Re: KASAN help, please (Re: [PATCH 00/16] Entry stuff, in decent shape now)

[Andy Lutomirski]

> On Nov 21, 2017, at 2:09 AM, Dmitry Vyukov <dvyukov@xxxxxxxxxx> wrote:
> 
>> On Mon, Nov 20, 2017 at 10:44 PM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>> On Mon, Nov 20, 2017 at 9:07 AM, Andy Lutomirski <luto@xxxxxxxxxx> wrote:
>>> This sets up stack switching, including for SYSCALL.  I think it's
>>> in decent shape.
>>> 
>>> Known issues:
>>> - KASAN is likely to be busted.  This could be fixed either by teaching
>>>   KASAN that cpu_entry_area contains valid stacks (I have no clue how
>>>   to go about doing this) or by rigging up the IST entry code to switch
>>>   RSP to point to the direct-mapped copy of the stacks before calling
>>>   into non-KASAN-excluded C code.
>>> 
>> 
>> I tried to fix the KASAN issue, and I'm doing something wrong.  I'm
>> building this tree:
>> 
>> https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/entry_stack&id=8319677bd04a1ab291ca71fe1da7aa023306e4a9
>> 
>> for 64 bits with KASAN on.  The relevant commit is:
>> 
>> https://git.kernel.org/pub/scm/linux/kernel/git/luto/linux.git/commit/?h=x86/entry_stack&id=a4bdb48c3469708b6b51e5ab90d27bf0c859000c
>> 
>> If I run tools/testing/selftests/single_step_syscall_32, then the
>> kernel goes into lala land and infinite loops.  The root cause seems
>> to we're hitting do_debug with RSP pointing into the fixmap,
>> specifically in the cpu_entry_area's exception stack, with a value of
>> roughly 0xffffffffff1bd108.  The KASAN instrumentation in do_debug is
>> then getting a page fault.  I think my KASAN setup code should be
>> populating the KASAN data there and, indeed, gdb seems to be able to
>> access the faulting address.  So I'm confused.
> 
> 
> Hi,
> 
> I don't have any great insights.
> 
> You have stack instrumentation turned on, right? And the fault happens
> on stack instrumentation?
> Stack instrumentation is turned on with gcc7+ I think. And as the
> result compiler adds redzones on stack and poisons/unpoisons shadow
> for them in function prologue/epilogue.

I found the problem.  I goofed in the setup code, so I ended up with a only zero page in the shadow.  Turns out that gdb can happily write to read only memory :(

> 
> The fact that KASAN instrumentation faults, but gdb can access it
> sounds strange. KASAN instrumentation is no magic, it just does not a
> normal memory load. Please check exact faulting address. KASAN can do
> accesses with large offset from RSP.
> 
> Does the fault happen before/after kasan_early_init? Before that there
> is a different bootstrap shadow mapped by kasan_map_early_shadow.
> 
> Does the fault happen on read access or write access? Stack
> instrumentation does write into shadow, but some parts of shadow are
> mapped with a single read-only page. Can gdb write to that address?
> 
> Is it possible that the stack has overflowed? I see that we increase
> EXCEPTION_STACK_ORDER by order 1 under KASAN (from 4k page to 8k
> pages), but it may be not enough. Normal stacks are increased from 16k
> to 32k.
> 
> Last stupid question: why is it -1 here:
> FIX_CPU_ENTRY_AREA_BOTTOM = FIX_CPU_ENTRY_AREA_TOP +
> (CPU_ENTRY_AREA_PAGES * NR_CPUS) - 1,
> ?
> Say CPU_ENTRY_AREA_PAGES=1 (we need only 1 page) and NR_CPUS=1, then
> the increment will be 0, which looks wrong for any case (must be at
> least 1, right?).