[Qemu-devel] More structured migration URIs?

[Qemu-devel] More structured migration URIs?

[Dr. David Alan Gilbert]

Hi,
  I keep thinking of things where it might make sense to add
options onto the migration URIs and wondered if it makes
sense to restructure the migration URIs; my proposal would be:

  a) Restructure tcp:hhhh:pppp into   protocol=tcp,host=hhhh,port=pppp
  b) Have a requirement that protocol= is the first entry in the list
  c) If it doesn't start protocol= then it's the old format.
  d) This would also change in the 'migrate' command to keep it symmetric

Eric/Daniel does this make sense for libvirt?

My current set of things I might want to add are:
  1) A flag saying if a return channel is needed
     For sockets qemu can open this afterwards when needed, but Dave Gibson's
     review of my postcopy world pointed out that it might not be that
     easy for all protocols to open the reverse later.
  2) Flags for opening multiple sockets/FDs - e.g. to pass the pages down
     a separate fd.

Thoughts?

Dave
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

Re: [Qemu-devel] More structured migration URIs?

[Daniel P. Berrange]

On Mon, Jan 05, 2015 at 12:14:25PM +0000, Dr. David Alan Gilbert wrote:
> Hi,
>   I keep thinking of things where it might make sense to add
> options onto the migration URIs and wondered if it makes
> sense to restructure the migration URIs; my proposal would be:
> 
>   a) Restructure tcp:hhhh:pppp into   protocol=tcp,host=hhhh,port=pppp
>   b) Have a requirement that protocol= is the first entry in the list
>   c) If it doesn't start protocol= then it's the old format.
>   d) This would also change in the 'migrate' command to keep it symmetric
> 
> Eric/Daniel does this make sense for libvirt?
> 
> My current set of things I might want to add are:
>   1) A flag saying if a return channel is needed
>      For sockets qemu can open this afterwards when needed, but Dave Gibson's
>      review of my postcopy world pointed out that it might not be that
>      easy for all protocols to open the reverse later.
>   2) Flags for opening multiple sockets/FDs - e.g. to pass the pages down
>      a separate fd.
> 
> Thoughts?

The QEMU migration URI is exposed via the libvirt API to applications, so
they can control the host/port of the target explicitly (to override
libvirt's default guess of a suitable host/port).

These extra things that you suggest adding are not things we neccessarily
want to expose to applications. So if QEMU changed the format, libvirt
would probably not accept this new format from applications - we would
force applications to always use the URI syntax and only allow host+port
to be specified. Internally libvirt would translate that to the new
key/value pair format, and add the extra flags / options as needed. We
would possibly add some options to our public API to configure extra
features as desired, separately from the URI.

More generally though, what is the advantage of encoding new things in
the migration URI, as opposed to adding new parameters to the QMP
migration command. The use of URIs in this scenario is really a hang
over from days of HMP. If we were designing the migration command today
I don't think we'd use URIs at all - we'd just have a QMP command with
explicit parameters for the hostname, the port number and anything else
we might want to set.  So if there are new features I'd be inclined to
just add more optional parameters to the QMP migration command and not
touch the URI format at all.

Regards,
Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|

Re: [Qemu-devel] More structured migration URIs?

[Dr. David Alan Gilbert]

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Mon, Jan 05, 2015 at 12:14:25PM +0000, Dr. David Alan Gilbert wrote:
> > Hi,
> >   I keep thinking of things where it might make sense to add
> > options onto the migration URIs and wondered if it makes
> > sense to restructure the migration URIs; my proposal would be:
> > 
> >   a) Restructure tcp:hhhh:pppp into   protocol=tcp,host=hhhh,port=pppp
> >   b) Have a requirement that protocol= is the first entry in the list
> >   c) If it doesn't start protocol= then it's the old format.
> >   d) This would also change in the 'migrate' command to keep it symmetric
> > 
> > Eric/Daniel does this make sense for libvirt?
> > 
> > My current set of things I might want to add are:
> >   1) A flag saying if a return channel is needed
> >      For sockets qemu can open this afterwards when needed, but Dave Gibson's
> >      review of my postcopy world pointed out that it might not be that
> >      easy for all protocols to open the reverse later.
> >   2) Flags for opening multiple sockets/FDs - e.g. to pass the pages down
> >      a separate fd.
> > 
> > Thoughts?
> 
> The QEMU migration URI is exposed via the libvirt API to applications, so
> they can control the host/port of the target explicitly (to override
> libvirt's default guess of a suitable host/port).

Ah OK, that's a shame.

> These extra things that you suggest adding are not things we neccessarily
> want to expose to applications. So if QEMU changed the format, libvirt
> would probably not accept this new format from applications - we would
> force applications to always use the URI syntax and only allow host+port
> to be specified. Internally libvirt would translate that to the new
> key/value pair format, and add the extra flags / options as needed. We
> would possibly add some options to our public API to configure extra
> features as desired, separately from the URI.

It all gets a bit more complicated when an application could specify
an arbitrary exec: uri through that API, rather than you knowing what it's
doing.

> More generally though, what is the advantage of encoding new things in
> the migration URI, as opposed to adding new parameters to the QMP
> migration command. The use of URIs in this scenario is really a hang
> over from days of HMP. If we were designing the migration command today
> I don't think we'd use URIs at all - we'd just have a QMP command with
> explicit parameters for the hostname, the port number and anything else
> we might want to set.  So if there are new features I'd be inclined to
> just add more optional parameters to the QMP migration command and not
> touch the URI format at all.

I was only interested in adding options to the -incoming side rather than the
'migrate' side, but thought if I was changing the URI syntax I may as well
make it consistent.   If you wanted to add options to the -incoming side
what would be easiest for you?

Dave

> 
> Regards,
> Daniel
> -- 
> |: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
> |: http://libvirt.org              -o-             http://virt-manager.org :|
> |: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
> |: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

Re: [Qemu-devel] More structured migration URIs?

[Daniel P. Berrange]

On Mon, Jan 05, 2015 at 12:37:23PM +0000, Dr. David Alan Gilbert wrote:
> * Daniel P. Berrange (berrange@redhat.com) wrote:
> > On Mon, Jan 05, 2015 at 12:14:25PM +0000, Dr. David Alan Gilbert wrote:
> > > Hi,
> > >   I keep thinking of things where it might make sense to add
> > > options onto the migration URIs and wondered if it makes
> > > sense to restructure the migration URIs; my proposal would be:
> > > 
> > >   a) Restructure tcp:hhhh:pppp into   protocol=tcp,host=hhhh,port=pppp
> > >   b) Have a requirement that protocol= is the first entry in the list
> > >   c) If it doesn't start protocol= then it's the old format.
> > >   d) This would also change in the 'migrate' command to keep it symmetric
> > > 
> > > Eric/Daniel does this make sense for libvirt?
> > > 
> > > My current set of things I might want to add are:
> > >   1) A flag saying if a return channel is needed
> > >      For sockets qemu can open this afterwards when needed, but Dave Gibson's
> > >      review of my postcopy world pointed out that it might not be that
> > >      easy for all protocols to open the reverse later.
> > >   2) Flags for opening multiple sockets/FDs - e.g. to pass the pages down
> > >      a separate fd.
> > > 
> > > Thoughts?
> > 
> > The QEMU migration URI is exposed via the libvirt API to applications, so
> > they can control the host/port of the target explicitly (to override
> > libvirt's default guess of a suitable host/port).
> 
> Ah OK, that's a shame.
> 
> > These extra things that you suggest adding are not things we neccessarily
> > want to expose to applications. So if QEMU changed the format, libvirt
> > would probably not accept this new format from applications - we would
> > force applications to always use the URI syntax and only allow host+port
> > to be specified. Internally libvirt would translate that to the new
> > key/value pair format, and add the extra flags / options as needed. We
> > would possibly add some options to our public API to configure extra
> > features as desired, separately from the URI.
> 
> It all gets a bit more complicated when an application could specify
> an arbitrary exec: uri through that API, rather than you knowing what it's
> doing.

True, the QMP command would need to support a union of different options
switched depending on the protocol type. Not sure if QMP can do that or
not ?

> > More generally though, what is the advantage of encoding new things in
> > the migration URI, as opposed to adding new parameters to the QMP
> > migration command. The use of URIs in this scenario is really a hang
> > over from days of HMP. If we were designing the migration command today
> > I don't think we'd use URIs at all - we'd just have a QMP command with
> > explicit parameters for the hostname, the port number and anything else
> > we might want to set.  So if there are new features I'd be inclined to
> > just add more optional parameters to the QMP migration command and not
> > touch the URI format at all.
> 
> I was only interested in adding options to the -incoming side rather than the
> 'migrate' side, but thought if I was changing the URI syntax I may as well
> make it consistent.   If you wanted to add options to the -incoming side
> what would be easiest for you?

Historically the use of -incoming would block use of the monitor commands
while the migration data was being loaded, but I'm not sure if that is
still the case. Libvirt has a had long time desire to be able to query
the 'info migration' on the target side to be able to report on progress
reading the migration data stream. Last time we discussed this all someone
mentioned that perhaps a long time option is switch to replace the use of
the '-incoming' arg, with a 'migrate-incoming' QMP command. I don't know
if that's practical or not.

Back to your question though, libvirt's handling of the URIs for the
-incoming side is completely separate from handling of URIs for the
outgoing side. So I've not objection to changes to the -incoming arg
syntax as that's completely isolated from applications/users.

Regards,
Daniel
-- 
|: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
|: http://libvirt.org              -o-             http://virt-manager.org :|
|: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
|: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|

Re: [Qemu-devel] More structured migration URIs?

[Dr. David Alan Gilbert]

* Daniel P. Berrange (berrange@redhat.com) wrote:
> On Mon, Jan 05, 2015 at 12:37:23PM +0000, Dr. David Alan Gilbert wrote:
> > * Daniel P. Berrange (berrange@redhat.com) wrote:
> > > On Mon, Jan 05, 2015 at 12:14:25PM +0000, Dr. David Alan Gilbert wrote:
> > > > Hi,
> > > >   I keep thinking of things where it might make sense to add
> > > > options onto the migration URIs and wondered if it makes
> > > > sense to restructure the migration URIs; my proposal would be:
> > > > 
> > > >   a) Restructure tcp:hhhh:pppp into   protocol=tcp,host=hhhh,port=pppp
> > > >   b) Have a requirement that protocol= is the first entry in the list
> > > >   c) If it doesn't start protocol= then it's the old format.
> > > >   d) This would also change in the 'migrate' command to keep it symmetric
> > > > 
> > > > Eric/Daniel does this make sense for libvirt?
> > > > 
> > > > My current set of things I might want to add are:
> > > >   1) A flag saying if a return channel is needed
> > > >      For sockets qemu can open this afterwards when needed, but Dave Gibson's
> > > >      review of my postcopy world pointed out that it might not be that
> > > >      easy for all protocols to open the reverse later.
> > > >   2) Flags for opening multiple sockets/FDs - e.g. to pass the pages down
> > > >      a separate fd.
> > > > 
> > > > Thoughts?
> > > 
> > > The QEMU migration URI is exposed via the libvirt API to applications, so
> > > they can control the host/port of the target explicitly (to override
> > > libvirt's default guess of a suitable host/port).
> > 
> > Ah OK, that's a shame.
> > 
> > > These extra things that you suggest adding are not things we neccessarily
> > > want to expose to applications. So if QEMU changed the format, libvirt
> > > would probably not accept this new format from applications - we would
> > > force applications to always use the URI syntax and only allow host+port
> > > to be specified. Internally libvirt would translate that to the new
> > > key/value pair format, and add the extra flags / options as needed. We
> > > would possibly add some options to our public API to configure extra
> > > features as desired, separately from the URI.
> > 
> > It all gets a bit more complicated when an application could specify
> > an arbitrary exec: uri through that API, rather than you knowing what it's
> > doing.
> 
> True, the QMP command would need to support a union of different options
> switched depending on the protocol type. Not sure if QMP can do that or
> not ?
> 
> > > More generally though, what is the advantage of encoding new things in
> > > the migration URI, as opposed to adding new parameters to the QMP
> > > migration command. The use of URIs in this scenario is really a hang
> > > over from days of HMP. If we were designing the migration command today
> > > I don't think we'd use URIs at all - we'd just have a QMP command with
> > > explicit parameters for the hostname, the port number and anything else
> > > we might want to set.  So if there are new features I'd be inclined to
> > > just add more optional parameters to the QMP migration command and not
> > > touch the URI format at all.
> > 
> > I was only interested in adding options to the -incoming side rather than the
> > 'migrate' side, but thought if I was changing the URI syntax I may as well
> > make it consistent.   If you wanted to add options to the -incoming side
> > what would be easiest for you?
> 
> Historically the use of -incoming would block use of the monitor commands
> while the migration data was being loaded, but I'm not sure if that is
> still the case. Libvirt has a had long time desire to be able to query
> the 'info migration' on the target side to be able to report on progress
> reading the migration data stream. Last time we discussed this all someone
> mentioned that perhaps a long time option is switch to replace the use of
> the '-incoming' arg, with a 'migrate-incoming' QMP command. I don't know
> if that's practical or not.

I've just checked, and qmp seems to be happy with -incoming  and we see:

{ "execute": "query-status" }
{"return": {"status": "inmigrate", "singlestep": false, "running": false}}

> Back to your question though, libvirt's handling of the URIs for the
> -incoming side is completely separate from handling of URIs for the
> outgoing side. So I've not objection to changes to the -incoming arg
> syntax as that's completely isolated from applications/users.

OK, thanks, I'll have a think about it.

Dave

> 
> Regards,
> Daniel
> -- 
> |: http://berrange.com      -o-    http://www.flickr.com/photos/dberrange/ :|
> |: http://libvirt.org              -o-             http://virt-manager.org :|
> |: http://autobuild.org       -o-         http://search.cpan.org/~danberr/ :|
> |: http://entangle-photo.org       -o-       http://live.gnome.org/gtk-vnc :|
--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK

Re: [Qemu-devel] More structured migration URIs?

[Eric Blake]

[-- Attachment #1: Type: text/plain, Size: 621 bytes --]

On 01/05/2015 05:45 AM, Daniel P. Berrange wrote:

>> It all gets a bit more complicated when an application could specify
>> an arbitrary exec: uri through that API, rather than you knowing what it's
>> doing.
> 
> True, the QMP command would need to support a union of different options
> switched depending on the protocol type. Not sure if QMP can do that or
> not ?

Yes, the qapi allows for QMP commands to have a union of secondary
parameters determined by the value of a discriminator key.

-- 
Eric Blake   eblake redhat com    +1-919-301-3266
Libvirt virtualization library http://libvirt.org


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 604 bytes --]