Re: [PATCH v5 02/10] x86: Compile-time asm code validation

[Pedro Alves <palves@redhat.com>]

On 06/12/2015 03:10 PM, Josh Poimboeuf wrote:
> On Fri, Jun 12, 2015 at 12:18:16PM +0100, Pedro Alves wrote:
>> On 06/11/2015 03:10 PM, Josh Poimboeuf wrote:
>>
>>> C would definitely make more sense when analyzing object code.  In fact,
>>> asmvalidate is written in C.  But then I guess we'd have to re-implement
>>> the .cfi stuff and populate the DWARF sections manually instead of
>>> letting the assembler do it.
>>
>> Was doing all this directly in the assembler considered?  That is,
>> e.g., add some knob that makes it error/warn in the same conditions
>> you're making the validator catch.  For tail calls, you'd e.g., add
>> some  new ".nonlocal" directive that you'd use to whitelist the
>> following jump.  And then if it's possible run a CFI generator
>> as a separate step over the source, it sounds like it should also
>> be possible to have the assembler do it instead too (again with
>> some new high level directive to trigger/help it).
> 
> In general I think doing these types of things in the assembler would be
> a good idea.  Missing or inaccurate debug data for asm code seems to be
> a common problem for other projects as well.  As Andy pointed out,
> they're doing similar things in musl [1].

Thanks for the pointer.

> So it might be useful to add
> an option to the assembler which validates that the code conforms to
> certain structural rules, and then inserts frame pointer and/or .cfi
> directives.

> That said, the kernel has much more custom features than other projects.
> There are some sneaky macros, like _ASM_EXTABLE and ALTERNATIVE, which
> hide code in various sections.  Unless we're able to somehow either stop
> using these macros or isolate them to a few places, I doubt that such a
> general purpose assembler option would work.

How does the asmvalidator handle these?

> [1] http://www.openwall.com/lists/musl/2015/05/31/5

Thanks,
Pedro Alves