[PATCH v2 01/13] arm64: stackleak: fix current_top_of_stack()

[Catalin Marinas]

On Wed, Apr 27, 2022 at 06:31:16PM +0100, Mark Rutland wrote:
> Due to some historical confusion, arm64's current_top_of_stack() isn't
> what the stackleak code expects. This could in theory result in a number
> of problems, and practically results in an unnecessary performance hit.
> We can avoid this by aligning the arm64 implementation with the x86
> implementation.
> 
> The arm64 implementation of current_top_of_stack() was added
> specifically for stackleak in commit:
> 
>   0b3e336601b82c6a ("arm64: Add support for STACKLEAK gcc plugin")
> 
> This was intended to be equivalent to the x86 implementation, but the
> implementation, semantics, and performance characteristics differ
> wildly:
> 
> * On x86, current_top_of_stack() returns the top of the current task's
>   task stack, regardless of which stack is in active use.
> 
>   The implementation accesses a percpu variable which the x86 entry code
>   maintains, and returns the location immediately above the pt_regs on
>   the task stack (above which x86 has some padding).
> 
> * On arm64 current_top_of_stack() returns the top of the stack in active
>   use (i.e. the one which is currently being used).
> 
>   The implementation checks the SP against a number of
>   potentially-accessible stacks, and will BUG() if no stack is found.
> 
> The core stackleak_erase() code determines the upper bound of stack to
> erase with:
> 
> | if (on_thread_stack())
> |         boundary = current_stack_pointer;
> | else
> |         boundary = current_top_of_stack();
> 
> On arm64 stackleak_erase() is always called on a task stack, and
> on_thread_stack() should always be true. On x86, stackleak_erase() is
> mostly called on a trampoline stack, and is sometimes called on a task
> stack.
> 
> Currently, this results in a lot of unnecessary code being generated for
> arm64 for the impossible !on_thread_stack() case. Some of this is
> inlined, bloating stackleak_erase(), while portions of this are left
> out-of-line and permitted to be instrumented (which would be a
> functional problem if that code were reachable).
> 
> As a first step towards improving this, this patch aligns arm64's
> implementation of current_top_of_stack() with x86's, always returning
> the top of the current task's stack. With GCC 11.1.0 this results in the
> bulk of the unnecessary code being removed, including all of the
> out-of-line instrumentable code.
> 
> While I don't believe there's a functional problem in practice I've
> marked this as a fix since the semantic was clearly wrong, the fix
> itself is simple, and other code might rely upon this in future.
> 
> Fixes: 0b3e336601b82c6a ("arm64: Add support for STACKLEAK gcc plugin")
> Signed-off-by: Mark Rutland <mark.rutland at arm.com>
> Cc: Alexander Popov <alex.popov at linux.com>
> Cc: Andrew Morton <akpm at linux-foundation.org>
> Cc: Andy Lutomirski <luto at kernel.org>
> Cc: Catalin Marinas <catalin.marinas at arm.com>
> Cc: Kees Cook <keescook at chromium.org>
> Cc: Will Deacon <will at kernel.org>

I thought this was queued already but I couldn't find it in -next. So:

Acked-by: Catalin Marinas <catalin.marinas at arm.com>