Re: [RFC PATCH] UML: add support for KASAN under x86_64

[Dmitry Vyukov]

On Fri, Jan 17, 2020 at 11:03 AM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
>
> On Fri, Jan 17, 2020 at 10:59 AM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> >
> > On Thu, Jan 16, 2020 at 10:39 PM Patricia Alfonso
> > <trishalfonso@xxxxxxxxxx> wrote:
> > >
> > > On Thu, Jan 16, 2020 at 1:23 AM Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> > > >
> > > > On Thu, Jan 16, 2020 at 10:20 AM Johannes Berg
> > > > <johannes@xxxxxxxxxxxxxxxx> wrote:
> > > > >
> > > > > On Thu, 2020-01-16 at 10:18 +0100, Dmitry Vyukov wrote:
> > > > > >
> > > > > > Looking at this problem and at the number of KASAN_SANITIZE := n in
> > > > > > Makefiles (some of which are pretty sad, e.g. ignoring string.c,
> > > > > > kstrtox.c, vsprintf.c -- that's where the bugs are!), I think we
> > > > > > initialize KASAN too late. I think we need to do roughly what we do in
> > > > > > user-space asan (because it is user-space asan!). Constructors run
> > > > > > before main and it's really good, we need to initialize KASAN from
> > > > > > these constructors. Or if that's not enough in all cases, also add own
> > > > > > constructor/.preinit array entry to initialize as early as possible.
> > > > >
> > >
> > > I am not too happy with the number of KASAN_SANITIZE := n's either.
> > > This sounds like a good idea. Let me look into it; I am not familiar
> > > with constructors or .preint array.
> > >
> > > > > We even control the linker in this case, so we can put something into
> > > > > the .preinit array *first*.
> > > >
> > > > Even better! If we can reliably put something before constructors, we
> > > > don't even need lazy init in constructors.
> > > >
> > > > > > All we need to do is to call mmap syscall, there is really no
> > > > > > dependencies on anything kernel-related.
> > > > >
> > > > > OK. I wasn't really familiar with those details.
> > > > >
> > > > > > This should resolve the problem with constructors (after they
> > > > > > initialize KASAN, they can proceed to do anything they need) and it
> > > > > > should get rid of most KASAN_SANITIZE (in particular, all of
> > > > > > lib/Makefile and kernel/Makefile) and should fix stack instrumentation
> > > > > > (in case it does not work now). The only tiny bit we should not
> > > > > > instrument is the path from constructor up to mmap call.
> > >
> > > This sounds like a great solution. I am getting this KASAN report:
> > > "BUG: KASAN: stack-out-of-bounds in syscall_stub_data+0x2a5/0x2c7",
> > > which is probably because of this stack instrumentation problem you
> > > point out.
> >
> > [reposting to the list]
> >
> > If that part of the code I mentioned is instrumented, manifestation
> > would be different -- stack instrumentation will try to access shadow,
> > shadow is not mapped yet, so it would crash on the shadow access.
> >
> > What you are seeing looks like, well, a kernel bug where it does a bad
> > stack access. Maybe it's KASAN actually _working_? :)
>
> Though, stack instrumentation may have issues with longjmp-like things.
> I would suggest first turning off stack instrumentation and getting
> that work. Solving problems one-by-one is always easier.
> If you need help debugging this, please post more info: patch, what
> you are doing, full kernel output (preferably from start, if it's not
> too lengthy).

I see syscall_stub_data does some weird things with stack (stack
copy?). Maybe we just need to ignore accesses there: individual
accesses, or whole function/file.