Re: reliable reproducer, was Re: core dump analysis

Hi Finn,

Am 20.04.2023 um 17:17 schrieb Finn Thain:
> On Thu, 20 Apr 2023, Michael Schmitz wrote:
>
>>>
>>> As with dash, the corruption lies the page boundary.
>>
>> Hence implies a page fault handled at the page boundary.
>>
>> Can you try and fault in as many of these stack pages as possible, ahead
>> of filling the stack? (Depending on how much RAM you have ...). Maybe we
>> would need to lock those pages into memory? Just to show that with no
>> page faults (but still signals) there is no corruption?
>>
>
> I modified the test program to execute rec() to full depth with no
> forking, then do it again with forking.
>
> root@(none):/root# while ./stack-test 5000 ; do : ; done
> starting recursion
> done.
> starting recursion with fork
> done.
> starting recursion
> done.
> starting recursion with fork
> Illegal instruction
> root@(none):/root#
>
> I can't get this to crash during the first descent. The second descent
> always crashes, given sufficient depth:
>
> root@(none):/root# while ./stack-test 50000 ; do : ; done
> starting recursion
> done.
> starting recursion with fork
> Illegal instruction
>
> So all the stack pages would have been faulted in well before the failure
> shows up. It appears to be the signal that's the problem and not the page
> fault. That's not surprising considering the PC in the signal frame in the
> dash crash was a MOVEM saving registers onto the stack.

Well. without locking the faulted in pages in memory we can't be sure 
they were not swapped back out. Unless I misunderstand what's involved 
in that ...

In my tests, increasing the depth does not cause a monotonous increase 
in fault probability. 16k depth only has four crashes, 8k had nine. I'll 
stick with 200000 for now.

Best try and narrow down how long this bug has been present - my kernel 
builds on the current system go back a little over ten years.

Cheers,

	Michael


>
> It's worth noting that the test program never crashes with a corrupted
> return address. Random corruption would have clobbered that address about
> 10% of the time, since the entire rec() stack frame is 9 long words. So it
> must be that a MOVEM went awry when a signal got delivered.
>