LLVM's safe-stack feature is a compiler mode intended to harden the generated code against stack-smashing attacks such as exploits of buffer overrun bugs.
The Clang/LLVM documentation page linked above describes the general scheme. The capsule summary is that that each thread has two stacks instead of the usual one: a “safe stack” and an “unsafe stack”. The unsafe stack is used for all purposes where a pointer into the stack memory might be used, while the safe stack is used only for purposes where no code should ever see a pointer into the stack memory. So, the unsafe stack is used for arrays or variables that are passed by reference to another function or have their addresses stored in the heap--memory that could be subject to buffer overrun or use-after-free bugs and their exploits. The safe stack is used for the compiler's register spills, and for the return address of a function call. Thus, for example, a simple buffer overrun bug cannot be exploited to overwrite the return address of the containing function, which is the basis of exploits and attacks using the so-called ROP (“return-oriented programming”) technique.
The Compatibility section of that page does not apply to Zircon (or Fuchsia). In Zircon user-mode code (including all of Fuchsia), the runtime support for SafeStack is included directly in the standard C runtime library, and everything works fine in shared libraries (DSOs).
The safe-stack and shadow-call-stack instrumentation schemes and ABIs are related and similar but also orthogonal. Each can be enabled or disabled independently for any function. Fuchsia‘s compiler ABI and libc always interoperate with code built with or without either kind of instrumentation, regardless of what instrumentation was or wasn’t used in the particular libc build.
In general, safe-stack does not affect the ABI. The machine-specific calling conventions are unchanged. It works fine to have some functions in a program built with safe-stack and some not. It doesn't matter if combining the two comes from directly compiled
.o files, from archive libraries (
.a files), or from shared libraries (
.so files), in any combination.
While there is some additional per-thread state (the unsafe stack pointer, see below under Implementation details), code not using safe-stack does not need to do anything about this state to keep it correct when calling, or being called by, code that does use safe-stack. The only potential exceptions to this are for code that is implementing its own kinds of non-local exits or context-switching (e.g. coroutines). The Zircon C library‘s
longjmp code saves and restores this additional state automatically, so anything that is based on
longjmp already handles everything correctly even if the code calling
longjmp doesn’t know about safe-stack.
This is enabled in the Clang compiler by the
-fsanitize=safe-stack command-line option. This is the default mode of the compiler for
*-fuchsia targets. To disable it for a specific compilation, use the
Zircon supports safe-stack for both user-mode and kernel code. In the Zircon build, safe-stack is always enabled when building with Clang (pass
variants = [ "clang" ] to
The essential addition to support safe-stack code is the unsafe stack pointer. In the abstract, this can be thought of as an additional register just like the machine‘s normal stack pointer register. The machine’s stack pointer register is used for the safe stack, just as it always has been. The unsafe stack pointer is used as if it were another register with a fixed purpose in the ABI, but of course the machines don't actually have a new register, and for compatibility safe-stack does not change the basic machine-specific calling conventions that assign uses to all the machine registers.
The C and C++ ABI for Zircon and Fuchsia stores the unsafe stack pointer in memory that's at a fixed offset from the thread pointer. The
<zircon/tls.h> header defines the offset for each machine.
For x86 user-mode, the thread pointer is the
fsbase, meaning access in assembly code looks like
%fs:ZX_TLS_UNSAFE_SP_OFFSET. For the x86 kernel, the thread pointer is the
gsbase, meaning access in assembly code looks like
For Aarch64 (ARM64), in C or C++ code,
__builtin_thread_pointer() returns the thread pointer. In user-mode, the thread pointer is in the
TPIDR_EL0 special register and must be fetched into a normal register (with
mrs *reg*, TPIDR_EL0) to access the memory, so it‘s not a single instruction in assembly code. In the kernel, it’s just the same but using the
TPIDR_EL1 special register instead.
Most code, even in assembly, does not need to think about safe-stack issues at all. The calling conventions are not changed. Using the stack for saving registers, finding return addresses, etc. is all the same with or without safe-stack. The main exception is code that is implementing something like a non-local exit or context switch. Such code may need to save or restore the unsafe stack pointer. Both the
longjmp function and C++
throw already handle this directly, so C or C++ code using those constructs does not need to do anything new.
The context-switch code in the kernel handles switching the unsafe stack pointer. On x86, this is explicit in the code:
%gs points to the
struct x86_percpu, which has a member
arch_context_switch copies this into the
unsafe_sp field of the old thread‘s
struct arch_thread and then copies the new thread’s
kernel_unsafe_sp. On ARM64, this is implicitly done by
set_current_thread, because that changes the
TPIDR_EL1 special register, which points directly into the per-thread
struct thread rather than a per-CPU structure like on x86.
New code implementing some new kind of non-local exit or context switch will need to handle the unsafe stack pointer similarly to how it handles the traditional machine stack pointer register. Any such code should use
#if __has_feature(safe_stack) to test at compile time whether safe-stack is being used in the particular build. That preprocessor construct can be used in C, C++, or assembly (
.S) source files.