Re: [PATCH v1 0/3] migration: Fixes to the 'background-snapshot' code

From: Andrey Gruzdev <andrey.gruzdev@virtuozzo.com>
To: Peter Xu <peterx@redhat.com>
Cc: qemu-devel@nongnu.org, Den Lunev <den@openvz.org>,
	Eric Blake <eblake@redhat.com>,
	Paolo Bonzini <pbonzini@redhat.com>,
	Juan Quintela <quintela@redhat.com>,
	"Dr . David Alan Gilbert" <dgilbert@redhat.com>,
	Markus Armbruster <armbru@redhat.com>,
	David Hildenbrand <david@redhat.com>
Subject: Re: [PATCH v1 0/3] migration: Fixes to the 'background-snapshot' code
Date: Thu, 25 Mar 2021 12:51:28 +0300	[thread overview]
Message-ID: <a4c76b56-a57c-8e4a-0a2c-78a6b9136c1d@virtuozzo.com> (raw)
In-Reply-To: <20210324154158.GC219069@xz-x1>

On 24.03.2021 18:41, Peter Xu wrote:
> On Wed, Mar 24, 2021 at 11:09:27AM +0300, Andrey Gruzdev wrote:
>>> I'm also looking into introducing UFFD_FEATURE_WP_UNALLOCATED so as to
>>> wr-protect page holes too for a uffd-wp region when the feature bit is set.
>>> With that feature we should be able to avoid pre-fault as what we do in the
>>> last patch of this series.  However even if that can work out, we'll still need
>>> this for old kernel anyways.
>> I'm curious this new feature is based on adding wr-protection at the level of VMAs,
>> so we won't miss write faults for missing pages?
> I think we can do it with multiple ways.
>
> The most efficient one would be wr-protect the range during uffd-wp
> registration, so as you said it'll be per-vma attribute.  However that'll
> change the general semantics of uffd-wp as normally we need registration and
> explicit wr-protect.  Then it'll still be pte-based for faulted in pages (the
> ones we wr-protected during registration will still be), however for the rest
> it'll become vma-based.  It's indeed a bit confusing.
>
> The other way is we can fault in zero page during UFFDIO_WRITEPROTECT.  However
> that's less efficient, since it's close to pre-fault on read but it's just
> slightly more cleaner than doing it in userspace.  When I rethink about this it
> may not worth it to do in kernel if userspace can achieve things similar.
>
> So let's stick with current solution; that idea may need more thoughts..
>
> Thanks,
>
Agree, let's stick with current solution. For the future I think having 
a registration
flag like WP_MISSING to induce per-vma wr-protection is not a bad choice.

The reason is that usage of UFFDIO_WRITEPROTECT ioctl is often 
asymmetrical; we usually
write-protect the whole registration range but un-protect by small chunks.

So if we stay with current current symmetric protect/un-protect API but 
add the registration
flag to handle protection for unpopulated pages - that may be worth to do.

-- 
Andrey Gruzdev, Principal Engineer
Virtuozzo GmbH  +7-903-247-6397
                 virtuzzo.com