Re: perf: fuzzer KASAN unwind_get_return_address

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Wed, Nov 16, 2016 at 03:49:43PM +0100, Peter Zijlstra wrote:
> On Wed, Nov 16, 2016 at 08:37:46AM -0600, Josh Poimboeuf wrote:
> > On Wed, Nov 16, 2016 at 02:03:37PM +0100, Peter Zijlstra wrote:
> > > On Tue, Nov 15, 2016 at 02:57:48PM -0600, Josh Poimboeuf wrote:
> > > > Would you mind posting a disassembly of unwind_get_return_address()?
> > > > Any idea how recreatable it is?  (In particular I'd be interested in
> > > > seeing this dump with the latest unwinder improvements in the -tip tree,
> > > > which dump the pt_regs associated with an interrupt.)
> > > 
> > > Fairly reproducable it seems, doesn't seem to include pt_regs dumps
> > > though :/
> > > 
> > > tip/master as of this morning.
> > 
> > Thanks.  This is actually a different issue than the one reported by
> > Vince.  In this case FRAME_POINTER is disabled, so it uses the "guess"
> > unwinder which scans every address on the stack, looking for text
> > addresses.  So the kasan errors are expected.
> > 
> > (The missing pt_regs are also expected: the guess unwinder doesn't show
> > them.)
> > 
> > I'll work up a patch to fix this.  I still have no idea what's causing
> > Vince's bug in the frame pointer unwinder.
> 
> Hurm,.. by the number of '?' entries in Vince's backtrace I was assuming
> it was without frame pointers.

When frame pointers are disabled, *all* the addresses are prefixed with
'?'.

When frame pointers are enabled, and there are a lot of '?' addresses,
it usually means the containing functions reserved a lot of stack space
and the printed addresses are mostly leftovers from previous runs.

> Let me enable those and run again, it didn't insta-trigger like it does
> without.

Thanks!

-- 
Josh