Re: perf: fuzzer KASAN unwind_get_return_address

[Dmitry Vyukov]

On Thu, Nov 17, 2016 at 3:36 PM, Josh Poimboeuf <jpoimboe@xxxxxxxxxx> wrote:
> On Thu, Nov 17, 2016 at 09:25:58AM -0500, Vince Weaver wrote:
>> On Thu, 17 Nov 2016, Josh Poimboeuf wrote:
>>
>> > On Thu, Nov 17, 2016 at 10:48:27AM +0100, Dmitry Vyukov wrote:
>> > > Just in case, there is currently a known KASAN false positive related
>> > > to longjmp's on GPFs. When a syscall hits GPF stack is unwound to
>> > > kernel entry point, this leaves a bunch of stray poisoned redzones on
>> > > the thread stack. They later cause false stack-out-of-bounds reports.
>> > >
>> > > But this does not seem to be the case here. Kernel is not tainted. And
>> > > shadow at the bottom of the reports looks sane.
>> > >
>> > > But if that's the case somehow, we will need to add
>> > > kasan_unpoison_remaining_stack() call before a longjmp like we did for
>> > > jprobe_return():
>> > > https://groups.google.com/d/msg/kasan-dev/Hzox58yZ4MU/TOdFoWMuBQAJ
>> >
>> > I'm pretty sure this isn't a KASAN false positive.  The unwinder does
>> > actually seem to be accessing a bad area of the stack, in the middle of
>> > a function's stack frame.
>>
>> I'm having trouble reproducing it on a few other machines I have fuzzing.
>> So there might be some kernel option contributing, I need to compare
>> .configs.
>>
>> Also the machine that easily triggers the problem I'm compiling with
>> gcc-5.4 where the machines I can't are using gcc-4.9.
>
> I believe KASAN only works with gcc 5 and later, so that would explain
> why you aren't seeing it with gcc 4.9.

Right. 4.9 has limited support for KASAN. It supports general
instrumentation, but only with CONFIG_KASAN_OUTLINE, and it does not
support stack poisoning. Which is required to detect stack OOBs.