Re: [PATCH] x86/suspend: fix false positive KASAN warning on suspend/resume

[Dmitry Vyukov <dvyukov@google.com>]

On Thu, Dec 1, 2016 at 6:13 PM, Josh Poimboeuf <jpoimboe@redhat.com> wrote:
> On Thu, Dec 01, 2016 at 05:51:52PM +0100, Dmitry Vyukov wrote:
>> On Thu, Dec 1, 2016 at 5:45 PM, Josh Poimboeuf <jpoimboe@redhat.com> wrote:
>> > On Thu, Dec 01, 2016 at 08:58:21AM -0600, Josh Poimboeuf wrote:
>> >> On Thu, Dec 01, 2016 at 12:05:34PM +0300, Andrey Ryabinin wrote:
>> >> >
>> >> >
>> >> > On 12/01/2016 02:10 AM, Josh Poimboeuf wrote:
>> >> > > Resuming from a suspend operation is showing a KASAN false positive
>> >> > > warning:
>> >> > >
>> >> >
>> >> > > KASAN instrumentation poisons the stack when entering a function and
>> >> > > unpoisons it when exiting the function.  However, in the suspend path,
>> >> > > some functions never return, so their stack never gets unpoisoned,
>> >> > > resulting in stale KASAN shadow data which can cause false positive
>> >> > > warnings like the one above.
>> >> > >
>> >> > > Reported-by: Scott Bauer <scott.bauer@intel.com>
>> >> > > Tested-by: Scott Bauer <scott.bauer@intel.com>
>> >> > > Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com>
>> >> > > ---
>> >> > >  arch/x86/kernel/acpi/sleep.c | 3 +++
>> >> > >  include/linux/kasan.h        | 7 +++++++
>> >> > >  2 files changed, 10 insertions(+)
>> >> > >
>> >> > > diff --git a/arch/x86/kernel/acpi/sleep.c b/arch/x86/kernel/acpi/sleep.c
>> >> > > index 4858733..62bd046 100644
>> >> > > --- a/arch/x86/kernel/acpi/sleep.c
>> >> > > +++ b/arch/x86/kernel/acpi/sleep.c
>> >> > > @@ -115,6 +115,9 @@ int x86_acpi_suspend_lowlevel(void)
>> >> > >   pause_graph_tracing();
>> >> > >   do_suspend_lowlevel();
>> >> > >   unpause_graph_tracing();
>> >> > > +
>> >> > > + kasan_unpoison_stack_below_sp();
>> >> > > +
>> >> >
>> >> > I think this might be too late. We may hit stale poison in the first C function called
>> >> > after resume (restore_processor_state()). Thus the shadow must be unpoisoned prior such call,
>> >> > i.e. somewhere in do_suspend_lowlevel() after .Lresume_point.
>> >>
>> >> Yeah, I think you're right.  Will spin a v2.
>> >
>> > So I tried calling kasan_unpoison_task_stack_below() from
>> > do_suspend_lowlevel(), but it hung on the resume.  Presumably because
>> > restore_processor_state() does some important setup which would be
>> > needed before calling into kasan_unpoison_task_stack_below().  For
>> > example, setting up the gs register.  So it's a bit of a catch-22.
>> >
>> > It could probably be fixed properly by rewriting do_suspend_lowlevel()
>> > to call restore_processor_state() with the temporary stack before
>> > switching to the original stack and doing the unpoison.
>> >
>> > (And there are some other issues with do_suspend_lowlevel() and I'd love
>> > to try taking a scalpel to it.  But I have too many knives in the air
>> > already to want to try to attempt that right now...)
>> >
>> > Unless somebody else wants to take a stab at it, my original patch is
>> > probably good enough for now, since restore_processor_state() doesn't
>> > seem to be triggering any KASAN warnings.
>>
>> restore_processor_state/__restore_processor_state does not seem to
>> have any local variables, so KASAN does not do any stack checks there.
>
> Actually, looking at the object code, it uses a lot of stack space and
> has several calls to __asan_report_load*() functions.  Probably due to
> inlining of other functions which have stack variables.

That can be loads of heap variables (or other non-stack data). KASAN
will emit these checks for lots of loads, but they don't necessary go
to stack.


>> We could disable KASAN instrumentation of the file, or of particular
>> functions.
>
> I don't think that would be sufficient unless it were disabled for
> __restore_processor_state() and all the functions it calls (and the
> functions they call, etc), which wouldn't necessarily be
> straightforward.
>
>> Or we could call kasan_unpoison_shadow() on the stack range
>> before switching to it.
>
> I tried that already, but it hung because restore_processor_state()
> hadn't been called yet (the catch-22 I mentioned aboved).

Ah, I see, we just can't execute normal C code at that point...