On 11/15/2021 11:08 AM, Alistair Delva wrote:
On Mon, Nov 15, 2021 at 11:04 AM Ondrej Mosnacek omosnace@redhat.com wrote:
On Mon, Nov 15, 2021 at 7:14 PM Alistair Delva adelva@google.com wrote:
Booting to Android userspace on 5.14 or newer triggers the following SELinux denial:
avc: denied { sys_nice } for comm="init" capability=23 scontext=u:r:init:s0 tcontext=u:r:init:s0 tclass=capability permissive=0
Init is PID 0 running as root, so it already has CAP_SYS_ADMIN. For better compatibility with older SEPolicy, check ADMIN before NICE.
But with this patch you in turn punish the new/better policies that try to avoid giving domains CAP_SYS_ADMIN unless necessary (using only the more granular capabilities wherever possible), which may now get a bogus sys_admin denial. IMHO the order is better as it is, as it motivates the "good" policy writing behavior - i.e. spelling out the capability permissions more explicitly and avoiding CAP_SYS_ADMIN.
IOW, if you domain does CAP_SYS_NICE things, and you didn't explicitly grant it that (and instead rely on the CAP_SYS_ADMIN fallback), then the denial correctly flags it as an issue in your policy and encourages you to add that sys_nice permission to the domain. Then when one beautiful hypothetical day the CAP_SYS_ADMIN fallback is removed, your policy will be ready for that and things will keep working.
Feel free to carry that patch downstream if patching the kernel is easier for you than fixing the policy, but for the upstream kernel this is just a step in the wrong direction.
I'm personally fine with this position, but I am curious why "never break userspace" doesn't apply to SELinux policies.
Because SELinux policy is configuration data, not system code. One is free to modify SELinux policy to suit one's whims without making any change to the Linux kernel or its APIs.
At the end of the day, booting 5.13 or older, we don't get a denial, and there's nothing for the sysadmin to do. On 5.14 and newer, we get denials. This is a common pattern we see each year: some new capability or permission is required where it wasn't required before, and there's no compatibility mechanism to grandfather in old policies.
This is an artifact of separating policy from mechanism. The capability mechanism does not suffer from this issue because it embodies its policy. SELinux, Smack, AppArmor and "containers" are vulnerable to it because they explicitly deny the kernel and kernel developers permission to make assumptions about how they define "policy".
So, we have to touch userspace. If this is just how things are, I can certainly update our init.te definitions.
If SELinux was a required kernel mechanism and the policy was included in the kernel tree you might be able to argue that kernel developers are responsible for changes to SELinux policy. But it ain't, and it isn't. By design.
Fixes: 9d3a39a5f1e4 ("block: grant IOPRIO_CLASS_RT to CAP_SYS_NICE") Signed-off-by: Alistair Delva adelva@google.com Cc: Khazhismel Kumykov khazhy@google.com Cc: Bart Van Assche bvanassche@acm.org Cc: Serge Hallyn serge@hallyn.com Cc: Jens Axboe axboe@kernel.dk Cc: Greg Kroah-Hartman gregkh@linuxfoundation.org Cc: Paul Moore paul@paul-moore.com Cc: selinux@vger.kernel.org Cc: linux-security-module@vger.kernel.org Cc: kernel-team@android.com Cc: stable@vger.kernel.org # v5.14+
block/ioprio.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/block/ioprio.c b/block/ioprio.c index 0e4ff245f2bf..4d59c559e057 100644 --- a/block/ioprio.c +++ b/block/ioprio.c @@ -69,7 +69,7 @@ int ioprio_check_cap(int ioprio)
switch (class) { case IOPRIO_CLASS_RT:
if (!capable(CAP_SYS_NICE) && !capable(CAP_SYS_ADMIN))
if (!capable(CAP_SYS_ADMIN) && !capable(CAP_SYS_NICE)) return -EPERM; fallthrough; /* rt has prio field too */-- 2.34.0.rc1.387.gb447b232ab-goog
-- Ondrej Mosnacek Software Engineer, Linux Security - SELinux kernel Red Hat, Inc.