Re: [PATCH v3 1/1] security: Add mechanism to safely (un)load LSMs after boot time

From: Casey Schaufler
Date: Sat Mar 31 2018 - 17:35:06 EST

On 3/30/2018 11:16 PM, Sargun Dhillon wrote:
> On Fri, Mar 30, 2018 at 2:39 PM, Casey Schaufler <casey@xxxxxxxxxxxxxxxx> wrote:
>> On 3/29/2018 7:33 PM, Sargun Dhillon wrote:
>>> On Thu, Mar 29, 2018 at 02:37:10PM -0700, Casey Schaufler wrote:
>>>> On 3/29/2018 2:14 PM, Sargun Dhillon wrote:
>>>>> This patch introduces a mechanism to add mutable hooks and immutable
>>>>> hooks to the callback chain. It adds an intermediary item to the
>>>>> chain which separates mutable and immutable hooks. Immutable hooks
>>>>> are then marked as read-only, as well as the hook heads. This does
>>>>> not preclude some hooks being able to be mutated (removed).
>>>>>
>>>>> It also wraps the hook unloading, and execution with an SRCU. One
>>>>> SRCU is used across all hooks, as the SRCU struct can be memory
>>>>> intensive, and hook execution time in general should be relatively
>>>>> short.
>>>>>
>>>>> Signed-off-by: Sargun Dhillon <sargun@xxxxxxxxx>
>>>>> Signed-off-by: Tetsuo Handa <penguin-kernel@xxxxxxxxxxxxxxxxxxx>
>>>>> ---
>>>>> include/linux/lsm_hooks.h | 23 ++---
>>>>> security/Kconfig | 2 +-
>>>>> security/apparmor/lsm.c | 2 +-
>>>>> security/commoncap.c | 2 +-
>>>>> security/security.c | 210 ++++++++++++++++++++++++++++++++++++++-------
>>>>> security/selinux/hooks.c | 5 +-
>>>>> security/smack/smack_lsm.c | 3 +-
>>>>> security/tomoyo/tomoyo.c | 3 +-
>>>>> security/yama/yama_lsm.c | 2 +-
>>>>> 9 files changed, 196 insertions(+), 56 deletions(-)
>>>>>
>>>>> diff --git a/include/linux/lsm_hooks.h b/include/linux/lsm_hooks.h
>>>>> index 09bc60fb35f1..689e5e72fb38 100644
>>>>> --- a/include/linux/lsm_hooks.h
>>>>> +++ b/include/linux/lsm_hooks.h
>>>>> @@ -1981,9 +1981,12 @@ extern struct security_hook_heads security_hook_heads;
>>>>> extern char *lsm_names;
>>>>>
>>>>> extern void security_add_hooks(struct security_hook_list *hooks, int count,
>>>>> - char *lsm);
>>>>> + char *lsm, bool is_mutable);
>>>>>
>>>>> -#ifdef CONFIG_SECURITY_SELINUX_DISABLE
>>>>> +#define __lsm_ro_after_init __ro_after_init
>>>>> +/* Currently required to handle SELinux runtime hook disable. */
>>>>> +#ifdef CONFIG_SECURITY_WRITABLE_HOOKS
>>>>> +#define __lsm_mutable_after_init
>>>>> /*
>>>>> * Assuring the safety of deleting a security module is up to
>>>>> * the security module involved. This may entail ordering the
>>>>> @@ -1996,21 +1999,9 @@ extern void security_add_hooks(struct security_hook_list *hooks, int count,
>>>>> * disabling their module is a good idea needs to be at least as
>>>>> * careful as the SELinux team.
>>>>> */
>>>>> -static inline void security_delete_hooks(struct security_hook_list *hooks,
>>>>> - int count)
>>>>> -{
>>>>> - int i;
>>>>> -
>>>>> - for (i = 0; i < count; i++)
>>>>> - hlist_del_rcu(&hooks[i].list);
>>>>> -}
>>>>> -#endif /* CONFIG_SECURITY_SELINUX_DISABLE */
>>>>> -
>>>>> -/* Currently required to handle SELinux runtime hook disable. */
>>>>> -#ifdef CONFIG_SECURITY_WRITABLE_HOOKS
>>>>> -#define __lsm_ro_after_init
>>>>> +extern void security_delete_hooks(struct security_hook_list *hooks, int count);
>>>>> #else
>>>>> -#define __lsm_ro_after_init __ro_after_init
>>>>> +#define __lsm_mutable_after_init __ro_after_init
>>>>> #endif /* CONFIG_SECURITY_WRITABLE_HOOKS */
>>>>>
>>>>> extern int __init security_module_enable(const char *module);
>>>>> diff --git a/security/Kconfig b/security/Kconfig
>>>>> index c4302067a3ad..a3b8b1142e6f 100644
>>>>> --- a/security/Kconfig
>>>>> +++ b/security/Kconfig
>>>>> @@ -32,7 +32,7 @@ config SECURITY
>>>>> If you are unsure how to answer this question, answer N.
>>>>>
>>>>> config SECURITY_WRITABLE_HOOKS
>>>>> - depends on SECURITY
>>>>> + depends on SECURITY && SRCU
>>>>> bool
>>>>> default n
>>>>>
>>>>> diff --git a/security/apparmor/lsm.c b/security/apparmor/lsm.c
>>>>> index 9a65eeaf7dfa..d6cca8169df0 100644
>>>>> --- a/security/apparmor/lsm.c
>>>>> +++ b/security/apparmor/lsm.c
>>>>> @@ -1155,7 +1155,7 @@ static int __init apparmor_init(void)
>>>>> goto buffers_out;
>>>>> }
>>>>> security_add_hooks(apparmor_hooks, ARRAY_SIZE(apparmor_hooks),
>>>>> - "apparmor");
>>>>> + "apparmor", false);
>>>>>
>>>>> /* Report that AppArmor successfully initialized */
>>>>> apparmor_initialized = 1;
>>>>> diff --git a/security/commoncap.c b/security/commoncap.c
>>>>> index 48620c93d697..fe4b0d9d44ce 100644
>>>>> --- a/security/commoncap.c
>>>>> +++ b/security/commoncap.c
>>>>> @@ -1363,7 +1363,7 @@ struct security_hook_list capability_hooks[] __lsm_ro_after_init = {
>>>>> void __init capability_add_hooks(void)
>>>>> {
>>>>> security_add_hooks(capability_hooks, ARRAY_SIZE(capability_hooks),
>>>>> - "capability");
>>>>> + "capability", false);
>>>>> }
>>>>>
>>>>> #endif /* CONFIG_SECURITY */
>>>>> diff --git a/security/security.c b/security/security.c
>>>>> index 3cafff61b049..2ddb64864e3e 100644
>>>>> --- a/security/security.c
>>>>> +++ b/security/security.c
>>>>> @@ -29,6 +29,11 @@
>>>>> #include <linux/backing-dev.h>
>>>>> #include <linux/string.h>
>>>>> #include <net/flow.h>
>>>>> +#include <linux/srcu.h>
>>>>> +#include <linux/mutex.h>
>>>>> +
>>>>> +#define SECURITY_HOOK_COUNT \
>>>>> + (sizeof(security_hook_heads) / sizeof(struct hlist_head))
>>>>>
>>>>> #define MAX_LSM_EVM_XATTR 2
>>>>>
>>>>> @@ -36,7 +41,10 @@
>>>>> #define SECURITY_NAME_MAX 10
>>>>>
>>>>> struct security_hook_heads security_hook_heads __lsm_ro_after_init;
>>>>> +EXPORT_SYMBOL_GPL(security_hook_heads);
>>>>> +
>>>>> static ATOMIC_NOTIFIER_HEAD(lsm_notifier_chain);
>>>>> +static DEFINE_MUTEX(security_hook_mutex);
>>>>>
>>>>> char *lsm_names;
>>>>> /* Boot-time LSM user choice */
>>>>> @@ -53,6 +61,103 @@ static void __init do_security_initcalls(void)
>>>>> }
>>>>> }
>>>>>
>>>>> +#ifdef CONFIG_SECURITY_WRITABLE_HOOKS
>>>>> +DEFINE_STATIC_SRCU(security_hook_srcu);
>>>>> +static struct security_hook_list null_hooks[SECURITY_HOOK_COUNT];
>>>>> +#define HAS_FUNC(SHL, FUNC) (SHL->hook.FUNC)
>>>> The HAS_FUNC() macro will work, but it's awkward outside of the
>>>> call_..._hook() macros. I think you should document how to use it
>>>> properly somewhere in here. There are enough cases where the
>>>> call_..._hook() macros aren't used that someone could have trouble
>>>> figuring out how to use it.
>>>>
>>>>
>>> What about something like:
>>>
>>> security/security.c | 58 ++++++++++++++++++++++++++++++++++++++---------------
>>> 1 file changed, 42 insertions(+), 16 deletions(-)
>>>
>>> diff --git a/security/security.c b/security/security.c
>>> index 2ddb64864e3e..bc14125cfc78 100644
>>> --- a/security/security.c
>>> +++ b/security/security.c
>>> @@ -62,9 +62,37 @@ static void __init do_security_initcalls(void)
>>> }
>>>
>>> #ifdef CONFIG_SECURITY_WRITABLE_HOOKS
>>> -DEFINE_STATIC_SRCU(security_hook_srcu);
>>> +/*
>>> + * With writable hooks, we setup a structure like this:
>>> + * +------+ +-----------+ +-----------+ +-----------+ +--------------+
>>> + * | | | | | | | | | |
>>> + * | HEAD +---> Immutable +---> Immutable +---> Null hook +---> Mutable Hook |
>>> + * | | | Hook 1 | | Hook 2 | | | | |
>>> + * +------+ +-----------+ +-----------+ +-----------+ +--------------+
>>> + * | | |
>>> + * v v v
>>> + * Callback Callback Callback
>>> + *
>>> + * The hooks before to null hook are marked only after kernel initialization.
>>> + * The null hook, as well as the hooks succeeding it are not marked read only,
>>> + * therefore allowing them be (un)loaded after initialization time.
>>> + *
>>> + * Since the null hook doesn't have a callback, we need to check if a hook
>>> + * is the null hook prior to invoking it.
>>> + */
>> I think a comment like this is helpful.
>>
>> Why not have two hook list heads, one for regular hooks and
>> one for mutable hooks? You can dispense with the "null hook"
>> handling.
>>
> The iterations gets really messy. The patch earlier had a lot of
> awkward gunk in it. The issue primarily came with deciding to iterate
> over the second set of hooks iff the first set of hooks failed. With
> the way the calls are setup, it's messy.

Hang on. I would think you'd only call the 2nd list if the first list
was completed successfully. And of course the void hooks always get
called regardless.

> What's your primary issue with the "null hooks"? If it's the null hook
> checks, I think I can actually remove those.

In the common case of one LSM and no mutable modules you
are making a check on every hook that will always have the
same answer. Performance impact *of any amount* that does
not add value in the common case has to be avoided.

This is a clever implementation. But it does not generalize.
If I wanted to add a third kind of hooks the mechanism would
not support it. If you add a 2nd list it's obvious what you'd
want to do to add a 3rd.

>
>>> static struct security_hook_list null_hooks[SECURITY_HOOK_COUNT];
>>> -#define HAS_FUNC(SHL, FUNC) (SHL->hook.FUNC)
>>> +DEFINE_STATIC_SRCU(security_hook_srcu);
>>> +
>>> +static inline bool is_null_hook(struct security_hook_list *shl)
>>> +{
>>> + union {
>>> + void *cb_ptr;
>>> + union security_list_options slo;
>>> + } hook_options;
>>> +
>>> + hook_options.slo = shl->hook;
>>> + return !hook_options.cb_ptr;
>>> +}
>> I like the HAS_FUNC() approach better.
> Just curious, why? I personally prefer small static inline functions
> over macros, if possible.

Minimize the complexity, even inside a macro/function. Your
is_null_hook() function is massively more complicated than the
HAS_FUNC() macro. I'm still not 100% sure I understand why you
have the wacky union, and how you can know that it will get you
the right result on all architectures.

>
>> What I think would work best is to have a separate list head
>> for the mutable hooks.
>>
>>
>>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
>