Linux Virtualization - Apr 2017 - [PATCH v9 0/5] Extend virtio-balloon for fast (de)inflating & fast live migration

On Thu, Apr 13, 2017 at 01:44:11PM -0700, Matthew Wilcox
wrote:
> On Thu, Apr 13, 2017 at 05:35:03PM +0800, Wei Wang wrote:
> > 2) transfer the guest unused pages to the host so that they
> > can be skipped to migrate in live migration.
> 
> I don't understand this second bit.  You leave the pages on the free
list,
> and tell the host they're free.  What's preventing somebody else
from
> allocating them and using them for something?  Is the guest semi-frozen
> at this point with just enough of it running to ask the balloon driver
> to do things?
There's missing documentation here.

The way things actually work is host sends to guest
a request for unused pages and then write-protects all memory.

So guest isn't frozen but any changes will be detected by host.

On 04/14/2017 09:50 AM, Michael S. Tsirkin wrote:
> On Thu, Apr 13, 2017 at 01:44:11PM -0700, Matthew Wilcox wrote:
>> On Thu, Apr 13, 2017 at 05:35:03PM +0800, Wei Wang wrote:
>>> 2) transfer the guest unused pages to the host so that they
>>> can be skipped to migrate in live migration.
>> I don't understand this second bit.  You leave the pages on the
free list,
>> and tell the host they're free.  What's preventing somebody
else from
>> allocating them and using them for something?  Is the guest semi-frozen
>> at this point with just enough of it running to ask the balloon driver
>> to do things?
> There's missing documentation here.
>
> The way things actually work is host sends to guest
> a request for unused pages and then write-protects all memory.
>
> So guest isn't frozen but any changes will be detected by host.
>
Probably it's better to say " transfer the info about the guest unused
pages
to the host so that the host gets a chance to skip the transfer of the 
unused
pages during live migration".

Best,
Wei

On Fri, Apr 14, 2017 at 10:28:32AM +0800, Wei Wang
wrote:
> On 04/14/2017 09:50 AM, Michael S. Tsirkin wrote:
> > On Thu, Apr 13, 2017 at 01:44:11PM -0700, Matthew Wilcox wrote:
> > > On Thu, Apr 13, 2017 at 05:35:03PM +0800, Wei Wang wrote:
> > > > 2) transfer the guest unused pages to the host so that they
> > > > can be skipped to migrate in live migration.
> > > I don't understand this second bit.  You leave the pages on
the free list,
> > > and tell the host they're free.  What's preventing
somebody else from
> > > allocating them and using them for something?  Is the guest
semi-frozen
> > > at this point with just enough of it running to ask the balloon
driver
> > > to do things?
> > There's missing documentation here.
> > 
> > The way things actually work is host sends to guest
> > a request for unused pages and then write-protects all memory.
> > 
> > So guest isn't frozen but any changes will be detected by host.
> > 
> 
> Probably it's better to say " transfer the info about the guest
unused pages
> to the host so that the host gets a chance to skip the transfer of the
> unused
> pages during live migration".
> 
> Best,
> Wei
IMHO this would not be helpful.
Most people don't know how does migration work, even if they did
this isn't tied to migration in any way.
It just makes people go "oh it's some virtualization mumbo jumbo".
We want people to be able to review and for that
interfaces need to be separate from the implementation.

IOW we must document what the interface promises not how it's used.


The promise is that pages have been unused at some time between when
host sent command and when guest completed it.  Host uses that by
tracking memory changes and then discarding changes made to pages
it gets from guest before it sent the command.

Say that and drop all mention of transfer, migration etc.

-- 
MST

On Fri, Apr 14, 2017 at 04:50:48AM +0300, Michael S. Tsirkin
wrote:
> On Thu, Apr 13, 2017 at 01:44:11PM -0700, Matthew Wilcox wrote:
> > On Thu, Apr 13, 2017 at 05:35:03PM +0800, Wei Wang wrote:
> > > 2) transfer the guest unused pages to the host so that they
> > > can be skipped to migrate in live migration.
> > 
> > I don't understand this second bit.  You leave the pages on the
free list,
> > and tell the host they're free.  What's preventing somebody
else from
> > allocating them and using them for something?  Is the guest
semi-frozen
> > at this point with just enough of it running to ask the balloon driver
> > to do things?
> 
> There's missing documentation here.
> 
> The way things actually work is host sends to guest
> a request for unused pages and then write-protects all memory.
... hopefully you mean "write protects all memory, then sends a request
for unused pages", otherwise there's a race condition.

And I see the utility of this, but does this functionality belong in
the balloon driver?  It seems like it's something you might want even
if you don't have the balloon driver loaded.  Or something you might
not want if you do have the balloon driver loaded.

On Fri, Apr 14, 2017 at 02:47:40AM -0700, Matthew Wilcox
wrote:
> On Fri, Apr 14, 2017 at 04:50:48AM +0300, Michael S. Tsirkin wrote:
> > On Thu, Apr 13, 2017 at 01:44:11PM -0700, Matthew Wilcox wrote:
> > > On Thu, Apr 13, 2017 at 05:35:03PM +0800, Wei Wang wrote:
> > > > 2) transfer the guest unused pages to the host so that they
> > > > can be skipped to migrate in live migration.
> > > 
> > > I don't understand this second bit.  You leave the pages on
the free list,
> > > and tell the host they're free.  What's preventing
somebody else from
> > > allocating them and using them for something?  Is the guest
semi-frozen
> > > at this point with just enough of it running to ask the balloon
driver
> > > to do things?
> > 
> > There's missing documentation here.
> > 
> > The way things actually work is host sends to guest
> > a request for unused pages and then write-protects all memory.
> 
> ... hopefully you mean "write protects all memory, then sends a
request
> for unused pages", otherwise there's a race condition.
Exactly.
> And I see the utility of this, but does this functionality belong in
> the balloon driver?
We have historically put all kind of memory-related functionality in the
balloon device. Consider for example memory statistics - seems related
conceptually. See patches 1-2: the new mechanism for reporting lists of
pages seems to be benefitial for both which seems to indicate using the
balloon for this is a good idea.
> It seems like it's something you might want even if you don't have
the
> balloon driver loaded.  Or something you might not want if you do have
> the balloon driver loaded.
Most of balloon functionality is kind of loosely coupled.  Yes we could
split it up but I'm not sure what would this buy us. What do you have
in mind?

-- 
MST