On Thu, Nov 30, 2023 at 07:00:58PM +0000, Catalin Marinas wrote:
My hope when looking at the arm64 patches was that we can completely avoid the kernel allocation/deallocation of the shadow stack since it doesn't need to do this for the normal stack either. Could someone please summarise why we dropped the shadow stack pointer after v1? IIUC there was a potential security argument but I don't think it was a very strong one. Also what's the threat model for this feature? I thought it's mainly mitigating stack corruption. If some rogue code can do syscalls, we have bigger problems than clone3() taking a shadow stack pointer.
As well as preventing/detecting corruption of the in memory stack shadow stacks are also ensuring that any return instructions are unwinding a prior call instruction, and that the returns are done in opposite order to the calls. This forces usage of the stack - any value we attempt to RET to is going to be checked against the top of the shadow stack which makes chaining returns together as a substitute for branches harder.
The concern Rick raised was that allowing user to pick the exact shadow stack pointer would allow userspace to corrupt or reuse the stack of an existing thread by starting a new thread with the shadow stack pointing into the existing shadow stack of that thread. While in isolation that's not too much more than what userspace could just do directly anyway it might compose with other issues to something more "interesting" (eg, I'd be a bit concerned about overlap with pkeys/POE though I've not thought through potential uses in detail).
I'm not against clone3() getting a shadow_stack_size argument but asking some more questions. If we won't pass a pointer as well, is there any advantage in expanding this syscall vs a specific prctl() option? Do we need a different size per thread or do all threads have the same shadow stack size? A new RLIMIT doesn't seem to map well though, it is more like an upper limit rather than a fixed/default size (glibc I think uses it for thread stacks but bionic or musl don't AFAIK).
I don't know what the userspace patterns are likely to be here, it's possible a single value for each process might be fine but I couldn't say that confidently. I agree that a RLIMIT does seem like a poor fit.
As well as the actual configuration of the size the other thing that we gain is that as well as relying on heuristics to determine if we need to allocate a new shadow stack for the new thread we allow userspace to explicitly request a new shadow stack. There was some corner case with IIRC posix_nspawn() mentioned where the heuristics aren't what we want for example.
Another dumb question on arm64 - is GCSPR_EL0 writeable by the user? If yes, can the libc wrapper for threads allocate a shadow stack via map_shadow_stack() and set it up in the thread initialisation handler before invoking the thread function?
No, GCSPR_EL0 can only be changed by EL0 through BL, RET and the new GCS instructions (push/pop and stack switch). Push is optional - userspace has to explicitly request that it be enabled and this could be prevented through seccomp or some other LSM. The stack switch instructions require a token at the destination address which must either be written by a higher EL or will be written in the process of switching away from a stack so you can switch back. Unless I've missed one every mechanism for userspace to update GCSPR_EL0 will do a GCS memory access so providing guard pages have been allocated wrapping to a different stack will be prevented.
We would need a syscall to allow GCSPR_EL0 to be written.