[Qemu-devel] Re: [PATCH 3/3] disk: don't read from disk until the guest starts

[Juan Quintela]

Kevin Wolf <address@hidden> wrote:
> Am 11.09.2010 16:04, schrieb Anthony Liguori:
>> This fixes a couple nasty problems relating to live migration.
>> 
>> 1) When dealing with shared storage with weak coherence (i.e. NFS), even if
>>    we re-read, we may end up with undesired caching.  By delaying any reads
>>    until we absolutely have to, we decrease the likelihood of any undesirable
>>    caching.
>> 
>> 2) When dealing with copy-on-read, the local storage acts as a cache.  We 
>> need
>>    to make sure to avoid any reads to avoid polluting the local cache.
>> 
>> Signed-off-by: Anthony Liguori <address@hidden>
>
> I think we should also delay even opening the image file at all to the
> latest possible point to avoid that new problems of this kind are
> introduced. Ideally, opening the image would be the very last thing we
> do before telling the migration source that we're set and it should
> close the images.
>
> Even better would be to only open the image when the source has already
> closed it (we could completely avoid the invalidation/reopen then), but
> I think you were afraid that we might lose the image on both ends.

But then we have formats like qcow2 that store the "number_of_sectors"
inside the file.

And we need the size of the disk to initialize some structures :(

About having the image opened in two different machines .....

- Having it opened in two places: makes it "interesting" for cocherence reasons.
- Not having it, makes the failure case of anything failing during
  migration ... interesting.

Real solution for "this" problem is to just send in the migration stream
the size/head/cylinder/sector ... values and not having to open the
image.

Problem here is that for qemu architecture, we have to know this values
before we call migration functions.

As far as I can see, until we are not able to create a machine from
inside the monitor (i.e. qemu initialization & machine creation are
decoupled), we are not going to be able to fix migration properly (using
that very same mechanism to create the machine during migration).

Later, Juan.