Linux Virtualization - Apr 2015 - [PATCH v2 1/6] virtio_balloon: transitional interface

On Wed, Apr 01, 2015 at 12:22:44PM +0200, Cornelia Huck
wrote:
> On Wed, 1 Apr 2015 11:43:46 +0200
> "Michael S. Tsirkin" <mst at redhat.com> wrote:
> 
> > On Wed, Apr 01, 2015 at 07:53:14PM +1030, Rusty Russell wrote:
> > > "Michael S. Tsirkin" <mst at redhat.com> writes:
> > > > On Wed, Apr 01, 2015 at 02:17:23PM +1030, Rusty Russell
wrote:
> > > >> I would leave the device *exactly* as is, ugly structure
packing and
> > > >> all.
> > > >
> > > > But why?  It's going to be used for years, might as well
make it clean?
> > > 
> > > Because the only spec which currently exists says to do that.
> > 
> > OK but the only spec which currently exists also says it's a
legacy only
> > device, so driver must not set VERSION_1.  So surely, we can make
minor
> > changes when VERSION_1 is set, like we did for other devices.
> 
> But we don't plan to replace the other devices, so it makes sense to do
> some changes for 1.0.
I'm not sure what the above says. Do you agree with
making minor changes in device behaviour?
Also to be clear, I think this is 1.1 material.
> > 
> > Let me post the latest patches I'm working on,
> > see what you think then.
> > 
> > >  We do
> > > need a new virtio memballoon spec, but it'll look nothing
like this
> > > anyway.
> > > 
> > > Cheers,
> > > Rusty.
> > 
> > I think it's going to have significantly different semantics, too,
> > so not much value in making that one work with current
> > drivers, right?
> > 
> 
> So why not just keep virtio-balloon as-is and just specify endianness
> etc. for 1.0? Keeps the old drivers going without hacks,
> and we can
> start with a fresh driver for the new virtio-balloon.
Well it doesn't really, we need cpu_to_virtio in a bunch of
places anyway.

So I kind of prefer making it clean, even just to avoid setting a bad
example for other devices.

Let me post the new patch where it's all fixed in a cleaner way, and
everyone can discuss whether it's too much work.

-- 
MST

On Wed, 1 Apr 2015 12:28:30 +0200
"Michael S. Tsirkin" <mst at redhat.com> wrote:
> On Wed, Apr 01, 2015 at 12:22:44PM +0200, Cornelia Huck wrote:
> > On Wed, 1 Apr 2015 11:43:46 +0200
> > "Michael S. Tsirkin" <mst at redhat.com> wrote:
> > 
> > > On Wed, Apr 01, 2015 at 07:53:14PM +1030, Rusty Russell wrote:
> > > > "Michael S. Tsirkin" <mst at redhat.com>
writes:
> > > > > On Wed, Apr 01, 2015 at 02:17:23PM +1030, Rusty Russell
wrote:
> > > > >> I would leave the device *exactly* as is, ugly
structure packing and
> > > > >> all.
> > > > >
> > > > > But why?  It's going to be used for years, might as
well make it clean?
> > > > 
> > > > Because the only spec which currently exists says to do
that.
> > > 
> > > OK but the only spec which currently exists also says it's a
legacy only
> > > device, so driver must not set VERSION_1.  So surely, we can make
minor
> > > changes when VERSION_1 is set, like we did for other devices.
> > 
> > But we don't plan to replace the other devices, so it makes sense
to do
> > some changes for 1.0.
> 
> I'm not sure what the above says. Do you agree with
> making minor changes in device behaviour?
The other way around.
> Also to be clear, I think this is 1.1 material.
Btw, I'd really like to see your proposed spec updates.
> 
> > > 
> > > Let me post the latest patches I'm working on,
> > > see what you think then.
> > > 
> > > >  We do
> > > > need a new virtio memballoon spec, but it'll look
nothing like this
> > > > anyway.
> > > > 
> > > > Cheers,
> > > > Rusty.
> > > 
> > > I think it's going to have significantly different semantics,
too,
> > > so not much value in making that one work with current
> > > drivers, right?
> > > 
> > 
> > So why not just keep virtio-balloon as-is and just specify endianness
> > etc. for 1.0? Keeps the old drivers going without hacks,
> > and we can
> > start with a fresh driver for the new virtio-balloon.
> 
> Well it doesn't really, we need cpu_to_virtio in a bunch of
> places anyway.
Of course, but what about keeping changes minimal?
> 
> So I kind of prefer making it clean, even just to avoid setting a bad
> example for other devices.
> 
> Let me post the new patch where it's all fixed in a cleaner way, and
> everyone can discuss whether it's too much work.
>

On Wed, Apr 01, 2015 at 12:57:48PM +0200, Cornelia Huck
wrote:
> On Wed, 1 Apr 2015 12:28:30 +0200
> "Michael S. Tsirkin" <mst at redhat.com> wrote:
> 
> > On Wed, Apr 01, 2015 at 12:22:44PM +0200, Cornelia Huck wrote:
> > > On Wed, 1 Apr 2015 11:43:46 +0200
> > > "Michael S. Tsirkin" <mst at redhat.com> wrote:
> > > 
> > > > On Wed, Apr 01, 2015 at 07:53:14PM +1030, Rusty Russell
wrote:
> > > > > "Michael S. Tsirkin" <mst at
redhat.com> writes:
> > > > > > On Wed, Apr 01, 2015 at 02:17:23PM +1030, Rusty
Russell wrote:
> > > > > >> I would leave the device *exactly* as is, ugly
structure packing and
> > > > > >> all.
> > > > > >
> > > > > > But why?  It's going to be used for years,
might as well make it clean?
> > > > > 
> > > > > Because the only spec which currently exists says to do
that.
> > > > 
> > > > OK but the only spec which currently exists also says
it's a legacy only
> > > > device, so driver must not set VERSION_1.  So surely, we can
make minor
> > > > changes when VERSION_1 is set, like we did for other
devices.
> > > 
> > > But we don't plan to replace the other devices, so it makes
sense to do
> > > some changes for 1.0.
> > 
> > I'm not sure what the above says. Do you agree with
> > making minor changes in device behaviour?
> 
> The other way around.
> 
> > Also to be clear, I think this is 1.1 material.
> 
> Btw, I'd really like to see your proposed spec updates.
Just sent.
> > 
> > > > 
> > > > Let me post the latest patches I'm working on,
> > > > see what you think then.
> > > > 
> > > > >  We do
> > > > > need a new virtio memballoon spec, but it'll look
nothing like this
> > > > > anyway.
> > > > > 
> > > > > Cheers,
> > > > > Rusty.
> > > > 
> > > > I think it's going to have significantly different
semantics, too,
> > > > so not much value in making that one work with current
> > > > drivers, right?
> > > > 
> > > 
> > > So why not just keep virtio-balloon as-is and just specify
endianness
> > > etc. for 1.0? Keeps the old drivers going without hacks,
> > > and we can
> > > start with a fresh driver for the new virtio-balloon.
> > 
> > Well it doesn't really, we need cpu_to_virtio in a bunch of
> > places anyway.
> 
> Of course, but what about keeping changes minimal?
> > 
> > So I kind of prefer making it clean, even just to avoid setting a bad
> > example for other devices.
> > 
> > Let me post the new patch where it's all fixed in a cleaner way,
and
> > everyone can discuss whether it's too much work.
> >