Re: Stock 2.6.27 kernel hanging on WP test

[Jeremy Fitzhardinge <jeremy@goop.org>]

On 10/08/09 15:56, Jeremy Fitzhardinge wrote:
> On 10/08/09 15:32, Jeremy Fitzhardinge wrote:
>   
>> When I boot Fedora 9 with a stock 2.7.26.25 32-bit-PAE kernel, it's
>> hanging forever in the WP test.  I'm running current xen-unstable
>> 1d7221667204.
>>
>> Another kernel (I think non-PAE this time) complained that the WP test
>> failed (ie, supervisor was allowed to write to a RO page without fault).
>>
>> The previous version of Xen I tried (only a couple of days ago), crashed
>> with "(XEN) Assertion 'v->arch.paging.mode' failed at multi.c:4069" when
>> starting the same domain, but I'm not sure if it got to the same point.
>>   
>>     
> Oh, I guess I should have pointed out this is an HVM domain.  PV is no
> problem.
>
> A 64-bit HVM Linux guest seems to die very quickly, but I'm not sure
> where its dying (I don't think 64-bit kernels bother testing for WP, so
> it could be another crash with the same root cause).
>   

Any thoughts about this?  When I look at the domain context with
debug-key 'd', the eip is on the write to the RO page doing the WP test,
and the VCPU is stuck in run state.  It looks like the Xen fault handler
is endlessly restarting the instruction rather than injecting a guest fault.

Presumably that path must mostly work or not much would run at all.  The
only thing that occurs to me is that some heuristic is putting Xen down
some unusual path that has a bug in it.  The fault is a bit unusual
because it's a deliberate write to a RO fixmap slot, high in the 32-bit
address space.

    J