[question] The source cannot recover, if the destination fails in the last round of live migration

[question] The source cannot recover, if the destination fails in the last round of live migration

[Kunkun Jiang]

Hi all,

Recently I am learning about the part of live migration.
I have a question about the last round.

When the pending_size is less than the threshold, it will enter
the last round and call migration_completion(). It will stop the
source and sent the remaining dirty pages and devices' status
information to the destination. The destination will load these
information and start the VM.

If there is an error at the destination at this time, it will exit
directly, and the source will not be able to detect the error
and recover. Because the source will not call
migration_detect_error().

Is my understanding correct?
Should the source wait the result of the last round of destination ?

Thanks,
Kunkun Jiang

Re: [question] The source cannot recover, if the destination fails in the last round of live migration

[Dr. David Alan Gilbert]

* Kunkun Jiang (jiangkunkun@huawei.com) wrote:
> Hi all,

Hi,

> Recently I am learning about the part of live migration.
> I have a question about the last round.
> 
> When the pending_size is less than the threshold, it will enter
> the last round and call migration_completion(). It will stop the
> source and sent the remaining dirty pages and devices' status
> information to the destination. The destination will load these
> information and start the VM.
> 
> If there is an error at the destination at this time, it will exit
> directly, and the source will not be able to detect the error
> and recover. Because the source will not call
> migration_detect_error().
> 
> Is my understanding correct?
> Should the source wait the result of the last round of destination ?

Try setting the 'return-path' migration capability on both the source
and destination;  I think it's that option will cause the destination to
send an OK/error at the end and the source to wait for it.

Dave

> Thanks,
> Kunkun Jiang
> 
> 
-- 
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

Re: [question] The source cannot recover, if the destination fails in the last round of live migration

[Kunkun Jiang]

Hi Dave,

On 2021/5/6 21:05, Dr. David Alan Gilbert wrote:
> * Kunkun Jiang (jiangkunkun@huawei.com) wrote:
>> Hi all,
> Hi,
>
>> Recently I am learning about the part of live migration.
>> I have a question about the last round.
>>
>> When the pending_size is less than the threshold, it will enter
>> the last round and call migration_completion(). It will stop the
>> source and sent the remaining dirty pages and devices' status
>> information to the destination. The destination will load these
>> information and start the VM.
>>
>> If there is an error at the destination at this time, it will exit
>> directly, and the source will not be able to detect the error
>> and recover. Because the source will not call
>> migration_detect_error().
>>
>> Is my understanding correct?
>> Should the source wait the result of the last round of destination ?
> Try setting the 'return-path' migration capability on both the source
> and destination;  I think it's that option will cause the destination to
> send an OK/error at the end and the source to wait for it.
Thank you for your reply!
The 'return-path' migration capability solved my question. 😁

But why not set it as the default? In my opinion, it is a basic ability
of live migration. We need it to judge whether the last round of the
destination is successful in the way of 'precopy'.

Looking forward to your reply.

Thanks,
Kunkun Jiang
> Dave
>
>> Thanks,
>> Kunkun Jiang
>>
>>

Re: [question] The source cannot recover, if the destination fails in the last round of live migration

[Peter Xu]

On Fri, May 07, 2021 at 05:46:44PM +0800, Kunkun Jiang wrote:
> Hi Dave,
> 
> On 2021/5/6 21:05, Dr. David Alan Gilbert wrote:
> > * Kunkun Jiang (jiangkunkun@huawei.com) wrote:
> > > Hi all,
> > Hi,
> > 
> > > Recently I am learning about the part of live migration.
> > > I have a question about the last round.
> > > 
> > > When the pending_size is less than the threshold, it will enter
> > > the last round and call migration_completion(). It will stop the
> > > source and sent the remaining dirty pages and devices' status
> > > information to the destination. The destination will load these
> > > information and start the VM.
> > > 
> > > If there is an error at the destination at this time, it will exit
> > > directly, and the source will not be able to detect the error
> > > and recover. Because the source will not call
> > > migration_detect_error().
> > > 
> > > Is my understanding correct?
> > > Should the source wait the result of the last round of destination ?
> > Try setting the 'return-path' migration capability on both the source
> > and destination;  I think it's that option will cause the destination to
> > send an OK/error at the end and the source to wait for it.
> Thank you for your reply!
> The 'return-path' migration capability solved my question. 😁
> 
> But why not set it as the default? In my opinion, it is a basic ability
> of live migration. We need it to judge whether the last round of the
> destination is successful in the way of 'precopy'.

I think it should be enabled as long as both sides support it; though may not
be suitable as default (at least in QEMU) so as to consider old binaries.

Thanks,

-- 
Peter Xu

Re: [question] The source cannot recover, if the destination fails in the last round of live migration

[Dr. David Alan Gilbert]

* Peter Xu (peterx@redhat.com) wrote:
> On Fri, May 07, 2021 at 05:46:44PM +0800, Kunkun Jiang wrote:
> > Hi Dave,
> > 
> > On 2021/5/6 21:05, Dr. David Alan Gilbert wrote:
> > > * Kunkun Jiang (jiangkunkun@huawei.com) wrote:
> > > > Hi all,
> > > Hi,
> > > 
> > > > Recently I am learning about the part of live migration.
> > > > I have a question about the last round.
> > > > 
> > > > When the pending_size is less than the threshold, it will enter
> > > > the last round and call migration_completion(). It will stop the
> > > > source and sent the remaining dirty pages and devices' status
> > > > information to the destination. The destination will load these
> > > > information and start the VM.
> > > > 
> > > > If there is an error at the destination at this time, it will exit
> > > > directly, and the source will not be able to detect the error
> > > > and recover. Because the source will not call
> > > > migration_detect_error().
> > > > 
> > > > Is my understanding correct?
> > > > Should the source wait the result of the last round of destination ?
> > > Try setting the 'return-path' migration capability on both the source
> > > and destination;  I think it's that option will cause the destination to
> > > send an OK/error at the end and the source to wait for it.
> > Thank you for your reply!
> > The 'return-path' migration capability solved my question. 😁
> > 
> > But why not set it as the default? In my opinion, it is a basic ability
> > of live migration. We need it to judge whether the last round of the
> > destination is successful in the way of 'precopy'.
> 
> I think it should be enabled as long as both sides support it; though may not
> be suitable as default (at least in QEMU) so as to consider old binaries.

Right, we can't break migration from old qemu's (I think libvirt will
use it if available by default).

(Maybe some day we should enable it by default and deprecate turning it
off)

Dave

> Thanks,
> 
> -- 
> Peter Xu
> 
-- 
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK