[Intel-gfx] [PATCH] drm/i915: Disable kmem_caches when KASAN is enabled

[Chris Wilson]

On Wed, Mar 15, 2017 at 11:46:24AM +0200, Mika Kuoppala wrote:
> Chris Wilson <chris at chris-wilson.co.uk> writes:
> 
> > kasan is very good at detecting use-after-free. However, our requests
> > are allocated from a rcu-typesafe slab that is important for performance
> > but makes kasan less effective. When enabling kasan we are intentionally
> > looking for memory errors, disable the use of our caches to improve the
> > likelihood of kasan spotting a bug.
> >
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  drivers/gpu/drm/i915/Kconfig.debug      | 13 +++++++++++++
> >  drivers/gpu/drm/i915/i915_gem.c         | 11 +++++++++--
> >  drivers/gpu/drm/i915/i915_gem_request.c | 25 ++++++++++++++++++++-----
> >  drivers/gpu/drm/i915/i915_vma.c         | 15 ++++++++++++---
> >  4 files changed, 54 insertions(+), 10 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/i915/Kconfig.debug b/drivers/gpu/drm/i915/Kconfig.debug
> > index e091809a9a9e..bd8e90e4dfb9 100644
> > --- a/drivers/gpu/drm/i915/Kconfig.debug
> > +++ b/drivers/gpu/drm/i915/Kconfig.debug
> > @@ -48,6 +48,19 @@ config DRM_I915_DEBUG_GEM
> >  
> >            If in doubt, say "N".
> >  
> > +config DRM_I915_DEBUG_KASAN
> > +        bool "Insert extra checks when using KASAN"
> > +        default n
> > +        depends on DRM_I915_WERROR
> > +        depends on KASAN
> > +        help
> > +	  Turns off use of kmem_caches to improve KASAN error detection,
> > +          and inserts extra santiy checks.
> 
> s/santiy/sanity.
> 
> Would i915.kmem_cache=0/1 be too costly?

More of the principle. Having module options means somebody will enable
it. Hiding it away underneath a bevy of expert config options makes it
much less likely. CI basically have to do two runs for each option -- if
we are using kasan, we want improved coverage, but we also need to test
that the default configuration works. It's probably best just to teach
kmem_cache itself to behave differently if it helps. I was just thinking
fine-grained swapping of behaviour would be benificial.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre