Re: [PATCH] x86/fgraph: Fix the unwind hint at the end of return_to_handler()

From: Jiri Olsa

Date: Thu Nov 19 2026 - 10:19:19 EST


On Wed, Nov 18, 2026 at 03:50:23PM -0800, Yifan Feng wrote:
> return_to_handler() reserves 8 bytes of stack for the return address
> and marks that with UNWIND_HINT_FUNC. The hint stays in effect after
> "addq $(FRAME_SIZE) + 8, %rsp" has dropped the slot again, so the ORC
> entry of the following call has CFA = sp+0:
>
> return_to_handler+0x41: sp+0 call .Ldo_rop
> return_to_handler+0x46: (und) int3
> return_to_handler+0x47: sp+8 mov %rdi, (%rsp)
>
> A stack trace taken from an interrupt which hits the call or the mov
> ends up at that entry, the stack pointer does not advance and the
> unwinder prints:
>
> WARNING: stack going in the wrong direction? at return_to_handler+0x46/0x50
>
> The return address is only in %rdi at that point. Mark the code as
> UNWIND_HINT_UNDEFINED again, as it was before the hint was added, so
> that the unwinder stops there without the warning.
>
> Fixes: 20a0bc10272f ("x86/fgraph,bpf: Fix stack ORC unwind from kprobe_multi return probe")
> Signed-off-by: Yifan Feng <evanfengff@xxxxxxxxx>

Reviewed-by: Jiri Olsa <jolsa@xxxxxxxxxx>

thanks,
jirka


> ---
> QEMU, function_graph + 20 kHz perf callchains: warning in 2s before, none in 10 min after.
> Found and written with the help of an AI coding assistant.
>
> arch/x86/kernel/ftrace_64.S | 2 ++
> 1 file changed, 2 insertions(+)
>
> diff --git a/arch/x86/kernel/ftrace_64.S b/arch/x86/kernel/ftrace_64.S
> index 4b20673a3395..8b488f99de02 100644
> --- a/arch/x86/kernel/ftrace_64.S
> +++ b/arch/x86/kernel/ftrace_64.S
> @@ -387,6 +387,8 @@ SYM_CODE_START(return_to_handler)
> movq RAX(%rsp), %rax
>
> addq $(FRAME_SIZE) + 8, %rsp
> + /* The return address is only in %rdi now, the unwinder cannot follow. */
> + UNWIND_HINT_UNDEFINED
>
> /*
> * Jump back to the old return address. This cannot be JMP_NOSPEC rdi
>
> base-commit: 20eef29b7c8465ac98d7dd6d6f2ac07a8d8056ca
> --
> 2.39.3 (Apple Git-146)
>