Re: [PATCH v13 0/5] UFFD write-tracking migration/snapshots

[David Hildenbrand]

> Am 11.02.2021 um 22:05 schrieb Peter Xu <peterx@redhat.com>:
> 
> On Thu, Feb 11, 2021 at 09:44:07PM +0100, David Hildenbrand wrote:
>> Let‘s take a look at init-on-free.
>> 
>> The guest zeroes a page and puts it onto a buddy freelist. Free page 
>> reporting code takes it off that list and reports it to the hypervisor. The 
>> hypervisor discards the physical page and tells the guest he‘s done 
>> processing the page. The guest re-places the page onto the free page list.
>> 
>> From that point on, the page can be re-allocated inside the guest and is 
>> assumed to be zero. On access, a fresh (zeroed) page is populated by the 
>> hypervisor. The guest won‘t re-zero the page, as it has the guarantee (from 
>> free page reporting) that the page remained zero.
>> 
>> Write-protecting the unpopulated page won‘t work as discussed.
> 
> IMHO no matter if it's init_on_alloc or init_on_free or both, as long as it's
> inited in some way then it means the guest OS wrote to it.  Then wr-protect
> will work..

The issue is when the discard happened before starting the snapshot. 
Write-protection won‘t work and the zeroed content won‘t be retained in the 
snapshot.

> 
> MADV_DONTNEED during live snapshot seems to be a separate topic as you
> mentioned in the other thread.  For that, I agree we'd better simply let
> virtio_balloon_inhibited() to return true for live snapshot too just like
> postcopy.

Yes, but other issue.

> 
> Thanks,
> 
> -- 
> Peter Xu
>