Re: [Qemu-devel] [PATCH 0/4] Add ignore-external migration capability

[Eduardo Habkost <ehabkost@redhat.com>]

On Fri, Jan 11, 2019 at 06:49:53PM +0300, Yury Kotov wrote:
> 10.01.2019, 23:12, "Dr. David Alan Gilbert" <dgilbert@redhat.com>:
> > * Yury Kotov (yury-kotov@yandex-team.ru) wrote:
> >>  Hi,
> >>
> >>  The series adds migration capability which allows to skip 'external' RAM blocks
> >>  during migration. External block is a RAMBlock which available from the outside
> >>  of current QEMU process (e.g. file in /dev/shm). It's useful for fast local
> >>  migration to update QEMU for the running guests.
> >
> > Hi Yury,
> >   There have been a few similar patch series around from people wanting
> > to do similar things.
> >   In particular Lai Jiangshan's https://lists.nongnu.org/archive/html/qemu-devel/2018-03/msg07511.html
> > and Cédric Le Goater wanted to skip regions for a different reason.
> >
> >   We merged some of Cédric's code last year so that we now
> > have the qemu_ram_is_migratable() function - and we should be reusing
> > that to skip things rather than adding a new check that we have to add
> > everywhere.
> >
> 
> I didn't see the series, so I'll check it, thanks!
> But I saw qemu_ram_is_migratable() function and corresponding patch.
> It's very close to my needs, but it works a bit different IIUC:
> 1. Not migratable blocks isn't validated (existence and size) during migration,
> 2. "Migratable" state is determined during the block creation time.
>    Such case isn't valid because of it:
>    * Source has one migratable and one not migratable RAM blocks,
>    * Target has the same (idstr) blocks, but both are not migratable.
>    Thus, target will not expect pages for not migratable blocks.
> 
> >   Also, ypu're skipping 'external' things, I think the other suggestion
> > was to skip 'shared' things (i.e. anything with share=0); skipping
> > share=on cases sounds easier to me.
> 
> I agree that introducing new term is a complication, but 'share' and 'external'
> terms have important differences (I'll describe it below).
> 
> Just to clarify:
> * 'share' means that other processes has an access to such memory,
> * 'external' means file backed memory.

If you use file backed memory with share=off, writes are not
propagated to the file (they are mapped with MAP_PRIVATE).  Would
you really want to skip file backed memory if it has share=off?

> 
> There is another use case I wanted to support (I had to write about it in
> the cover letter, sorry..):
> 1. Migrate source VM to file and kill source,
> 2. Start target VM and migrate it from file.
> In such case source VM may have memory-backend-ram with share=off, it's ok.
> 
> Thus, in the new migration capability I want to migrate memory that meets
> three conditions:
> 1. The source will not use the memory after migration ends,
> 2. The source may exit before target starts (migrate to file),
> 3. The target has an access to the memory.
> 
> I think 'external' fits them better than 'share'.
> 

In either case, defining "external" seems tricky.  A memory
region might be backed by a file on tmpfs or hugetlbfs that was
deleted, which makes the file "internal" for practical purposes.
QEMU has no way to tell if (3) is really true.

-- 
Eduardo