[PATCH] riscv/ftrace: Fix the problem modules cannot find _mcount
alankao at andestech.com
Tue May 8 19:36:05 PDT 2018
On Tue, May 08, 2018 at 09:11:42AM -0400, Steven Rostedt wrote:
> On Tue, 8 May 2018 11:21:57 +0800
> Alan Kao <alankao at andestech.com> wrote:
> > Enabling ftrace and module support at the same time fails the kernel
> > build process, because modules cannot find the _mcount symbol. This
> > patch fixes this issue.
> I think you have a much bigger issue.
Yes, we do, but not this one. Let me state more detail later.
> > Signed-off-by: Alan Kao <alankao at andestech.com>
> > Cc: Greentime Hu <greentime at andestech.com>
> > Cc: Zong Li <zong at andestech.com>
> > ---
> > arch/riscv/kernel/mcount.S | 2 +-
> > 1 file changed, 1 insertion(+), 1 deletion(-)
> > diff --git a/arch/riscv/kernel/mcount.S b/arch/riscv/kernel/mcount.S
> > index ce9bdc57a2a1..5721624886a1 100644
> > --- a/arch/riscv/kernel/mcount.S
> > +++ b/arch/riscv/kernel/mcount.S
> > @@ -126,5 +126,5 @@ do_trace:
> > RESTORE_ABI_STATE
> > ret
> > ENDPROC(_mcount)
> > -EXPORT_SYMBOL(_mcount)
> > #endif
> > +EXPORT_SYMBOL(_mcount)
> How does this work? How do you export _mcount if _mcount isn't even
_mcount is defined.
This EXPORT_SYMBOL and the _mcount body was inside a
"#ifndef CONFIG_DYNAMIC_FTRACE" section, so if the config has dynamic ftrace
disabled, _mcount is visible.
For the dynamic ftrace case, there is a code snippet in this file:
> #ifdef CONFIG_DYNAMIC_FTRACE
> .global _mcount
> .set _mcount, ftrace_stub
so the _mcount symbol is visible in the kernel, but not for modules.
As a result, a module build always fails, because _mcount is neither exported,
nor aliased to a defined symbol.
That's the simple purpose of this patch: we just want to eliminate the errors
during the kernel build process.
> -- Steve
Talking about some bigger issues here, there will be twofold.
1. This patch lacks call-site collecting mechanism.
This feature requires a pattern check in recordmcount.pl. With this,
section __mcount_loc is properly constructed, and the call-sites will be
registered during the loading stage.
However, the loading will then fail at the first try of make_nop, due to
the instruction check. This is thus the second issue.
2. The instruction check for kernel texts is not applicable to module texts.
The check expects a call instruction to _mcount, but here it is a call to
the .plt section of the module. After referencing the implementation of arm64,
I think it would need much more work to have make_nop's and make_call's behavior
Quick summary: This patch ensures that a kernel build will not fail because of
the _mcount-missing problem. But ftrace cannot trace loading modules for now
due to the stated reasons.
More information about the linux-riscv