Re: [PATCH v2 02/10] x86: assembly, FUNC_START for fn, DATA_START for data

From: Josh Poimboeuf
Date: Wed Mar 22 2017 - 10:12:12 EST


On Wed, Mar 22, 2017 at 08:46:16AM +0100, Ingo Molnar wrote:
>
> * Jiri Slaby <jslaby@xxxxxxx> wrote:
>
> > On 03/22/2017, 08:25 AM, Ingo Molnar wrote:
> > >
> > > * Pavel Machek <pavel@xxxxxx> wrote:
> > >
> > >> Hi!
> > >>
> > >>> -ENTRY(saved_rbp) .quad 0
> > >>> -ENTRY(saved_rsi) .quad 0
> > >>> -ENTRY(saved_rdi) .quad 0
> > >>> -ENTRY(saved_rbx) .quad 0
> > >>> +SYM_DATA_START(saved_rbp) .quad 0
> > >>> +SYM_DATA_START(saved_rsi) .quad 0
> > >>> +SYM_DATA_START(saved_rdi) .quad 0
> > >>> +SYM_DATA_START(saved_rbx) .quad 0
> > >>
> > >> Does it make sense to call it SYM_DATA_*START* when there's no
> > >> corresponding end?
> > >
> > > That looks like a bug - I think we should strive for them to always be in pairs.
> > >
> > > Jiri, Josh, could objtool help here perhaps, to detect 'non-terminated'
> > > SYM_*_START() uses? This could be done by emitting debug data into a special
> > > section and then analyzing that section for unpaired entries. The section can be
> > > discarded in the final link, it won't show up in the kernel image.
> >
> > It should be easier than that. No introduction of other info needed --
> > every global symbol without a ".type" or ".size" (i.e. SYM_*_END) should
> > be a bug now.
>
> I'm all for that!

It would be easy to add this checking to objtool since it already reads
the symbol table. The hard part is figuring out the logistics. :-)

- Should the warnings be on by default?

- Part of the "objtool check" command or something else?

- Separate config option or just include it with
CONFIG_STACK_VALIDATION?

- Should all asm files be checked, including those currently skipped by
objtool with OBJECT_FILES_NON_STANDARD?

> Can we detect double ends as well - i.e. do a build check of the full syntax of
> these symbol definition primitives?

Detecting double ends would be a little trickier. The second SYM_*_END
supersedes the first, so that information isn't in the ELF symbol table.

We could use a special section to annotate all the macro uses and have
objtool do the checking, similar to what you suggested earlier.

Or, here's a much easier way to do it, without involving objtool:

--- a/include/linux/linkage.h
+++ b/include/linux/linkage.h
@@ -138,9 +138,17 @@
name:
#endif

+#ifndef CHECK_DUP_SYM_END
+#define CHECK_DUP_SYM_END(name) \
+ .pushsection .discard.sym_func_end ASM_NL \
+ SYM_END_##name: .byte 0 ASM_NL \
+ .popsection
+#endif
+
/* SYM_END -- use only if you have to */
#ifndef SYM_END
#define SYM_END(name, sym_type) \
+ CHECK_DUP_SYM_END(name) ASM_NL \
.type name sym_type ASM_NL \
.size name, .-name
#endif


If there's an extra SYM_*_END, the build fails. For example, if I add
an extra SYM_FUNC_END(\name) to the THUNK macro:

AS arch/x86/entry/thunk_64.o
arch/x86/entry/thunk_64.S: Assembler messages:
arch/x86/entry/thunk_64.S:42: Error: symbol `SYM_END_trace_hardirqs_on_thunk' is already defined
arch/x86/entry/thunk_64.S:43: Error: symbol `SYM_END_trace_hardirqs_off_thunk' is already defined
arch/x86/entry/thunk_64.S:47: Error: symbol `SYM_END_lockdep_sys_exit_thunk' is already defined
arch/x86/entry/thunk_64.S:51: Error: symbol `SYM_END____preempt_schedule' is already defined
arch/x86/entry/thunk_64.S:52: Error: symbol `SYM_END____preempt_schedule_notrace' is already defined
scripts/Makefile.build:395: recipe for target 'arch/x86/entry/thunk_64.o' failed

--
Josh