Re: [kvm-unit-tests PATCH v2 2/4] x86: Better backtraces for leaf functions

[Mathias Krause <minipli@grsecurity.net>]

On 9/15/25 23:54, Mathias Krause wrote:
> Leaf functions are problematic for backtraces as they lack the frame
> pointer setup epilogue. If such a function causes a fault, the original
> caller won't be part of the backtrace. That's problematic if, for
> example, memcpy() is failing because it got passed a bad pointer. The
> generated backtrace will look like this, providing no clue what the
> issue may be:
> 
> 	STACK: @401b31 4001ad
>   0x0000000000401b31: memcpy at lib/string.c:136 (discriminator 3)
>         	for (i = 0; i < n; ++i)
>       > 		a[i] = b[i];
> 
>   0x00000000004001ac: gdt32_end at x86/cstart64.S:127
>         	lea __environ(%rip), %rdx
>       > 	call main
>         	mov %eax, %edi
> 
> By abusing profiling, we can force the compiler to emit a frame pointer
> setup epilogue even for leaf functions, making the above backtrace
> change like this:
> 
> 	STACK: @401c21 400512 4001ad
>   0x0000000000401c21: memcpy at lib/string.c:136 (discriminator 3)
>         	for (i = 0; i < n; ++i)
>       > 		a[i] = b[i];
> 
>   0x0000000000400511: main at x86/hypercall.c:91 (discriminator 24)
> 
>       > 	memcpy((void *)~0xbadc0de, (void *)0xdeadbeef, 42);
> 
>   0x00000000004001ac: gdt32_end at x86/cstart64.S:127
>         	lea __environ(%rip), %rdx
>       > 	call main
>         	mov %eax, %edi
> 
> Above backtrace includes the failing memcpy() call, making it much
> easier to spot the bug.
> 
> Enable "fake profiling" if supported by the compiler to get better
> backtraces. The runtime overhead should be negligible for the gained
> debugability as the profiling call is actually a NOP.
> 
> Signed-off-by: Mathias Krause <minipli@grsecurity.net>
> ---
> One may argure that the "ifneq ($(KEEP_FRAME_POINTER),) ... endif"
> wrapping isn't needed, and that's true. However, it simplifies toggling
> that variable, if there'll ever be a need for it.
> 
>  x86/Makefile.common | 11 +++++++++++
>  1 file changed, 11 insertions(+)
> 
> diff --git a/x86/Makefile.common b/x86/Makefile.common
> index 5663a65d3df4..be18a77a779e 100644
> --- a/x86/Makefile.common
> +++ b/x86/Makefile.common
> @@ -43,6 +43,17 @@ COMMON_CFLAGS += -O1
>  # stack.o relies on frame pointers.
>  KEEP_FRAME_POINTER := y
>  
> +ifneq ($(KEEP_FRAME_POINTER),)
> +# Fake profiling to force the compiler to emit a frame pointer setup also in
> +# leaf function (-mno-omit-leaf-frame-pointer doesn't work, unfortunately).
> +#
> +# Note:
> +# We need to defer the cc-option test until -fno-pic or -no-pie have been
> +# added to CFLAGS as -mnop-mcount needs it. The lazy evaluation of CFLAGS
> +# during compilation makes this do "The Right Thing."
> +LATE_CFLAGS += $(call cc-option, -pg -mnop-mcount, "")
> +endif
> +
>  FLATLIBS = lib/libcflat.a
>  
>  ifeq ($(CONFIG_EFI),y)

Paolo, can you please comment on this one, so the fixes for ARM and
AArch64 are no longer blocked and, ideally, this series can be merged?

Thanks,
Mathias