RE: [virtio-dev] Re: [PATCH 0/5] virtio: introduce SUSPEND bit and vq state

[Parav Pandit <parav@nvidia.com>]

> From: Jason Wang <jasowang@redhat.com>
> Sent: Thursday, September 21, 2023 9:50 AM

> Parav, I think I've clarified several times:
> 
> migration using the admin command is probably fine in some use cases.
>
This is definitely, was not clear to me.
I am 100% clear now.
 
> What's not fine, is:
> 
> Mandate the admin command to be the only way for migration.
>
For sure, my series did not mandate that either.
I kept asking if we both can converge it will be really good to merge two use cases, we should.
If we cannot because of technical issues, than both methods to exists to address two different use cases.

> Are we on the same page for my concern now?
> 
Yes.

> > What is the advantage of descriptor posting using virtqueue. It is the way of
> virtio spec...
> >
> > > > Well, it is one way to achieve it.
> > > > There may be different way to do all bulk data transfer without
> > > > admin
> > > commands.
> > >
> > > Why is virtqueue the only way to do bulk data transferring? Can't
> > > DMA be initiated by other-way?
> > >
> >
> > Sure, what is the disadvantage of existing mechanism of virtqueue that can do
> following.
> > 1. Ability to do DMA
> > 2. agnostic of the DMA who do not want to do DMA
> 
> I don't understand this.
>
Admin commands can work without DMA, right because they are transported using admin queue.
 
> > 3. Ability to multiple command executions in parallel
> 
> Each device has their self-contained interface, why can't the commands be
> executed in parallel.
> 
Within the device it cannot if the interface is synchronous.

> > 4. Non blocking interface for driver that does not require any kind of
> > polling
> 
> Are you saying the interrupt can only work for virtqueue?
>
No. I am saying if one has to invent a interface that satisfy above needs, it will become a virtqueue.
And if it is not, one should list those disadvantages and cost of new interface, and explain its benefits.
Such interface should be generic one too.
 
> >
> > Why to invent new DMA scheme which at does all the 4 tasks?
> 
> It's simply because admin virtqueue can not work for all the cases. I think you've
> agreed on this, no?
>
I think it may work for nested case as well at the cost of replicating it on each device and adding special plumbing to isolate it, so that guest cannot issue driver notifications to it.

> > First please list down disadvantages of admin queue + show all 4 things are
> achieved using new DMA interface.
> > That will help to understand why new dma interface is needed.
> 
> I can give you a simple example. For example, what happens if we want to
> migrate the owner? Having another owner for this owner is not a good answer.

That is the nesting, donât see any difference with other nesting.