Re: [RFC PATCH 00/10] x86: undwarf unwinder

From: Ingo Molnar
Date: Thu Jun 01 2017 - 09:50:22 EST



* Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:

> On Thu, Jun 01, 2017 at 08:08:24AM +0200, Ingo Molnar wrote:
> >
> > * Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> >
> > > Here's the contents of the undwarf.txt file which explains the 'why' in
> > > more detail:
> >
> > Ok, so the code quality looks pretty convincing to me - the new core 'undwarf'
> > unwinder code is a _lot_ more readable than any of the Dwarf based attempts
> > before.
> >
> > That we control the debug info generation at build time is icing on the cake to
> > me.
> >
> > One thing I'd like to see on the list of benefits side of the equation is a size
> > comparison of kernel .text, with frame pointers vs. undwarf, on 64-bit kernels.
>
> Ok, will do a text size comparison. The only difficulty I encountered
> there is that the 'size' tool considers the .undwarf section to be text
> for some reason. So the "text" size grew considerably :-)

One trick I sometimes use is to only size some of the key builtin.o files.

> > Being able to generate more optimal code in the hottest code paths of the kernel
> > is the _real_, primary upstream kernel benefit of a different debuginfo method -
> > which has to be weighed against the pain of introducing a new unwinder. But this
> > submission does not talk about that aspect at all, which should be fixed I think.
>
> Actually I devoted an entire one-sentence paragraph to performance in
> the documentation:
>
> The simpler debuginfo format also enables the unwinder to be relatively
> fast, which is important for perf and lockdep.
>
> But I'll try to highlight that a little more.

That's not what I meant! The speedup comes from (hopefully) being able to disable
CONFIG_FRAME_POINTER, which:

- creates simpler/faster function prologues and epilogues - no managing of RBP
needed

- gives one more generic purpose register to work from. This matters less on
64-bit kernels but it's a small effect.

I've seen numbers of 1-2% of instruction count reduction in common kernel
workloads, which would be pretty significant on well cached workloads.

Thanks,

Ingo