This is the mail archive of the
systemtap@sourceware.org
mailing list for the systemtap project.
Re: [RFC][ PATCH -tip v2 0/7] kprobes: Kprobes jump optimization support
- From: Andi Kleen <andi at firstfloor dot org>
- To: Masami Hiramatsu <mhiramat at redhat dot com>
- Cc: Ingo Molnar <mingo at elte dot hu>, Ananth N Mavinakayanahalli <ananth at in dot ibm dot com>, lkml <linux-kernel at vger dot kernel dot org>, "H. Peter Anvin" <hpa at zytor dot com>, Frederic Weisbecker <fweisbec at gmail dot com>, Jim Keniston <jkenisto at us dot ibm dot com>, Srikar Dronamraju <srikar at linux dot vnet dot ibm dot com>, Christoph Hellwig <hch at infradead dot org>, Steven Rostedt <rostedt at goodmis dot org>, Anders Kaseorg <andersk at ksplice dot com>, Tim Abbott <tabbott at ksplice dot com>, systemtap <systemtap at sources dot redhat dot com>, DLE <dle-develop at lists dot sourceforge dot net>
- Date: Tue, 23 Jun 2009 18:34:51 +0200
- Subject: Re: [RFC][ PATCH -tip v2 0/7] kprobes: Kprobes jump optimization support
- References: <20090622212255.5384.53732.stgit@localhost.localdomain> <87vdmn179n.fsf@basil.nowhere.org> <4A40DDAD.6020202@redhat.com>
Masami Hiramatsu <mhiramat@redhat.com> writes:
> Hi Andi,
>
> Andi Kleen wrote:
>> Masami Hiramatsu <mhiramat@redhat.com> writes:
>>> The gcc's crossjumping unifies equivalent code by inserting indirect
>>> jumps which jump into other function body. It is hard to know to where
>>> these jumps jump, so I decided to disable it when setting
>>> CONFIG_OPTPROBES=y.
>>
>> That sounds quite bad. Tail call optimization is an important optimization
>> that especially on kernel style code (lots of indirect pointers
>> and sometimes deep call chains) is very useful. It would be quite
>> sad if production kernels would lose that optimization.
>
> I think the crossjumping is not the tail call optimization,
> http://gcc.gnu.org/onlinedocs/gcc-3.4.6/gccint/Passes.html
Statement didn't make sense then. The RTL crossjump pass you're referring
AFAIK does not jump into other functions, it only optimizes jumps
inside a function (unless you're talking about inlines)
>> Also tail calls in C should always jump directly to another function,
>> so they shouldn't be particularly complex to manage.
>
> Tail call jumps directly into the head of another function,
> not the middle. Thus it is safe.
cross jumping does neither.
>
>>> I also decided not to optimize probes when it is in functions which
>>> will cause exceptions, because the exception in the kernel will jump
>>> to a fixup code and the fixup code jumps back to the middle of the
>>> same function body.
>>
>> Note that not only exceptions do that, there are a few other cases
>> where jumps in and out of out of line sections happen. You might
>> need a more general mechanism to detect this.
>
> As far as I can see (under arch/x86), Almost all fixup entries are
> defined with ex_table entries, and others jump to the head of
> symbols(or functions). The jumps which jump into the middle of
> some functions are what I need to find, and, as far as I know,
> those fixup jumps are used with exception tables. Of course,
> I might miss some fixup codes, in that case, please let me know:-)
One case for example are out of line sections generated by gcc itself
with the right options.
-andi
--
ak@linux.intel.com -- Speaking for myself only.