Re: [PATCH] Linux Kernel Markers 0.5 for Linux 2.6.17 (with probe management)

[Karim Yaghmour <karim@opersys.com>]

Mathieu Desnoyers wrote:
> First of all, I think that specific architecture-specific optimisations can and
> should be integrated in a more generic portable framework.

No disagreement there. If Ingo would care to comment, I think it might
be an acceptable compromise to have x86 fully use kprobes/djprobes
immediately, and the other archs could walk there at their rate.
Practically, some stuff in include/asm-i386/markers.h and
include/asm-x86_64/markers.h would contain the binary modifiable stuff
and include/asm-generic/markers.h could contain a platform-independent
fallback.

> Hrm, your comment makes me think of an interesting idea :
> 
> .align
> jump_address:
>   near jump to end
> setup_stack_address:
>   setup stack
>   call empty function
> end:
> 
> So, instead of putting nops in the target area, we fill it with a useful
> function call. Near jump being 2 bytes, it might be much easier to modify.
> If necessary, making sure the instruction is aligned would help to change it
> atomically. If we mark the jump address, the setup stack address and the end
> tag address with symbols, we can easily calculate (portably) the offset of the
> near jump to activate either the setup_stack_address or end tags.

That's another possibility. It seems more C friendly than the simple
short-jump+3bytes.

Ingo?

Karim
-- 
President  / Opersys Inc.
Embedded Linux Training and Expertise
www.opersys.com  /  1.866.677.4546