Re: [PATCH v14 0/6] Introduce the STACKLEAK feature and a test for it

[Ingo Molnar <mingo@kernel.org>]

* Alexander Popov <alex.popov@linux.com> wrote:

> On 16.07.2018 01:44, Ingo Molnar wrote:
> > 
> > * Kees Cook <keescook@chromium.org> wrote:
> > 
> >> On Thu, Jul 12, 2018 at 2:22 PM, Alexander Popov <alex.popov@linux.com> wrote:
> >>> On 12.07.2018 23:50, Ingo Molnar wrote:
> >>>> Let's make sure informed users have an easy runtime way out
> >>>> from the worst of the overhead that doesn't involve "recompile your distro
> >>>> kernel". Also, make it easier to measure and fingerpoint the overhead...
> >>>
> >>> Would you like the following solution?
> >>>
> >>> I'll create the CONFIG_STACKLEAK_RUNTIME_DISABLE config option, which would be
> >>> similar to CONFIG_SECURITY_SELINUX_DISABLE. That option will provide a sysctl
> >>> switch disabling stackleak_erase(), which provides the most of performance penalty.
> >>
> >> I don't think CONFIG/sysctl is the way to go. I'd recommend making it
> >> a boot arg and using a static key, similar to what's happening to
> >> hardened_usercopy:
> > 
> > Why no sysctl? It's a PITA to reboot systems just to turn a stupid knob off.
> > 
> > Also, it's _much_ easier to measure performance impact when there's a sysctl.
> 
> Yes, you are right.
> 
> But I looked carefully and now see the troubles which sysctl would bring us.
> Each 'task_struct' has 'lowest_stack', which must be initialized before enabling
> STACKLEAK. So runtime enabling via sysctl is not plain and may bring race
> conditions.

Firstly, a sysctl could still allow it to be *disabled*, once - which is the most 
important usecase of the sysctl anyway.

Secondly, in the first iteration this could be kept included unconditionally:

	current->lowest_stack = current_top_of_stack() - THREAD_SIZE/64;

... which would keep it initialized and wouldn't be racy, right?

I.e. only the most expensive part of the function, the scanning, would be turned 
off via the sysctl. I submit that this will avoid all measurable aspects of the 1% 
kbuild performance overhead.

A more involved approach can be done in the future if warranted, and the feature 
could be disabled/enabled more thoroughly - but the runtime sysctl would be 
acceptable for me for now, as a first iteration.

> On the other hand, a boot param + static key that can only disable STACKLEAK 
> during __init are much more robust. I'm preparing the patch and performance 
> measurements.

A boot parameter does *not* address the concern I outlined. Why force admins 
reboot a box just to disable something that causes overhead?

Thanks,

	Ingo