Re: [PATCH v2] kasan: add memory corruption identification for software tag-based mode

[Walter Wu]

On Tue, 2019-06-11 at 13:39 +0200, Dmitry Vyukov wrote:
> I should have been asked this earlier, but: what is your use-case?
We need KASAN to help us to detect memory corruption at mobile phone. It
is powerful tool.

> Could you use CONFIG_KASAN_GENERIC instead? Why not?
> CONFIG_KASAN_GENERIC already has quarantine.
> 
We hope to use tag-based KASAN, because it consumes more less
memory(1/16) than generic KASAN(1/8), but we also hope the tag-based
KASAN report is easy read and able to identify the use-after-free or
out-of-bound.


> On Tue, Jun 11, 2019 at 1:32 PM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> >
> > On Tue, Jun 11, 2019 at 12:44 PM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
> > >
> > > On Tue, 2019-06-11 at 10:47 +0200, Dmitry Vyukov wrote:
> > > > On Tue, Jun 11, 2019 at 9:05 AM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Mon, 2019-06-10 at 13:46 +0200, Dmitry Vyukov wrote:
> > > > > > On Mon, Jun 10, 2019 at 9:28 AM Walter Wu <walter-zh.wu@xxxxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Fri, 2019-06-07 at 21:18 +0800, Dmitry Vyukov wrote:
> > > > > > > > > diff --git a/include/linux/kasan.h b/include/linux/kasan.h
> > > > > > > > > index b40ea104dd36..be0667225b58 100644
> > > > > > > > > --- a/include/linux/kasan.h
> > > > > > > > > +++ b/include/linux/kasan.h
> > > > > > > > > @@ -164,7 +164,11 @@ void kasan_cache_shutdown(struct kmem_cache *cache);
> > > > > > > > >
> > > > > > > > >  #else /* CONFIG_KASAN_GENERIC */
> > > > > > > > >
> > > > > > > > > +#ifdef CONFIG_KASAN_SW_TAGS_IDENTIFY
> > > > > > > > > +void kasan_cache_shrink(struct kmem_cache *cache);
> > > > > > > > > +#else
> > > > > > > >
> > > > > > > > Please restructure the code so that we don't duplicate this function
> > > > > > > > name 3 times in this header.
> > > > > > > >
> > > > > > > We have fixed it, Thank you for your reminder.
> > > > > > >
> > > > > > >
> > > > > > > > >  static inline void kasan_cache_shrink(struct kmem_cache *cache) {}
> > > > > > > > > +#endif
> > > > > > > > >  static inline void kasan_cache_shutdown(struct kmem_cache *cache) {}
> > > > > > > > >
> > > > > > > > >  #endif /* CONFIG_KASAN_GENERIC */
> > > > > > > > > diff --git a/lib/Kconfig.kasan b/lib/Kconfig.kasan
> > > > > > > > > index 9950b660e62d..17a4952c5eee 100644
> > > > > > > > > --- a/lib/Kconfig.kasan
> > > > > > > > > +++ b/lib/Kconfig.kasan
> > > > > > > > > @@ -134,6 +134,15 @@ config KASAN_S390_4_LEVEL_PAGING
> > > > > > > > >           to 3TB of RAM with KASan enabled). This options allows to force
> > > > > > > > >           4-level paging instead.
> > > > > > > > >
> > > > > > > > > +config KASAN_SW_TAGS_IDENTIFY
> > > > > > > > > +       bool "Enable memory corruption idenitfication"
> > > > > > > >
> > > > > > > > s/idenitfication/identification/
> > > > > > > >
> > > > > > > I should replace my glasses.
> > > > > > >
> > > > > > >
> > > > > > > > > +       depends on KASAN_SW_TAGS
> > > > > > > > > +       help
> > > > > > > > > +         Now tag-based KASAN bug report always shows invalid-access error, This
> > > > > > > > > +         options can identify it whether it is use-after-free or out-of-bound.
> > > > > > > > > +         This will make it easier for programmers to see the memory corruption
> > > > > > > > > +         problem.
> > > > > > > >
> > > > > > > > This description looks like a change description, i.e. it describes
> > > > > > > > the current behavior and how it changes. I think code comments should
> > > > > > > > not have such, they should describe the current state of the things.
> > > > > > > > It should also mention the trade-off, otherwise it raises reasonable
> > > > > > > > questions like "why it's not enabled by default?" and "why do I ever
> > > > > > > > want to not enable it?".
> > > > > > > > I would do something like:
> > > > > > > >
> > > > > > > > This option enables best-effort identification of bug type
> > > > > > > > (use-after-free or out-of-bounds)
> > > > > > > > at the cost of increased memory consumption for object quarantine.
> > > > > > > >
> > > > > > > I totally agree with your comments. Would you think we should try to add the cost?
> > > > > > > It may be that it consumes about 1/128th of available memory at full quarantine usage rate.
> > > > > >
> > > > > > Hi,
> > > > > >
> > > > > > I don't understand the question. We should not add costs if not
> > > > > > necessary. Or you mean why we should add _docs_ regarding the cost? Or
> > > > > > what?
> > > > > >
> > > > > I mean the description of option. Should it add the description for
> > > > > memory costs. I see KASAN_SW_TAGS and KASAN_GENERIC options to show the
> > > > > memory costs. So We originally think it is possible to add the
> > > > > description, if users want to enable it, maybe they want to know its
> > > > > memory costs.
> > > > >
> > > > > If you think it is not necessary, we will not add it.
> > > >
> > > > Full description of memory costs for normal KASAN mode and
> > > > KASAN_SW_TAGS should probably go into
> > > > Documentation/dev-tools/kasan.rst rather then into config description
> > > > because it may be too lengthy.
> > > >
> > > Thanks your reminder.
> > >
> > > > I mentioned memory costs for this config because otherwise it's
> > > > unclear why would one ever want to _not_ enable this option. If it
> > > > would only have positive effects, then it should be enabled all the
> > > > time and should not be a config option at all.
> > >
> > > Sorry, I don't get your full meaning.
> > > You think not to add the memory costs into the description of config ?
> > > or need to add it? or make it not be a config option(default enabled)?
> >
> > Yes, I think we need to include mention of additional cost into _this_
> > new config.