Re: [PATCH 0/8] livepatch: klp-convert tool

[Torsten Duwe <duwe@lst.de>]

On Fri, Oct 20, 2017 at 08:24:32AM -0500, Josh Poimboeuf wrote:
> On Fri, Oct 20, 2017 at 02:44:32PM +0200, Torsten Duwe wrote:
> > On Thu, Oct 19, 2017 at 06:00:54PM +0200, Miroslav Benes wrote:
> > > On Thu, 19 Oct 2017, Josh Poimboeuf wrote:
> > > > 
> > > > Sounds nice, though I wonder what the obstacles are?
> > > 
> > > Those GCC optimizations you mentioned below and which I didn't connect to 
> > > klp-convert itself.
> > 
> > I have a bad feeling about the IPA stuff in general. An obj-based approach
> > is cool in a way that it still works, and is sure to work, if the IPA
> > assumptions that led to the optimisations still hold, but as soon as they
> > break, you're screwed big time.
> 
> Huh?  The obj-based approach (e.g., kpatch, bin-diff) inherently detects
> such changes.  Or am I misunderstanding?  If so, please elaborate.

If e.g. the old callee does not use a caller-saved reg, neither does the new
code, there is no reason to actually save it. IPA will detect this and spare
the reg save/restore. If now by any coincidence the new function needs that
register, what can you do? Patch all callers as well? You'll likely end up
with a far-too-big live patch or by coding something in assembler :-(

> > For -fpatchable-function-entries I switched
> > off IPA-RA, as especially on RISC there's _nothing_ you can do between
> > functions without at least one scratch reg. But for live patching, I'd like
> > the kernel to be compiled in the first place with 100% ABI adherence, IOW
> > all IPA optimisations turned off. Does anyone have numbers on the performance
> > impact?
> 
> I agree that would be the best option.

I'm glad to hear this!-)

> > For example, if the patched functions and the fixes meet its criteria, you
> > could use CMBC (http://www.cprover.org/cbmc/) to _prove_ that the live patch
> > changes exactly what you claim to, and nothing else.
> 
> Can it also prove that the patch is applied in a safe manner?

I assume this strongly depends on the test cases you build around it.
OTOH the more code you include in the test, the more likely you'll violate
the cbmc preconditions. So I guess the simple answer is "no". But this might
change in the future.

	Torsten