Mimi Zohar <zohar@xxxxxxxxxxxxxxxxxx> writes: > On Thu, 2018-05-03 at 16:38 -0500, Eric W. Biederman wrote: >> Mimi Zohar <zohar@xxxxxxxxxxxxxxxxxx> writes: >> >> > [Cc'ing Kees and kernel-hardening] >> > >> > On Thu, 2018-05-03 at 15:13 -0500, Eric W. Biederman wrote: >> >> Mimi Zohar <zohar@xxxxxxxxxxxxxxxxxx> 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
