Re: [PATCH 4/5] Xen/MCE: Abort live migration when vMCE occur

[George Dunlap <george.dunlap@eu.citrix.com>]

On 19/10/12 21:32, Liu, Jinsong wrote:
>> Wouldn't your patch 5 be sufficient to deal with this case?  It seems
>> like the broken page would get marked as such, and then get marked
>> broken on the receiving side, wouldn't it?
>>
>>   -George
> Seems no, patch 4 is to handle the case mce occur during migration --> under such case the broken page would mapped (at that time the page is a good page) and copy to target; While patch 5 is to handle the case mce occur beofre migration --> under such case the broken page would not mapped and so would not copy to target.

In the "during migration", there are actually two cases to consider:
1. The page breaks before the domain save code maps it.
2. The page breaks after the domain save code has mapped it once

Patch 5 will detect a broken page when it tries to map it, and send it 
as type "broken", without data.

So in the case of #1, it will be taken care of by patch 5 without any 
changes.

In the case of #2, it seems like we could probably modify patch 5 to 
handle it.  If we mark a page dirty, then the domain save code will try 
to send it again.  When it tries to map it, it will discover that the 
page has been marked "broken", and will send it as a "broken" page, 
without data.  As long as the domain restore code marks the 
already-received page as "broken" when it receives this message, then 
everything should work as normal.

What do you think?

  -George