Re: [PATCH v5] x86: use builtins to read eflags

[Linus Torvalds]

On Fri, Mar 18, 2022 at 2:48 PM Andrew Cooper <Andrew.Cooper3@xxxxxxxxxx> wrote:
>
> As such, I'm not sure how current_stack_pointer can work as intended in
> all cases...

So as mentioned, the kernel doesn't really care, since for the kernel
that inline asm use is more of a "get proper backtraces" thing. The
compiler thinks the function is a leaf function, doesn't set up a
frame for a call that  happens inside the inline asm.

The code *works* without it, but the frame annotations aren't right.

And obviously we don't actually *(change* the stack pointer. Or
rather, a call will do exactly as with pushf/pop: rsp gets updated
there but gets put right back.

See commit 317c2ce77d8a ("x86/alternatives: Add stack frame dependency
to alternative_call_2()") for some details.

And yes, that trick then caused problems with clang, and we did
f5caf621ee35 ("x86/asm: Fix inline asm call constraints for Clang")
that created the current situation.

It would be lovely to have some explicit model for "I want the frame
to have been set up for backtraces", but here we are. Marking '%rsp
used makes the compiler understand it's not a leaf function.

And while we have other uses for it that then use the actual value,
those don't care about the exact value of the stack pointer register,
they just want "give me a pointer that is contained within the current
stack", because we control the stack allocation and do funky things
there. So "any random stack pointer value in this function" is
perfectly fine and expected.

But for user mode, it would probably be a great idea to also have a "I
cannot use a redzone in this function" thing. The kernel can't use it
because we have nested exceptions, but maybe some day even the kernel
could make use of (controlled) red-zoning.

            Linus