Re: perf: fuzzer KASAN unwind_get_return_address

[Josh Poimboeuf]

On Tue, Nov 15, 2016 at 02:05:50PM -0500, Vince Weaver wrote:
> On Tue, 15 Nov 2016, Peter Zijlstra wrote:
> 
> > On Tue, Nov 15, 2016 at 12:43:56PM -0500, Vince Weaver wrote:
> > > 
> > > Running on my haswell machine with the imc/uncore patch applied, the 
> > > perf_fuzzer next tripped over this issue.
> > > 
> > > [  202.034495] BAD LUCK: lost 371 message(s) from NMI context!
> > > [  202.034496] ==================================================================
> > > [  202.048327] BUG: KASAN: stack-out-of-bounds in unwind_get_return_address+0x35/0x80 at addr ffff8800cff0bd90
> > > [  202.058826] Read of size 8 by task perf_fuzzer/16254
> > > [  202.064186] page:ffffea00033fc2c0 count:1 mapcount:0 mapping:          (null) index:0x0^Ac
> > > [  202.073068] flags: 0x1ffff8000000400(reserved)
> > > [  202.077885] page dumped because: kasan: bad access detected
> > > [  202.083880] CPU: 4 PID: 16254 Comm: perf_fuzzer Not tainted 4.9.0-rc5+ #5
> > > [  202.091204] Hardware name: LENOVO 10AM000AUS/SHARKBAY, BIOS FBKT72AUS 01/26/2014
> > > [  202.099181]  ffff8800cff0b1d8^Ac ffffffff816bb796^Ac ffff8800cff0b270^Ac ffff8800cff0bd90^Ac
> > > [  202.107896]  ffff8800cff0b260^Ac ffffffff812fbe95^Ac 00007ffc9d1ab480^Ac 0000000000000000^Ac
> > > [  202.116638]  ffffffff8125117d^Ac 0000000000000092^Ac 0000000000000000^Ac ffff8800cff0b7c0^Ac
> > > [  202.125339] Call Trace:
> > > [  202.127994]  <NMI>  [<ffffffff816bb796>] dump_stack+0x63/0x8d
> > > [  202.134184]  [<ffffffff812fbe95>] kasan_report_error+0x495/0x4c0
> > > [  202.140680]  [<ffffffff8125117d>] ? perf_output_begin+0x28d/0x4c0
> > > [  202.147228]  [<ffffffff812fc319>] kasan_report+0x39/0x40
> > > [  202.152987]  [<ffffffff81095ce5>] ? unwind_get_return_address+0x35/0x80
> > > [  202.160094]  [<ffffffff812fa8fe>] __asan_load8+0x5e/0x70
> > > [  202.165859]  [<ffffffff81095ce5>] unwind_get_return_address+0x35/0x80
> > 
> > Josh, any ideas?
> 
> From what I can tell this maps to:
> 
> unsigned long unwind_get_return_address(struct unwind_state *state)
> {
>         unsigned long addr;
>         unsigned long *addr_p = unwind_get_return_address_ptr(state);
> 
>         if (unwind_done(state))
>                 return 0;
> 
> >>        addr = ftrace_graph_ret_addr(state->task, &state->graph_idx, *addr_p,
>                                      addr_p);
> 
>         return __kernel_text_address(addr) ? addr : 0;
> }

Hi Vince,

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.)

-- 
Josh