed to an
> >> > IMA specific hook, other LSMs can piggy back on top of it. Currently,
> >> > both load_pin and SELinux are gating the kernel module syscalls based
> >> > on security_kernel_read_file.
> >> >
> >> > If there was a similar option for the kernel image, I'm pretty sure
> >> > other LSMs would use it.
> >> >
> >> > From an IMA perspective, there needs to be some method for only
> >> > allowing signed code to be loaded, executed, etc. - kernel modules,
> >> > kernel image/initramfs, firmware, policies.
> >>
> >> What is the IMA perspective. Why can't IMA trust appropriately
> >> authorized userspace?
> >
> > Suppose a system owner wants to define a system wide policy that
> > requires all code be signed - kernel modules, firmware, kexec image &
> > initramfs, executables, mmapped files, etc - without having to rebuild
> > the kernel. Without a call in kexec_load that isn't possible.
>
> Of course it is. You just make it a requirement that before an
> executable will be signed it will be audited to see that it doesn't
> call sys_kexec_load. Signing presumably means something. So it should
> not be hard to enforce a policy like that on a specialty system call
> that most applications will never call.
Initially I'm hoping that files will simply come signed, providing
file provenance. Anything else is gravy.
> >> >> If you don't trust userspace that needs to be spelled out very clearly.
> >> >> You need to talk about what your threat models are.
> >> >>
> >> >> If the only justification is so that that we can't boot windows if
> >> >> someone hacks into userspace it has my nack because that is another kind
> >> >> of complete non-sense.
> >> >
> >> > The usecase is the ability to gate the kexec_load usage in stock
> >> > kernels.
> >>
> >> But kexec_load is already gated. It requires CAP_SYS_BOOT.
> >
> > It isn't a matter of kexec_load already being gated, but of wanting a
> > single place for defining a system wide policy, as described above.
>
> Signing is only a tool to enforce a policy. Signing by itself is not a
> policy. Enforcing any quality controls in the signed executables should
> trivially prevent kexec_load from being used.
Existing kernels might not support the newer kexec_file_load syscall,
so packages are currently being built to try one syscall and fallback
to using the other one. In this case, it has nothing to do with
quality control.
Mimi