Re: [PATCH uq/master -v2 2/2] KVM, MCE, unpoison memory address across reboot

[Jan Kiszka <jan.kiszka@web.de>]

[-- Attachment #1: Type: text/plain, Size: 1954 bytes --]

On 2011-02-10 01:27, Huang Ying wrote:
> On Wed, 2011-02-09 at 16:00 +0800, Jan Kiszka wrote:
>> On 2011-02-09 04:00, Huang Ying wrote:
>>> In Linux kernel HWPoison processing implementation, the virtual
>>> address in processes mapping the error physical memory page is marked
>>> as HWPoison.  So that, the further accessing to the virtual
>>> address will kill corresponding processes with SIGBUS.
>>>
>>> If the error physical memory page is used by a KVM guest, the SIGBUS
>>> will be sent to QEMU, and QEMU will simulate a MCE to report that
>>> memory error to the guest OS.  If the guest OS can not recover from
>>> the error (for example, the page is accessed by kernel code), guest OS
>>> will reboot the system.  But because the underlying host virtual
>>> address backing the guest physical memory is still poisoned, if the
>>> guest system accesses the corresponding guest physical memory even
>>> after rebooting, the SIGBUS will still be sent to QEMU and MCE will be
>>> simulated.  That is, guest system can not recover via rebooting.
>>
>> Yeah, saw this already during my test...
>>
>>>
>>> In fact, across rebooting, the contents of guest physical memory page
>>> need not to be kept.  We can allocate a new host physical page to
>>> back the corresponding guest physical address.
>>
>> I just wondering what would be architecturally suboptimal if we simply
>> remapped on SIGBUS directly. Would save us at least the bookkeeping.
> 
> Because we can not change the content of memory silently during guest OS
> running, this may corrupts guest OS data structure and even ruins disk
> contents.  But during rebooting, all guest OS state are discarded.

I was not talking about remapping more than just the pages that became
inaccessible, just like you do now. But I guess the problem is rather
that insane guests continuing to access those pages before reboot should
also still receive MCEs.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 259 bytes --]