kprobes trampoline_p registration

kprobes trampoline_p registration

[Mike Grundy]

Hi - I noticed in 2.6.17-rc6 in the i386 kprobes.c the registration of the
trampoline probe in arch_init_kprobes was removed. It looks like that pile of
assembly in kretprobe_trampoline_holder does the clean up instead of using the
regular mechanisms. Any particular reason for this? Should the other arches
consider a similar mechanism?
-- 
Thanks
Mike

=========================================
Michael Grundy - grundym@us.ibm.com
Advanced Linux Response Team (ALRT)

If at first you don't succeed, call in an air strike.

Re: kprobes trampoline_p registration

[Ananth N Mavinakayanahalli]

On Mon, Jun 12, 2006 at 05:30:07AM -0400, Mike Grundy wrote:
> Hi - I noticed in 2.6.17-rc6 in the i386 kprobes.c the registration of the
> trampoline probe in arch_init_kprobes was removed. It looks like that pile of
> assembly in kretprobe_trampoline_holder does the clean up instead of using the
> regular mechanisms. Any particular reason for this? Should the other arches
> consider a similar mechanism?

Mike,

That is the result of Masami's kprobe booster patchset and is applicable
to i386 (there is a patch floating around for ia64 too).

Other archs that don't have a booster implementation (powerpc for
instance) still use the older trampoline probe mechanism for function
return probes. Until such time the other archs have the booster
implemented, they'll need the trampoline probe registered as part of
arch_init_kprobes().

Ananth

Re: kprobes trampoline_p registration

[Masami Hiramatsu]

Hi,

That change is to reduce the overhead of kretprobe. The original
implementation of the kretprobe uses 2 kprobes. On i386 arch,
2nd kprobe can be replaced by assembly code which works much faster
than kprobe.
You can see how it works in detail here:
http://sources.redhat.com/ml/systemtap/2006-q1/msg00342.html

And our team is currently porting this kind of improvements to
IA64 and x86-64.

Best regards,

Mike Grundy wrote:
> Hi - I noticed in 2.6.17-rc6 in the i386 kprobes.c the registration of the
> trampoline probe in arch_init_kprobes was removed. It looks like that pile of
> assembly in kretprobe_trampoline_holder does the clean up instead of using the
> regular mechanisms. Any particular reason for this? Should the other arches
> consider a similar mechanism?


-- 
Masami HIRAMATSU
2nd Research Dept.
Hitachi, Ltd., Systems Development Laboratory
E-mail: hiramatu@sdl.hitachi.co.jp