On Sun, Jun 14, 2015 at 11:50 PM, Theodore Ts'o <tytso at mit.edu> wrote: > From experimentation and from looking at the sources, it appears that > the signature checking is only done in the kexec_file_load(2) system > all, and not in the kexec_load(2) system call. And I understand why > -- the signature is not sent from userspace to the kernel in the older > kexec_load(2) system call. > > The problem is that if you use an old version of kexec, it will use > the old kexec_load(2) system call, and even though > CONFIG_KEXEC_VERIFY_SIG is enabled, kexec_load(2) will happily load an > unsigned kernel, and then "kexec -e" will happily boot into it. > > Correct me if I am wrong, but this appears to be a hole in Secure Boot > you could drive a Mack Truck through. Yes, which is why most of the distro vendors carry an out-of-tree patch that disables the old kexec in an SB setup. It would be nice if we could merge said patches. However, they depend on Matthew's secure_modules/trusted_kernel/<whatever name that works> patchset which has gotten little movement since we came up with a tentative agreement at LPC 2013. > (I noticed this because Debian is still using a kexec-tools from the > stone ages, version 2.0.7, and I was wondering **why** I was able to > kexec boot completely unsigned kernels.) > > It would appear to me that if CONFIG_KEXEC_VERIFY_SIG is enabled, the > old kexec_load(2) system call should be disabled (and a warning be > placed in the Kconfig help that the user should have at least verision > 2.X of kexec-tools if they enable this kernel option). > > Am I missing something? Those sound like fine suggestions to me. josh
